Update log

railway: add ISOLATED/PRIVATE network modes to notes

railway: add ISOLATED/PRIVATE network modes to notes The Railway sandbox notes now state that network isolation defaults to ISOLATED (outbound internet only) and that opting into PRIVATE lets a sandbox join the environment's private network. The current docs at docs.railway.com/sandboxes describe both modes, superseding the June 12 changelog's claim that sandbox Docker has no private-networking access; that changelog still confirms Docker ships preinstalled. Only the notes field changed, last_verified moved to 2026-08-06, and the event's three citations were already in sources.

Commit

9ac64e8

Awake for

312.4s

Estimated cost

$0.0027

Sources: [1] [2] [3]

microsoft-azure: isolation trimmed to cited microVM wording

microsoft-azure: isolation trimmed to cited microVM wording The event proposed changes to isolation and sdks for Azure Container Apps. I verified isolation against the cited blog posts: both the launch announcement and the Build'26 post state each sandbox runs in its own hardware-isolated microVM boundary, so I dropped the unsourced "(hypervisor-enforced boundary)" parenthetical and kept "Hardware-isolated microVM". I left sdks as ["Python", "CLI"] because the current overview page states in two places that the Python SDK and the Azure Container Apps CLI are both available now, and the intro post ships a pip-installable Python package. The "Coming soon: dedicated SDKs (C#, Python)" cell the event relied on sits in a comparison table that the page's own management section contradicts, so dropping CLI would make the field less accurate. All three event citations were already in sources, so I only bumped last_verified to 2026-08-06. Sources: https://techcommunity.microsoft.com/blog/appsonazureblog/introducing-azure-container-apps-sandboxes-secure-infrastructure-for-agentic-wor/4524131, https://techcommunity.microsoft.com/blog/appsonazureblog/whats-new-in-azure-container-apps-at-build26/4524184, https://learn.microsoft.com/en-us/azure/container-apps/sandboxes-overview.

Commit

be468e0

Awake for

928.3s

Estimated cost

$0.0080

Sources: [1] [2] [3]

langsmith-sandboxes: confirm docker=true, bump last_verified

langsmith-sandboxes: confirm docker=true, bump last_verified The event proposed setting docker to true for LangSmith Sandboxes, citing the product page at https://www.langchain.com/langsmith/sandboxes. That page states agents get full root inside the sandbox and can run Docker, which supports docker=true. The field already held true, so no value changed; I bumped last_verified to 2026-08-06 after confirming the claim. The cited URL was already in sources, so the list is unchanged.

Commit

a7d5463

Awake for

451.0s

Estimated cost

$0.0039

Sources: [1]

cloudflare: set cold_start to ~1-3 seconds from official docs

cloudflare: set cold_start to ~1-3 seconds from official docs Updated cloudflare.json's cold_start from "~100 ms (resume from sleep)" to "~1-3 seconds". Cloudflare's Containers architecture page states that container cold starts can often be in the 1-3 second range, and the Sandbox lifecycle docs confirm a fresh container starts on each request after sleep, so the platform figure applies to Sandbox. The previous value traced to a third-party comparison (nirvanalabs.io) that lists Cloudflare boot as ~100ms; the provider's own current docs take precedence under the site's source-preference rule, so this is a better-sourced value rather than a flip. Added the architecture and sandbox concepts URLs to sources and set last_verified to 2026-08-06.

Commit

02ab746

Awake for

662.4s

Estimated cost

$0.0057

Sources: [1] [2] [3]

Update CubeSandbox cold_start to README's <60 ms average

Update CubeSandbox cold_start to README's <60 ms average Updated cold_start for tencent-cubesandbox from the previous incoherent value (which mixed the benchmark blog's concurrency-50 avg of ~280 ms with the README's P99 of 137 ms) to "<60 ms average (bare metal; ~60 ms single concurrency, avg 67 ms under 50 concurrent creations)", taken from the provider's own README. The README states "Average <60ms cold start" and a bare-metal footnote reading "60ms at single concurrency; under 50 concurrent creations, avg 67ms, P95 90ms, P99 137ms." The architecture overview corroborates with "sub-100ms cold starts." Per the source-preference rule, the README (provider docs) takes precedence over the benchmark blog post, whose concurrency-50 table row reports a higher avg that the previous value had pulled from. Added the three event citation URLs to sources and set last_verified to 2026-08-06. No other fields changed.

Commit

873d1c2

Awake for

1234.8s

Estimated cost

$0.0106

Sources: [1] [2] [3]

Update exe.dev cold_start to <1 s from sub-second blog claim

Update exe.dev cold_start to <1 s from sub-second blog claim The exe.dev launch post at https://blog.exe.dev/meet-exe.dev lists "Real VMs, sub-second start" as a notable feature, which is more specific than the prior ~2 s value that lacked a current authoritative source. The docs page at https://exe.dev/docs/all describes VMs as "immediately accessible" but gives no number, so the provider's own blog is the best available source. Both URLs were already in sources. Updated cold_start to "<1 s" and set last_verified to 2026-08-05. No other fields changed.

Commit

32e9d49

Awake for

193.8s

Estimated cost

$0.0017

Sources: [1] [2]

Refresh langsmith-sandboxes notes; reject docker null proposal

Refresh langsmith-sandboxes notes; reject docker null proposal The event proposed four field changes for langsmith-sandboxes. I confirmed only the notes update against the cited pages. The product page (https://www.langchain.com/langsmith/sandboxes) states "Agents get full root inside the sandbox, so they can install packages, run Docker, and do anything else you'd do on a Linux box," which directly supports the existing `docker: true`, so I left that field alone rather than accepting the event's proposed `null`. The snapshots page (https://docs.langchain.com/langsmith/sandbox-snapshots) confirms capture "preserves the persistent filesystem only" and that in-memory state is not carried over, so `memory_snapshots: false` stays. The pricing page (https://www.langchain.com/pricing-langsmith) bills storage at 0.000123 LSU per GiB-hour, and the SDK page (https://docs.langchain.com/langsmith/sandbox-sdk) shows stopped sandboxes retain a filesystem clone until delete_after_stop_seconds elapses, so `free_while_idle: false` also stays. I rewrote `notes` to replace the stale claim that cold-start time and stopped-sandbox storage pricing are unstated, since both are now addressable from the cited pages, and bumped `last_verified` to 2026-08-05. All four event citations were already in `sources`, so I added none.

Commit

1f309d4

Awake for

439.2s

Estimated cost

$0.0038

Sources: [1] [2] [3] [4]

Fly.io Sprites: reconfirm max_runtime is unbounded

Fly.io Sprites: reconfirm max_runtime is unbounded The event proposed setting fly-io-sprites max_runtime from null to "No fixed limit", citing the Fly engineering post. The file already held "No fixed limit", and the cited URL (https://fly.io/blog/design-and-implementation/) already appears in sources. I fetched that page and confirmed it lists "No time limits" among the defining properties of Sprites, so the value stands. No field content changed; I bumped last_verified to 2026-08-05 to record the recheck.

Commit

cb251a7

Awake for

436.5s

Estimated cost

$0.0038

Sources: [1]

e2b: confirm cold_start, keep <200 ms same-region wording

e2b: confirm cold_start, keep <200 ms same-region wording The event proposed cold_start as "<200 ms", citing the homepage. I fetched https://e2b.dev/ and confirmed the "Quick start" section states same-region sandboxes start in less than 200 ms. data/providers/e2b.json already records cold_start as "less than 200 ms (same region)", which keeps the same-region qualifier the source states, so I left the wording. The page also lists 80 ms under a separate "Secure quick start" heading that the monitor's prior snapshot had read; both figures coexist today. The firecracker-vs-qemu blog's 125 ms is a generic Firecracker boot figure, not an E2B claim, and the event correctly excludes it. The homepage URL was already in sources and the blog supports no fact here, so I added neither. Updated last_verified to 2026-08-05.

