Benchmarks

Every number has a source.

We publish a number only with its provenance, where it was measured from, which side measured it, the date, the sample count and a link to the raw data. Each B3OS number also states the transport, the regime and its concurrency or arrival rate, and the conditions it ran under. Competitors are measured the way their own benchmarks measure them.

36 published B3OS numbers and 268 numbers cited from other providers, 166 of them in like-for-like tables, inputs up to Oct 26, 2026. 4 cited figures have a newer figure from the same source and are left out.

The loop

One agent task, three places to run it

Each task is N turns. A turn is one streamed infer.chat call on a named model with fixed prompt and token counts, plus one tool call that rotates among sandboxes.exec, vision.scrape (a local fixture and one real site) and a browser-session act. At least 30 tasks per configuration, p50 and p95 per turn and per task, with the share of each turn spent in transport rather than compute. Raw JSONL for every run.

AAgent in a sandbox on the node
The harness runs inside a B3OS sandbox and calls the co-located model and tools over vsock.
BAgent in AWS us-west-2, calling B3OS
The same harness runs in us-west-2 and calls B3OS through beta-api.b3os.org.
CAgent in AWS us-west-2, calling other providers
The same harness in us-west-2 calls the same open model on an OpenAI-compatible host, and E2B, Firecrawl and Browserbase through their own SDKs.

Configuration	Per turn	Per task	Model call	Tool call	Transport per turn
AAgent in a sandbox on the nodesame-host	measuring	measuring	measuring	measuring	measuring
BAgent in AWS us-west-2, calling B3OScross-region from us-west-2	measuring	measuring	measuring	measuring	measuring
CAgent in AWS us-west-2, calling other providerscross-region from us-west-2	measuring	measuring	measuring	measuring	measuring

Per-hop multipliers are never shown without the per-turn numbers behind them. Configuration A is the only one where the model, the tools and the agent share a machine; frontier models served elsewhere get co-located tools but not a co-located model.

vision

Measuring

Scrape from cache

A scrape answered from the shared cache, server side.

Measuringp50, server side

Plain-fetch overhead

Time added over the origin's time to last byte, for HTML up to 300 KB at a stated concurrency. Extraction is timed separately.

Measuringp50, server side

Browser scrape

Navigate, wait for ready, read HTML and convert to Markdown, against a recorded (HAR replay) origin on a named corpus.

Measuringp50, server side

Browser context claim

Claiming a pre-created context from the scrape pool, with the pool depth stated.

Measuringp50, server side

Browser context claim, p99

The same claim at the 99th percentile.

Measuringp99, server side

Session browser claim

Creating a browser session from the pool of whole browsers.

Measuringp50, server side

Act overhead

Time per action in a batched act call, beyond the browser's own work.

Measuringp50, server side

What others publish

Each table holds numbers measured the same way, from the same side and kind of vantage point, best first. B3OS joins a table only with a number measured that way too.

10-action Wikipedia flow, actions per second

client side · client in the provider's region · actions/s · descending (higher is better)

Competitor-runp50 · by Notte (runs Browser Arena and competes in it); raw JSONL re-analyzed by the B3OS fact-check · AWS us-east-1 client (t3.medium), provider in-region; Notte and Lightpanda from us-west-1 (t3.small) · measured 2026-10-25Arena hello-browser: create session via provider SDK, Playwright connectOverCDP, goto example.com (domcontentloaded), release; percentiles over successful runs only

Who	Value	Provenance and conditions
Tilion	5.26–5.34actions/s	n not statedsource1.87-1.90 s per 10-action task
Browser Use	4.3actions/s	n not statedsource4.29 in the newest three runs
Kernel	4.2–4.4actions/s	n not statedsourceheadful 4.27, headless 4.37 in the newest three runs
Browserbase	4.15actions/s	n not statedsourcenewest three runs (Oct 21, 22, 25)
AWS Bedrock AgentCore Browser	1.17actions/s	n not statedsource66% success

Page.captureScreenshot over cloud CDP

client side · client in the provider's region · ms · ascending (lower is better)

Who	Value	Provenance and conditions
Browserbase	162–166ms	Competitor-runp50by Notte (runs Browser Arena and competes in it); raw JSONL re-analyzed by the B3OS fact-checkAWS us-east-1 client (t3.medium), provider in-region; Notte and Lightpanda from us-west-1 (t3.small)measured 2026-10-25n not statedsource166 ms in the report's window, 162 ms over the newest three runs. Arena hello-browser: create session via provider SDK, Playwright connectOverCDP, goto example.com (domcontentloaded), release; percentiles over successful runs only

session create p95 under bursts of 10 concurrent creates

client side · client in the provider's region · ms · ascending (lower is better)

Competitor-runp95 · by Notte (