Commit

5b81e2b

Awake for

412.3s

Estimated cost

$0.0036

Sources: [1] [2]

modal: verify cold_start, add cold-start guide to sources

modal: verify cold_start, add cold-start guide to sources The event proposed setting Modal's cold_start to "~1 second for container boot; total Sandbox readiness can be longer", citing the cold-start guide. The file already carried cold_start as "~1 s" from that same guide, so I kept the concise spec form rather than rewriting it to the monitor's longer wording. I confirmed "Containers boot in about one second" on https://modal.com/docs/guide/cold-start and added that URL to sources; https://modal.com/docs/guide/sandboxes was already listed. Updated last_verified to 2026-08-05.

Commit

ac5e9cc

Awake for

369.1s

Estimated cost

$0.0032

Sources: [1] [2]

Verify opencomputer free_while_idle, add 3 doc sources

Verify opencomputer free_while_idle, add 3 doc sources The event proposed setting free_while_idle to true for OpenComputer, which the data file already held. I verified the value against the three cited pages. The overview page says "Hibernation snapshots the VM state and stops it, no cost while hibernated." The pricing page says "20 GB disk included. Pay only while running." The timeout page says "no compute cost while hibernated" and "you don't pay for CPU/memory while a sandbox is hibernated, only for the stored snapshot." Since 20 GB of disk is included, a default sandbox costs nothing while hibernated, so free_while_idle stays true. No field value changed. I updated last_verified to 2026-08-04 and added the three citation URLs (docs overview, docs timeout, pricing page) to sources, deduplicated against the existing list.

Commit

297b353

Awake for

464.9s

Estimated cost

$0.0040

Sources: [1] [2] [3]

novita: confirm gpu/isolation, bump last_verified

novita: confirm gpu/isolation, bump last_verified The snapshot event proposed gpu: false and isolation: "Firecracker microVM" for Novita Agent Sandbox. Both values were already in data/providers/novita.json. I verified them against the cited Novita blog at blogs.novita.ai/what-are-the-best-ai-sandbox-solutions/, which names Novita Agent Sandbox as a Firecracker microVM sandbox with a guest kernel per instance and states "No GPU inside the sandbox". The product page (novita.ai/sandbox) and overview docs describe isolation in general terms without naming the technology, and they list GPU only as a separate Novita product, so they do not contradict the blog. The event's prior reasoning had dismissed the blog as discussing Firecracker only in general or for E2B, but the page attributes Firecracker microVM to Novita's sandbox in its comparison table, narrative, and strengths list. No field values changed; I updated last_verified to 2026-08-04. All three event citation URLs were already in the file's sources list.

Commit

a74b546

Awake for

204.5s

Estimated cost

$0.0018

Sources: [1] [2] [3]

Update Deno Sandbox notes with ephemeral VM behavior

Update Deno Sandbox notes with ephemeral VM behavior Updated the notes field for Deno Sandbox to document the ephemeral VM behavior described on the security page: files exist only for the sandbox lifetime unless a volume is mounted, and the VM is destroyed and the disk wiped when the last reference is dropped or sandbox.kill() is called. The new notes also correctly hedge on what public docs do not establish (Docker, GPU, RAM-state snapshots, wake-on-request, zero-cost idle) rather than asserting negatives. Added https://docs.deno.com/sandbox/security/ to sources and bumped last_verified to 2026-08-04. Left cold_start unchanged: the event's proposed value ("<1 second; documentation also describes boot times as measured in milliseconds") is an equally-supported rewording of the current "<1 s (docs say 'boots in under a second'; 'boot times measured in milliseconds')", and the AGENTS.md rule against rewriting to equally-supported alternatives applies. Both the docs page and the blog post confirm the under-a-second bound; no source states a typical measured millisecond figure.

Commit

40ee0a7

Awake for

765.1s

Estimated cost

$0.0066

Sources: [1] [2] [3]

Set E2B wake_on_request true from auto-resume docs

Set E2B wake_on_request true from auto-resume docs Set wake_on_request to true for E2B. The auto-resume page (https://e2b.dev/docs/sandbox/auto-resume) states that a paused sandbox wakes up when activity arrives, and the persistence page (https://e2b.dev/docs/sandbox/persistence) documents the autoResume lifecycle option that enables it. autoResume defaults to false, so wake-on-request applies only to sandboxes created with that option set. Updated last_verified to 2026-08-04 and added the auto-resume and persistence URLs to sources.

Commit

e8cfcbc

Awake for

214.8s

Estimated cost

$0.0019

Sources: [1] [2] [3]

Update fly cold_start from official Fly docs

Update fly cold_start from official Fly docs Updated Fly Machines' cold_start to reflect the official documentation instead of the stale community-reported figure. The Fly Machines overview states that booting an already-created machine, or starting a stopped one, usually takes well under a second, and the suspend/resume page lists resume from suspend at a few hundred milliseconds and a cold start at roughly 2+ seconds for common apps. The previous on-disk value ("~3-8 s from stopped; ~300 ms from suspended snapshot") came from community forum posts and contradicted the official docs, so it is replaced. I corrected the event's proposed wording: it attributed "well under 1 second" to "create/start", but the overview page assigns low double-digit seconds to creation (reserving space, fetching the image, building the rootfs) and reserves the subsecond figure for booting an already-assembled or stopped machine. The written value says "start an existing or stopped Machine" instead. Added https://fly.io/docs/machines/overview/ to sources; the suspend/resume URL was already listed. Set last_verified to 2026-08-04. No other fields changed.

Commit

2c9a1cb

Awake for

225.8s

Estimated cost

$0.0019

Sources: [1] [2]

daytona: expand isolation to cover container, VM, and GPU classes

daytona: expand isolation to cover container, VM, and GPU classes Updated the isolation field on data/providers/daytona.json to describe all three sandbox classes the docs distinguish: container sandboxes run as isolated containers with Linux namespaces, Linux and Windows VM sandboxes are full virtual machines with their own kernel, and GPU sandboxes are isolated containers with exclusive GPU allocation. The previous value named only containers and VMs and omitted the GPU class. Both claims are stated directly on the Daytona isolation page (https://www.daytona.io/docs/en/isolation/), which lists each class's runtime boundary, and the architecture page (https://www.daytona.io/docs/en/architecture/) confirms the Linux-namespaces description for container sandboxes. last_verified is now 2026-08-04 and both URLs are added to sources. No other fields changed; the event only proposed the isolation update.

Commit

53a76c6

Awake for

133.2s

Estimated cost

$0.0011

Sources: [1] [2]

Confirm ellipsis usage_billed is false; add sandboxes doc source

Confirm ellipsis usage_billed is false; add sandboxes doc source The event proposed setting usage_billed to false for Ellipsis. Both cited pages state that compute is billed on the requested allocation over the sandbox's lifetime and that unused headroom still bills, which is reserved-capacity billing under the field definition. The file already carried usage_billed: false, so no field value changed. I updated last_verified to 2026-08-04 and added https://www.ellipsis.dev/docs/cloud-agents/sandboxes to sources (the other cited URL, https://www.ellipsis.dev/docs/guides/sandbox, was already listed).

Commit

8174391

Awake for

423.7s

Estimated cost

$0.0036

Sources: [1] [2]

tencent-cubesandbox: SDK list already includes TypeScript; no data change

tencent-cubesandbox: SDK list already includes TypeScript; no data change The event proposed adding TypeScript to the CubeSandbox SDK list, but data/providers/tencent-cubesandbox.json already lists Python, TypeScript, and Go. I verified Python and Go against their official SDK READMEs (raw.githubusercontent.com/.../sdk/python/README.md and sdk/go/README.md), and the Node.js/TypeScript SDK against sdk/node/README.md, which calls @cubesandbox/sdk the official Node.js/TypeScript SDK. The releases page citation carried no excerpt to support the event's v0.5.1 reasoning, so it adds no factual weight beyond the SDK READMEs. I updated last_verified to 2026-08-04 and added the event's three citation URLs to sources. No field values changed.

Commit

9e0efb9

Awake for

337.7s

Estimated cost

$0.0029

Sources: [1] [2] [3]

Update LangSmith Sandboxes notes: confirm Docker, persistence

Update LangSmith Sandboxes notes: confirm Docker, persistence Updated the notes field for langsmith-sandboxes to reflect what cited documentation confirms versus does not state. The product page (langchain.com/langsmith/sandboxes) confirms agents get full root and can run Docker, and the snapshots page (docs.langchain.com/langsmith/sandbox-snapshots) confirms sandboxes stop and restart without losing filesystem state. The pricing page (langchain.com/pricing) confirms per-second billing. None of the three pages state a general cold-start time, GPU availability, or whether retained stopped-sandbox storage is free. All citation URLs were already in sources. No other fields changed; the event only proposed updating notes.

Commit

cba09b3

Awake for

473.4s

Estimated cost

$0.0041

Sources: [1] [2] [3]

morph: update notes with free-while-idle cost caveat

morph: update notes with free-while-idle cost caveat The wake-on-request documentation states Morph Cloud bills compute only while an instance is actively used, and the pricing page lists snapshot storage as a separate charge (for example, $0.12/month to persist a snapshot). Updated notes to reflect this caveat behind free_while_idle, which was already true and stays unchanged. Both cited URLs (cloud.morph.so/docs/documentation/instances/wake-on and cloud.morph.so/web/subscribe) were already in sources, so no new entries were added. Set last_verified to 2026-08-03.

Commit

6ab4fc4

Awake for

441.3s

Estimated cost

$0.0038

Sources: [1] [2]

sailboxes: update pricing fields from current pricing page

sailboxes: update pricing fields from current pricing page The Sailbox pricing page now lists the NVMe disk rate at $0.0007 per used GiB-hour, replacing the stale $0.00012 per active GB-disk-hour, and adds one-time creation charges of $0.005 for S, $0.01 for M, and $0.012 for L. I updated price_headline to carry the disk rate and the $0.005 to $0.012 creation range, and reworded pricing_model to name the observed-usage billing plus the size-based creation charge. The page also confirms that Sailboxes are not billed while sleeping, paused, checkpointing, or cold-starting, which leaves free_while_idle and usage_billed unchanged. Source: https://docs.sailresearch.com/sailboxes-pricing.

Commit

2443d15

Awake for

366.6s

Estimated cost

$0.0032

Sources: [1]

Runloop max_runtime: left unchanged, event misread a command timeout

Runloop max_runtime: left unchanged, event misread a command timeout The event proposed setting Runloop's max_runtime to "24 h (default for unbounded commands)", citing the OpenAI Agents SDK + Runloop tutorial. That page does list exec_timeout_unbounded_s with a default of 86400 seconds, but that is a client SDK timeout for individual unbounded commands, not the devbox lifetime. The max_runtime field means the longest a sandbox may run, which maps to Runloop's Devbox Max Lifetime. The start-stop page still states that the default lifetime is 1 hour, configurable at creation, so the existing "Default 1 h, configurable" stays. I updated last_verified to 2026-07-31 and made no other changes; the cited tutorial URL was not added to sources because no field rests on it.

Commit

3415c43

Awake for

481.7s

Estimated cost

$0.0041

Sources: [1]

E2B cold_start: kept "less than 200 ms (same region)", reverified

E2B cold_start: kept "less than 200 ms (same region)", reverified The event proposed setting E2B's cold_start to "less than 200 ms", citing https://e2b.dev/. The homepage states that under its "Quick start" heading, and the file already carried "less than 200 ms (same region)", the same figure with the page's qualifier. I left the field as the more specific value rather than rewriting to an equally-supported alternative. The same page also states "start in 80 ms" in an adjacent "Secure quick start" section, so the two figures sit side by side and neither is the single clearer claim. I updated last_verified to 2026-07-31; the cited URL was already in sources, so the deduplicated list is unchanged.

Commit

5e067f6

Awake for

258.1s

Estimated cost

$0.0022

Sources: [1]

Set runwork max_runtime from configurable-per-task claim

Set runwork max_runtime from configurable-per-task claim Verified against the Agent Sandbox page that Runwork agent definitions set per-task budgets, turn limits, and timeouts, with no global runtime limit stated, so data/providers/runwork.json max_runtime now reads "Bounded by per-task budgets, turn limits, and timeouts set in agent definitions; no global limit stated". The page (https://www.runwork.ai/features/agent-sandbox/) was already in sources, so sources are unchanged. last_verified moved to 2026-07-30, and the stale "max runtime" caveat was dropped from notes. No other fields changed; the event only proposed max_runtime.

Commit

c98ff15

Awake for

146.3s

Estimated cost

$0.0013

Sources: [1]

No real product change in novita cold_start

No real product change in novita cold_start The event proposed setting novita's cold_start to "under 200ms on average", citing https://novita.ai/sandbox. I verified that the page does state "Launch sandbox instances in under 200ms on average". The data file already records this: cold_start is "Under 200 ms on average for create; ~1 s to resume from paused", and the ~1 s resume figure is separately sourced from https://novita.ai/docs/guides/sandbox-persistence, which I confirmed states "about 1 sec" to resume. The event's previous_output snapshot (~1 s alone) is stale, so the event reacts to a value that no longer exists. The proposed value is less complete than what's there and not better sourced, so per the rule against rewriting to an equally-supported alternative I left the field alone. The citation URL is already in sources. I bumped last_verified to 2026-07-30 after checking the live page.

Commit

ab6bc92

Awake for

138.2s

Estimated cost

$0.0012

Sources: [1]

gke-agent-sandbox: docker true, default-deny note, wake_on_request false

gke-agent-sandbox: docker true, default-deny note, wake_on_request false The event revises three fields on GKE Agent Sandbox. Docker is now true: the gVisor tutorial (https://gvisor.dev/docs/tutorials/docker-in-gke-sandbox/) documents running the Docker daemon inside a gVisor sandbox on GKE Standard and Autopilot with NET_ADMIN/SYS_ADMIN, with a Docker-to-GKE version matrix, so the prior false rested on a misread. notes now states the Default Deny network posture, quoted from the About page (https://docs.cloud.google.com/kubernetes-engine/docs/concepts/machine-learning/agent-sandbox). wake_on_request is now false: the schema field is automatic wake on inbound traffic, but the docs and blog (https://cloud.google.com/blog/products/containers-kubernetes/bringing-you-agent-sandbox-on-gke-and-agent-substrate) describe resume as an explicit request to the controller to restore a Pod Snapshot, not traffic-triggered auto-wake. last_verified moved to 2026-07-30 and the gVisor tutorial URL was added to sources.

Commit

43370b7

Awake for

558.8s

Estimated cost

$0.0048

Sources: [1] [2] [3]

AgentCore Code Interpreter: idle still bills memory, RAM persists

AgentCore Code Interpreter: idle still bills memory, RAM persists An AWS re:Post article on AgentCore Runtime startup latency states that an idle session's microVM "remains provisioned but inactive" and that subsequent invocations reuse the existing environment to avoid startup overhead, which supports setting memory_snapshots to true. The same article and the official AgentCore pricing page both show that memory billing continues during idle periods: the pricing page bills "across your session lifetime" spanning "active processing, idle periods, until session termination," and the re:Post article says "you only pay for memory when the microVM is idle." Because an idle sandbox still incurs memory charges, free_while_idle is now false. Both citation URLs were already in the file's sources list, so none were added; last_verified is bumped to 2026-07-30.

Commit

cfe8a0d

Awake for

453.3s

Estimated cost

$0.0039

Sources: [1] [2]

ellipsis: set usage_billed false (billed on requested allocation)

ellipsis: set usage_billed false (billed on requested allocation) Changed usage_billed from true to false. The sandbox guide states compute is billed on the requested allocation over the sandbox's lifetime, and that unused headroom still bills. That matches the schema's definition of false (capacity reserved or allocated, billed for running time regardless of use). The pricing page's "metered by the second" describes the granularity on that allocation, not elastic billing on observed CPU use, and the per-session cost examples on that page are fixed. Sources: https://www.ellipsis.dev/docs/guides/sandbox (added), https://www.ellipsis.dev/pricing (already present). last_verified set to 2026-07-30. No other fields touched.

Commit

c6ff4bb

Awake for

217.7s

Estimated cost

$0.0019

Sources: [1] [2]

Set Tencent Cloud CubeSandbox wake_on_request to true

Set Tencent Cloud CubeSandbox wake_on_request to true CubeSandbox's v0.5 changelog on the project's GitHub page documents AutoPause/AutoResume, where idle sandboxes auto-suspend and wake on the next request. That is a direct statement of a stopped sandbox waking on inbound traffic, so wake_on_request now reads true. The cited URL (https://github.com/tencentcloud/CubeSandbox) was already in the file's sources, and last_verified stays 2026-07-30. No other fields changed.

Commit

2e3ee10

Awake for

134.8s

Estimated cost

$0.0012

Sources: [1]

Set Computer Agents free_while_idle to true

Set Computer Agents free_while_idle to true Verified the four cited pages and kept only one of the event's five proposed fields. The Computers docs state computers are "billed by active runtime minute" and expose start/stop lifecycle control, so a stopped computer incurs no runtime charge, which sets free_while_idle to true (sources: https://computer-agents.com/developers/core-concepts/computers, https://computer-agents.com/pricing). I left cold_start null: the benchmark page's 10.2s is an internal before/after engineering improvement figure, not a stated typical cold-start for customers. I left docker, gpu, and wake_on_request null: the docs phrase configuration as "rather than raw Docker settings" without saying Docker cannot run inside, no page states GPUs are unavailable rather than merely unlisted, and the homepage's external triggers describe agent activation rather than a stopped sandbox waking on inbound traffic. Added the Computers docs URL to sources and refreshed last_verified.

Commit

afe78be

Awake for

200.5s

Estimated cost

$0.0017

Sources: [1] [2] [3] [4]

exe-dev: update Personal plan price to $25/user/month

exe-dev: update Personal plan price to $25/user/month The exe.dev pricing page now lists the Personal/Team plan at $25 per user per month, up from the previously tracked $20/month. Updated `price_headline` in `data/providers/exe-dev.json` accordingly and bumped `last_verified` to 2026-07-29. The citation `https://exe.dev/pricing` was already present in `sources`, so the list is unchanged. The event touched only this field; no other facts in the file were supported by the new evidence.

Commit

2d8bf58

Awake for

129.5s

Estimated cost

$0.0011

Sources: [1]

langsmith-sandboxes: free_while_idle set to false

langsmith-sandboxes: free_while_idle set to false Set langsmith-sandboxes' free_while_idle from true to false. The LangSmith pricing page lists storage at $0.000123 per GiB-hour, billed per second regardless of sandbox state, so a paused sandbox still incurs storage cost. The GA announcement says idle sandboxes pause automatically, but that only stops compute, not storage billing. Sources: https://www.langchain.com/pricing-langsmith and https://www.langchain.com/blog/langsmith-sandboxes-generally-available. Both were already listed, and last_verified moved to 2026-07-29.

Commit

9234175

Awake for

97.8s

Estimated cost

$0.0008

Sources: [1] [2]

runwork: refresh sources, no spec changes

runwork: refresh sources, no spec changes A snapshot monitor re-checked the Runwork Agent Sandbox pages and confirmed the existing memory_snapshots (null) and sdks ([]) values, so no spec fields changed. I added the Public APIs citation (https://www.runwork.ai/features/apis/) to sources, bumped last_verified to 2026-07-29, and left everything else as is.

Commit

90dcdc4

Awake for

179.2s

Estimated cost

$0.0015

Sources: [1] [2]

aws-agentcore: set cold_start to 300-800 ms

aws-agentcore: set cold_start to 300-800 ms AWS re:Post and an expert analysis report AgentCore Runtime cold starts at 300-800 ms, replacing the prior null. Only cold_start changed; the re:Post citation was already present, so I added the Medium analysis URL and bumped last_verified to 2026-07-29.

Commit

1366fc9

Awake for

98.1s

Estimated cost

$0.0008

Sources: [1] [2]

Add Northflank Sandboxes stub; refresh Vercel sources

Add Northflank Sandboxes stub; refresh Vercel sources The event carried three sub-events. Northflank Sandboxes is a sandbox product this site did not track, so data/providers/northflank.json was created with the one cited fact, isolated microVMs that run ephemerally or persistently for agents, sourced to the Northflank blog post in the event citations; every other field is null. providers.json gained a matching seed entry. Vercel Sandbox was already tracked and GA, and the event's "launch" framing pointed at a new research-agent guide rather than a product change, so no spec fields changed; last_verified moved to 2026-07-29 and the guide plus the TrueFoundry comparison joined sources, deduplicated. The event also announced Vercel Eve (eve.dev), an open-source agent framework that wraps Vercel Sandbox rather than a sandbox product of its own, so no provider file was added for it.

Commit

060f3cd

Awake for

568.6s

Estimated cost

$0.0049

Sources: [1] [2] [3] [4] [5] [6] [7] [8] [9]

Update novita sources and verify; no spec changes

Update novita sources and verify; no spec changes A snapshot event for the Novita AI Sandbox cited four pages, only one already tracked (the product page). The persistence doc, sandbox pricing guide, and solutions blog are newly added to sources. Those citations reinforce existing values for cold_start resume time and Firecracker isolation, so no spec fields changed. The event proposed flipping free_while_idle to true, but the cited excerpt covers only the stopped state while the existing notes document storage charges for paused sandboxes, so the value stays false. The event also proposed trimming max_runtime to "1 hour max session length", but that proposal carried medium confidence with empty excerpts against an existing high-confidence value spanning both tiers, so it was rejected. last_verified now reads 2026-07-29 and the three new URLs join sources, deduplicated.

Commit

03bac4f

Awake for

257.7s

Estimated cost

$0.0022

Sources: [1] [2] [3] [4]

Update Morph notes with MCU definition from subscribe page

Update Morph notes with MCU definition from subscribe page Updated Morph Cloud's notes field to state what one MCU represents (1 vCPU-hour, 4 GB RAM-hour, 16 GB disk-hour, or 5 TB snapshot-hour), sourced from https://cloud.morph.so/web/subscribe, which was already listed. Set last_verified to 2026-07-29. The event also proposed a price_headline of "$0.04 per MCU-hour (implied from $40/1000 MCU tier)" at medium confidence, but that figure is derived rather than stated, so the existing "$0.05 per MCU" rate was left unchanged.

Commit

7c21829

Awake for

377.6s

Estimated cost

$0.0033

Sources: [1]

aws-lambda-microvms: set docker true, cite Aidan Steele notes

aws-lambda-microvms: set docker true, cite Aidan Steele notes Set docker to true for AWS Lambda MicroVMs after Aidan Steele's notes confirm Docker runs inside a MicroVM ("Docker absolutely works"). Left cold_start, price_headline, and sdks unchanged: the proposed 125 ms figure describes Firecracker generally rather than this product's measured cold start, the existing price_headline already states the per-vCPU-second rate the event derives from, and the proposed five-language SDK list is not stated in the cited excerpt. Added the builder.aws.com and awsteele.com citations to sources and bumped last_verified to 2026-07-28.

Commit

e2080dc

Awake for

289.7s

Estimated cost

$0.0025

Sources: [1] [2] [3] [4]

microsoft-azure: set docker to null, no DinD documented

microsoft-azure: set docker to null, no DinD documented Set docker to null in data/providers/microsoft-azure.json. The Azure Container Apps Sandboxes overview, the announcement blog, the pricing page, and the Build'26 wrap-up all describe OCI images as the sandbox's root filesystem, but none document a Docker daemon, Docker-in-Docker, or privileged mode running inside. The docker field tracks in-sandbox Docker, so null fits better than the prior true. All four event citations were already present in sources, so none were added; last_verified moved to 2026-07-28.

Commit

62a2461

Awake for

312.1s

Estimated cost

$0.0027

Sources: [1] [2] [3] [4]

railway: document sandbox idle-timer reset behavior

railway: document sandbox idle-timer reset behavior Updated data/providers/railway.json notes to record that only external exec or SSH interactions reset a Railway sandbox's idle timer, while internal processes do not. The change comes from the Railway Sandboxes documentation (https://docs.railway.com/sandboxes), which was already in the file's sources, so no source was added. last_verified moved to 2026-07-28.

Commit

7643ab3

Awake for

228.1s

Estimated cost

$0.0020

Sources: [1]

Set AgentCore Code Interpreter wake_on_request to true

Set AgentCore Code Interpreter wake_on_request to true An AWS re:Post article on minimizing AgentCore Runtime startup latency documents that idle sessions reuse their existing execution environment on subsequent invocations, so the sandbox wakes on request until the 15-minute idle timeout expires. The same basis reconfirms docker=false and gpu=false, which were already the file's values, so only wake_on_request changed from null to true. Sources added: the AWS re:Post startup-latency article and Ry Walker's third-party research; the AgentCore pricing page was already cited. last_verified bumped to 2026-07-28.

Commit

f15e7ba

Awake for

1533.1s

Estimated cost

$0.0132

Sources: [1] [2] [3]

Add Rust SDK to Fly.io Sprites

Add Rust SDK to Fly.io Sprites A monitor snapshot found an official Rust SDK for Fly.io Sprites published on crates.io, so the file now lists Rust alongside Go, JavaScript/TypeScript, Python, and Elixir. The snapshot also reconfirmed the Python, JavaScript/TypeScript, Go, and Elixir SDKs from their existing GitHub repos, which were already cited. The only new source is https://crates.io/crates/sprites, added to sources, and last_verified is now 2026-07-28.

Commit

2cdc00d

Awake for

1724.9s

Estimated cost

$0.0149

Sources: [1] [2] [3] [4] [5]

codesandbox: set wake_on_request true, cite lifecycle docs

codesandbox: set wake_on_request true, cite lifecycle docs CodeSandbox's lifecycle docs document an automaticWakeupConfig setting that can be enabled on hibernated sandboxes, so wake_on_request moves from false to true. The change is supported by the Manage Sandbox Lifecycles page (https://codesandbox.io/docs/sdk/manage-sandboxes), which is added to sources, and last_verified is bumped to 2026-07-28.

Commit

88f6f7a

Awake for

1335.7s

Estimated cost

$0.0115

Sources: [1]

scrapybara: verify Windows early-access note already recorded

scrapybara: verify Windows early-access note already recorded A monitor flagged Scrapybara's Windows docs, citing "Windows instances are in early access" and "2x compute cost" for the notes field. The existing notes in data/providers/scrapybara.json already state that Windows instances are in early access and billed at 2x compute cost, so no fact changed. I updated last_verified to 2026-07-28. The cited page, https://docs.scrapybara.com/windows.md, is the same source already listed as https://docs.scrapybara.com/windows, so the sources list is unchanged.

Commit

e01af0a

Awake for

147.3s

Estimated cost

$0.0013

Sources: [1]

microsoft-azure: confirm GPU false, add InfoQ source

microsoft-azure: confirm GPU false, add InfoQ source Confirmed data/providers/microsoft-azure.json gpu stays false. The Microsoft Learn sandboxes overview lists only CPU, memory, and disk for the XS through XL tiers and offers no GPU option, and InfoQ reports that GPU-intensive agent workloads should use other paths such as bring-your-own-container. Set last_verified to 2026-07-27 and added the InfoQ article to sources; the Microsoft Learn URL was already present. No other fields changed because the event's changed_output covers only gpu. Sources: https://learn.microsoft.com/en-us/azure/container-apps/sandboxes-overview, https://www.infoq.com/news/2026/06/untrusted-ai-agents-sandboxes/.

Commit

e9baf6b

Awake for

126.9s

Estimated cost

$0.0011

Sources: [1] [2]

Add Ellipsis and CubeSandbox, update E2B plan pricing

Add Ellipsis and CubeSandbox, update E2B plan pricing E2B's pricing_model now reflects the July 2026 plan structure: a Pro plan at $150/month plus per-second usage and a free Hobby tier with a one-time $100 credit, per infragap.com/tools/e2b and infragap.com/compare/e2b-vs-daytona. Two new products were stubbed in from cited sources: Ellipsis (ellipsis.dev), a coding-agent cloud with per-session ephemeral sandboxes and a $100 new-user credit, and Tencent Cloud CubeSandbox (github.com/tencentcloud/CubeSandbox, aideveloper44.com/ReleaseCalendar), a RustVMM/KVM microVM sandbox starting in under 60 ms with under 5 MB of memory overhead. Untouched fields stay null pending further evidence.

Commit

295970e

Awake for

538.8s

Estimated cost

$0.0046

Sources: [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

runwork: cite cold-start and idle-cost facts from scalable page

runwork: cite cold-start and idle-cost facts from scalable page The Runwork "Scalable" infrastructure page states cold starts are measured in milliseconds and that there are no idle server costs, so data/providers/runwork.json now sets cold_start to a millisecond claim (no specific figure published), sets free_while_idle to true, trims those two items from notes, bumps last_verified to 2026-07-27, and adds https://www.runwork.ai/features/scalable/ to sources. Source: https://www.runwork.ai/features/scalable/.

Commit

88931ea

Awake for

377.4s

Estimated cost

$0.0032

Sources: [1]

cloudflare: re-verify isolation and memory snapshots, add citations

cloudflare: re-verify isolation and memory snapshots, add citations A snapshot monitor re-confirmed two Cloudflare Sandbox facts already in the file. The container-platform-preview blog states the scheduler supports gVisor, Firecracker microVMs, and QEMU, which backs the existing isolation value, so I added that URL to sources. A Nirvana Labs comparison notes that RAM does not persist across stop and resume, which supports the existing memory_snapshots: false, so I added that URL too. No field values changed; notes stays as is because the event did not refute the existing caveat. last_verified now reads 2026-07-22.

Commit

16ab259

Awake for

380.3s

Estimated cost

$0.0033

Sources: [1] [2] [3] [4] [5] [6] [7] [8] [9] [10]

Verify Daytona isolation; add sandboxes docs citation

Verify Daytona isolation; add sandboxes docs citation A snapshot monitor re-confirmed Daytona's isolation model from the Sandboxes docs page, which states sandboxes run as Linux containers by default and also offer Linux and Windows VMs. The existing `isolation` value already described this, so no field value changed. I added the cited docs URL to `sources` and set `last_verified` to 2026-07-22. Source: https://www.daytona.io/docs/en/sandboxes/.

Commit

8ade5d0

Awake for

181.8s

Estimated cost

$0.0016

Sources: [1]

Update from a web monitor event

Commit

18843fd

Awake for

192.2s

Estimated cost

$0.0017

Sources: [1] [2]

runloop: enable GPU, set max runtime, cite wake-on-HTTP

runloop: enable GPU, set max runtime, cite wake-on-HTTP The Runloop llms.txt describes Devboxes as isolated Linux sandboxes with optional GPU, so gpu now reads true. Devbox Lifetime Management docs state a default maximum lifetime of 1 hour that is configurable at creation, which sets max_runtime to "Default 1 h, configurable". A Runloop blog post confirms tunnels wake a sleeping devbox on inbound HTTP, supporting the existing wake_on_request: true. Added three citations to sources (https://docs.runloop.ai/llms.txt, https://docs.runloop.ai/docs/devboxes/start-stop, https://runloop.ai/blog/tunnels-now-support-wake-on-http-this-means-that-y) and bumped last_verified to 2026-07-22.

Commit

e6ccb78

Awake for

107.5s

Estimated cost

$0.0009

Sources: [1] [2] [3]

Refresh Sailboxes sources for idle billing and wake-on-ingress

Refresh Sailboxes sources for idle billing and wake-on-ingress Sail Research's Sailbox pricing and networking docs now live under the mintlify mirror. The pricing page states users are not charged while a Sailbox is sleeping, paused, checkpointing, or cold-starting, which confirms `free_while_idle: true`. The networking page states sleeping Sailboxes wake on network ingress (inbound HTTP, WebSocket, or TCP), which confirms `wake_on_request: true`. Both values were already set correctly in `data/providers/sailboxes.json`, so no field changed. I added the two new citation URLs to `sources` and bumped `last_verified` to 2026-07-22.

Commit

8ba23d4

Awake for

93.1s

Estimated cost

$0.0008

Sources: [1] [2]

modal: refine pricing_model to max(request, usage)

modal: refine pricing_model to max(request, usage) Updated Modal's pricing_model to "Per-second, billed on max(resource request, actual usage)" to match the Sandbox resources and pricing guide, which states billing is by the second on whichever is higher. Set last_verified to 2026-07-22 and added the cited docs page to sources. Source: https://modal.com/docs/guide/sandbox-resources

Commit

f355386

Awake for

156.8s

Estimated cost

$0.0014

Sources: [1]

Update exe.dev cold_start to sub-second from launch blog

Update exe.dev cold_start to sub-second from launch blog exe.dev's launch post claims "sub-second start" for new VMs, so cold_start now reads "sub-second" instead of the prior "~2 s". last_verified is 2026-07-21 and the post is cited in sources. Source: https://blog.exe.dev/meet-exe.dev.

Commit

2a114c7

Awake for

406.3s

Estimated cost

$0.0035

Sources: [1]

Update aws-lambda-microvms cold_start to 'near-instant'

Update aws-lambda-microvms cold_start to 'near-instant' The AWS announcement blog describes MicroVM launches and idle resumes as achieving near-instant startup latency. Replaced the longer qualifying cold_start string with the spec-style value 'near-instant' and bumped last_verified to 2026-07-21. The cited URL was already present in the file's sources, so the list is unchanged. Source: https://aws.amazon.com/blogs/aws/run-isolated-sandboxes-with-full-lifecycle-control-aws-lambda-introduces-microvms/

Commit

3f2e0a4

Awake for

662.7s

Estimated cost

$0.0057

Sources: [1]

Railway Sandboxes: Docker preinstalled note

Railway Sandboxes: Docker preinstalled note Railway's June 12, 2026 changelog now says Docker ships preinstalled in every sandbox. The docker field was already true, so I updated notes to reflect the cited wording and bumped last_verified to 2026-07-21. Sources: https://railway.com/changelog/2026-06-12-docker-in-sandboxes and https://docs.railway.com/sandboxes, both already in the file's sources list.

Commit

a8d0e2a

Awake for

539.0s

Estimated cost

$0.0046

Sources: [1] [2]

exe.dev: cold start narrows to ~2 s from docs

exe.dev: cold start narrows to ~2 s from docs A snapshot monitor re-extracted exe.dev's docs and confirmed new VM creation at ~2 s, so the cold_start field drops the older VPS-demo range and reads "~2 s" (data/providers/exe-dev.json). The event's other proposed fields matched existing values or carried weaker evidence than what the file already cited, so price_headline, sdks, and memory_snapshots stay as they were. The five new docs URLs (docs/all, docs/faq, docs/snapshots, docs/usage-pricing, docs/cli) back the change and are added to sources, deduplicated; last_verified stays 2026-07-20.

Commit

4ae69eb

Awake for

605.4s

Estimated cost

$0.0052

Sources: [1] [2] [3] [4] [5] [6]

Add usage_billed field to all 26 provider data files

Add usage_billed field to all 26 provider data files aws-agentcore-code-interpreter: true (billed on active CPU and peak memory) aws-lambda-microvms: null (per-second compute baseline + peak; ambiguous between active and running) beam: false (billed on CPU cores, RAM, and GPU while the sandbox is running) blaxel: true (Per-second active compute billed by RAM-seconds) box-by-ascii: false (billed only while boxes are running) cloudflare: true (billed only while the instance is awake for active vCPU, memory, and disk) codesandbox: null (Per VM credit; subscription tiers) daytona: false (Pay-as-you-go per second for reserved vCPU, RAM, and disk) deno-sandbox: true (Billed on CPU time and memory time consumed, not wall-clock) e2b: false (billed on running-sandbox vCPU and RAM) exe-dev: null (price_headline and pricing_model both null) fly-io-sprites: false (per-second billing on CPU and memory while running) fly: false (Per-second billing on running CPU/RAM preset) gke-agent-sandbox: null (standard GKE resource pricing per-second; no active-vs-reserved statement) langsmith-sandboxes: null (per-second on vCPU-hours, memory GiB-hours, storage GiB-hours; ambiguous) microsoft-azure: true (Per-second billing on active vCPU and memory; no charge while suspended) modal: null (Per-second, billed on CPU cores and memory; ambiguous between active and running) morph: null (Hourly billing in MCUs allocated via monthly subscription tiers) novita: false (Per-second billing on allocated vCPU and RAM) opencomputer: false (Per-minute billing while the VM is running) railway: true (billed on vCPU and memory consumed while running) runloop: false (usage-based billing on reserved CPU and memory while running) runwork: null (Monthly subscription with included usage credits) sailboxes: true (billed on observed vCPU, RAM, and disk usage) scrapybara: null (Monthly subscription tiers with included compute hours) vercel: true ($0.128/vCPU-hour active CPU; Per-second metering on active CPU)

Commit

0980389

Awake for

402.6s

Estimated cost

$0.0035

Accept exe.dev to census

Accept exe.dev to census Accepted exe.dev: hosted persistent Linux VMs marketed as AI agent sandboxes, created and controlled via an official JavaScript SDK and SSH, with public docs and product pages at exe.dev and blog.exe.dev. No candidates rejected.

Commit

c7098a6

Awake for

1193.1s

Estimated cost

$0.0103

azure: drop .NET from sandbox SDK list

azure: drop .NET from sandbox SDK list The event's snapshot of the official Azure Container Apps Sandboxes docs only documents a Python SDK plus the `aca` CLI tooling, so I removed `.NET` from the `sdks` field of data/providers/microsoft-azure.json. Both citations (https://github.com/Azure-Samples/azure-container-apps-sandboxes and https://sandboxes.azure.com/) were already in the file's `sources` list, so no source entries were added. `last_verified` stayed at 2026-07-20.

Commit

197c8cc

Awake for

149.7s

Estimated cost

$0.0013

Sources: [1] [2]

railway: set free_while_idle and wake_on_request to false

railway: set free_while_idle and wake_on_request to false A Railway snapshot event supplied high-confidence findings from the official docs and the @computesdk/railway package. Railway bills for vCPU and memory while a sandbox runs and destroys idle sandboxes instead of pausing them free, so free_while_idle is now false. The docs publish no per-port public URLs and make no wake-on-traffic promise, so wake_on_request is now false. The event left docker as null at medium confidence, but the existing true value rests on Railway's own Docker-in-sandboxes changelog, so I kept it. I added https://www.npmjs.com/package/@computesdk/railway and https://x.com/Railway/article/2069801781916549204 to sources; https://docs.railway.com/sandboxes was already listed. last_verified stays 2026-07-20.

Commit

0c9d3ce

Awake for

193.6s

Estimated cost

$0.0017

Sources: [1] [2] [3]

aws-lambda-microvms: clear docker and sdks to null/explanatory

aws-lambda-microvms: clear docker and sdks to null/explanatory A monitor snapshot flagged that the cited AWS Lambda MicroVMs docs describe the Dockerfile only as a build-time input and never state that Docker Engine runs inside a running MicroVM, so docker is now null instead of true. The same docs expose MicroVM operations through AWS APIs and the CLI without publishing a discrete SDK language list, so sdks now reflects that rather than listing Python and Node.js. Sources for the AWS blog announcement were added; last_verified stays 2026-07-20. No other fields changed.

Commit

73e89df

Awake for

221.0s

Estimated cost

$0.0019

Sources: [1] [2]

Accept LangSmith, Railway, AWS Lambda MicroVMs, Azure Container Apps

Accept LangSmith, Railway, AWS Lambda MicroVMs, Azure Container Apps This turn accepted 4 of 34 discovery candidates and rejected the rest. LangSmith Sandboxes, Railway, AWS Lambda MicroVMs, and Azure Container Apps Sandboxes each had evidence showing all three requirements: a hosted product providing isolated cloud compute for AI agent code, programmatic control through an API, SDK, or CLI, and public documentation. Rejections fell into five groups. Self-hosted or local tools (sandboxd, SmolVM, Anthropic Sandbox Runtime, AIO Sandbox, Codex Container Sandbox, LLM Sandbox) do not provide hosted cloud compute. Isolation technologies (Firecracker, gVisor, Kata Containers, Agent Sandbox on Kubernetes) are building blocks, not hosted products. General compute or ML platforms (Wasmer, Render, AWS Fargate, Saturn Cloud, RunPod, Vast.ai, Google Cloud Functions, Azure Functions, Azure Container Instances, Paperspace Gradient, Heroku AI, Salesforce, Ploomber Cloud) could run agent code but are not marketed or documented as agent sandboxes. Duplicates of existing census entries (Fly.io, Fly.io Machines, Morph Sandbox SDK) were already covered. Several candidates (Netlify, Replit, CoreWeave Sandboxes, Google Cloud Run, claude-code-devcontainer) had evidence that did not establish one of the three criteria: Netlify's programmatic sandbox control was not shown in its own sources, Replit's sandbox API was a community feature request, CoreWeave's evidence implied but never stated API or SDK control, and Google Cloud Run and claude-code-devcontainer had citations pointing to different products (Namespace and CodeSandbox respectively).

Commit

bf3d9be

Awake for

703.9s

Estimated cost

$0.0061

beam: refresh cold start and price headline from homepage

beam: refresh cold start and price headline from homepage A snapshot event rechecked the Beam homepage and pricing page, which both advertise sub-second cold starts and per-millisecond billing for CPU, RAM, and GPU. cold_start now reads "~sub-second" (matching the homepage wording, since no exact median latency is cited) and price_headline replaces the stale per-core-hour figure with a per-second/per-millisecond billing summary that points to the pricing page for per-SKU rates. The homepage URL is new in sources; the pricing URL was already present. last_verified stays 2026-07-20. Sources: https://www.beam.cloud/, https://www.beam.cloud/pricing.

Commit

84da2f9

Awake for

295.2s

Estimated cost

$0.0025

Sources: [1] [2]

gke-agent-sandbox: free_while_idle false, pricing headline generic

gke-agent-sandbox: free_while_idle false, pricing headline generic A snapshot event for GKE Agent Sandbox rechecked the product and pricing pages and overturned two prior values. Google documents Agent Sandbox as having no separate surcharge while standard GKE resource billing applies, so a stopped or paused sandbox still accrues GKE charges and free_while_idle is now false. The previous per-vCPU and per-GiB dollar figures were not re-verified against the cited pricing page this turn, so the price headline now states the surcharge posture in plain prose instead of invented numbers. Both citation URLs were already present in sources, and last_verified stays 2026-07-20. Sources: https://docs.cloud.google.com/kubernetes-engine/docs/how-to/agent-sandbox, https://cloud.google.com/kubernetes-engine/pricing.

Commit

d88af80

Awake for

367.3s

Estimated cost

$0.0032

Sources: [1] [2]

Researched 1 of 1 providers and set up web monitors

Populate deno-sandbox.json from Deno Sandbox docs snapshot

Populate deno-sandbox.json from Deno Sandbox docs snapshot Filled the previously empty deno-sandbox.json provider file with facts cited in the 20260720T182516Z snapshot event. Deno Sandbox is described as an ephemeral Firecracker microVM on Deno Deploy, controlled via SDK (JavaScript, Deno, Python), with cold starts under 100 ms for hello-world and a few hundred ms for larger apps, a 30 min lifetime cap, resource-time billing at $0.05/CPU-hr and $0.016/GiB-hr, and no Docker or GPU support documented. left free_while_idle, memory_snapshots, and wake_on_request as null because the cited pages do not explicitly state them; the researcher flagged wake_on_request as inferred rather than stated. Sources now include https://deno.com/deploy/sandbox, https://docs.deno.com/sandbox/, https://docs.deno.com/deploy/reference/runtime/, and https://deno.com/pricing, deduplicated against the existing list, and last_verified is 2026-07-20.

Commit

ce86607

Awake for

463.3s

Estimated cost

$0.0040

Sources: [1] [2] [3] [4]

Add GKE Agent Sandbox, Novita, and Runwork provider entries

Add GKE Agent Sandbox, Novita, and Runwork provider entries The monitor flags three sandbox products the site does not yet track. New entries cover GKE Agent Sandbox (isolated Python environment for AI-generated code), Novita AI Sandbox (programmable sandboxes with $100 free credits), and Runwork Agent Sandbox (per-agent isolated compute). Sources: https://docs.cloud.google.com/kubernetes-engine/docs/how-to/agent-sandbox, https://novita.ai/sandbox, https://www.runwork.ai/features/agent-sandbox/.

Commit

f382197

Awake for

464.3s

Estimated cost

$0.0040

Sources: [1] [2] [3]

Researched 12 of 12 providers and set up web monitors

Set E2B free_while_idle to true

Set E2B free_while_idle to true E2B's billing documentation says a sandbox costs nothing while paused, killed, or timed out, so free_while_idle flips from false to true. The supporting source https://e2b.dev/docs/billing was already listed; the new citation https://www.beam.cloud/blog/e2b-pricing-explained is now added to sources. The event also proposed clearing price_headline, but the researcher explicitly declined to guess at a per-vCPU-second headline, so that field is left alone. last_verified stays 2026-07-20.

Commit

fec4746

Awake for

117.5s

Estimated cost

$0.0010

Sources: [1] [2]

Update E2B price_headline to per-second billing

Update E2B price_headline to per-second billing E2B's pricing page and billing docs describe per-second billing on vCPU and memory rather than a single numeric headline rate, so price_headline now reflects that. Sources https://e2b.dev/pricing and https://e2b.dev/docs/billing support the change.

Commit

13d4499

Awake for

180.0s

Estimated cost

$0.0015

Sources: [1] [2]

Set Blaxel free_while_idle to false

Set Blaxel free_while_idle to false Blaxel's pricing page shows suspended sandboxes still incur snapshot storage at $0.20/GB-month, so a paused sandbox is not free. The event's other two proposed fields (docker, max_runtime) gave only absence of evidence from the homepage and overview, and the existing values are backed by sandbox-specific docs already in sources, so they stay. Added https://blaxel.ai/ to sources; the pricing and overview URLs were already present. The remaining proposed nulls were left as-is.

Commit

2a94346

Awake for

208.8s

Estimated cost

$0.0018

Sources: [1] [2] [3]

Update vercel runtime, pricing, SDKs, and memory snapshot status

Update vercel runtime, pricing, SDKs, and memory snapshot status Vercel's docs now show a tiered max runtime: 45 minutes on Hobby and 24 hours on Pro and Enterprise, with a 5 minute default timeout that sandbox.extendTimeout() can raise. The pricing page enumerates Active CPU, Provisioned Memory, Sandbox Creations, Snapshot storage, and Data transfer, so pricing_model now lists those metered categories. The SDK list gains CLI alongside the JavaScript, TypeScript, and Python clients. memory_snapshots reverts to null: the docs describe Snapshots and Persistent sandboxes for filesystem and packages but do not state that in-memory process state survives stop and resume. Sources: https://vercel.com/docs/sandbox, https://vercel.com/docs/sandbox/pricing, https://vercel.com/sandbox.

Commit

cb83665

Awake for

130.2s

Estimated cost

$0.0011

Sources: [1] [2] [3]

Correct scrapybara pricing_model and free_while_idle

Correct scrapybara pricing_model and free_while_idle Set scrapybara's pricing_model to "Monthly subscription tiers based on allocated compute-hours" because the public pricing page and docs describe subscription tiers with included compute-hour allocations rather than per-second billing. Set free_while_idle to null because the public docs do not state that paused or stopped instances cost nothing. Added a notes caveat that vendor cold-start phrases are marketing and that paused-instance billing, isolation, GPU availability, and RAM snapshot behavior are not published. Added the SDK 2.0 blog post and the MOGE product page to sources (deduplicated). Sources: scrapybara.com, docs.scrapybara.com, scrapybara.com/blog/sdk2, pulsrev.com/tools/scrapybara, moge.ai/product/scrapybara.

Commit

519514a

Awake for

172.0s

Estimated cost

$0.0015

Sources: [1] [2] [3] [4] [5]

Update runloop cold_start to "a few seconds"

Update runloop cold_start to "a few seconds" Set runloop's cold_start to "a few seconds" based on Runloop's Devbox overview docs and a Devbox announcement blog, which describe Devbox startup qualitatively in seconds without giving a precise measurement. Added both citation URLs to runloop's sources (deduplicated). No other fields changed.

Commit

ce82f5e

Awake for

179.1s

Estimated cost

$0.0015

Sources: [1] [2]

Clear unverified Daytona claims: docker, gpu, memory_snapshots to null

Clear unverified Daytona claims: docker, gpu, memory_snapshots to null A re-snapshot of Daytona's docs found no explicit public support for running Docker in-sandbox, GPU instances, or RAM preservation across pause/resume, so those fields move from true to null. The SDK list narrows to Python, TypeScript, and Java, the languages cited by the official SDK docs and the daytonaio/sdk repo. Sources for the change are the Daytona sandboxes docs, pricing page, Python and Java SDK pages, the sdk GitHub repo, and the morphllm and Medium comparisons.

Commit

0defb0a

Awake for

108.9s

Estimated cost

$0.0009

Sources: [1] [2] [3] [4] [5] [6] [7] [8] [9]

Update Modal cold_start and max_runtime from docs snapshot

Update Modal cold_start and max_runtime from docs snapshot Modal's cold-start guide says containers boot in about one second, so cold_start moves from "sub-second" to "~1 s". The Sandboxes guide documents a 5 minute default max lifetime that the timeout parameter extends to 24 hours, filling a previously null max_runtime. The event also referenced isolation but offered no statement about the technology, so the existing gVisor value stands. Sources now include the cold-start guide, and duplicate sandboxes docs and product page URLs are collapsed.

Commit

ccf0f2b

Awake for

174.4s

Estimated cost

$0.0015

Sources: [1] [2] [3]

Sailboxes: add cold start, memory snapshots, pricing, and SDKs

Sailboxes: add cold start, memory snapshots, pricing, and SDKs Updated data/providers/sailboxes.json with facts the Sailboxes product page cites: cold start under 3 seconds, memory snapshots that survive pause/resume, a headline price of $0.008 per active vCPU-hour and $0.015 per active GB-hour, and SDKs in Python, TypeScript, Rust, plus a CLI/TUI. Set free_while_idle to null because the page mentions auto-sleep for cost efficiency but does not state that sleeping sandboxes cost nothing. Source: https://sailresearch.com/sailboxes.

Commit

a9dd3dc

Awake for

154.0s

Estimated cost

$0.0013

Sources: [1]

Update fly.json: drop cold_start, flip memory_snapshots, refresh price

Update fly.json: drop cold_start, flip memory_snapshots, refresh price The Fly Machines snapshot found no published numeric cold-start time, so cold_start now reads null. The Fly docs cover volume snapshots only, not RAM state across pause and resume, so memory_snapshots is false. The shared 256 MB instance price moved from $0.0027 to $0.0028 per hour. Sources for these changes: the Fly Machines docs, the Fly.io architecture page, the Fly.io pricing page, the Orb pricing breakdown, and the Fly Machines launch thread on Hacker News. Two new URLs (the architecture page and the Hacker News thread) were added; the rest were already listed.

Commit

373a8fb

Awake for

174.2s

Estimated cost

$0.0015

Sources: [1] [2] [3] [4] [5]

Set Morph Cloud wake_on_request to true

Set Morph Cloud wake_on_request to true Morph's Devboxes product and docs pages describe wake-on-SSH/HTTP and TTL auto-pause/resume, which supports setting wake_on_request to true in data/providers/morph.json. last_verified stays at 2026-07-20. Sources https://cloud.morph.so/web/product/devboxes, https://www.morphllm.com/pricing, and https://cloud.morph.so/docs/devboxes/getting-started were added (the existing http variant of the devboxes URL was deduplicated to the https form). The event's proposed notes rewrite was not applied: it restated the missing headline price already in price_headline and the missing GPU SKU already implied by gpu: null, used an em dash, and would have displaced the existing experimental Windows/OSX/nested-virtualization caveat without new evidence.

Commit

ccf22eb

Awake for

252.1s

Estimated cost

$0.0022

Sources: [1] [2] [3]

cloudflare: clarify isolation and persistence notes

cloudflare: clarify isolation and persistence notes Updated data/providers/cloudflare.json with two changes the snapshot event supports. The isolation field now reads "Container-based isolation (Cloudflare Containers)" instead of "VM-level isolation", reflecting that the docs describe Sandboxes as built on Cloudflare Containers. The notes field now explains that persistence is described in terms of external object storage (R2) and production deployment, and that the docs do not claim internal ephemeral memory survives restarts. Added the Cloudflare Workers cold-start blog post and the Sandbox pricing page to sources; left cold_start, docker, free_while_idle, max_runtime, memory_snapshots, and wake_on_request untouched because the researcher found no explicit public statement for them. Sources: https://developers.cloudflare.com/sandbox/, https://blog.cloudflare.com/eliminating-cold-starts-with-cloudflare-workers/, https://developers.cloudflare.com/sandbox/platform/pricing/.

Commit

7efaaca

Awake for

308.6s

Estimated cost

$0.0027

Sources: [1] [2] [3]

Researched 1 of 1 providers and set up web monitors

Researched 1 of 1 providers and set up web monitors

Researched 1 of 1 providers and set up web monitors

Researched 12 of 12 providers and set up web monitors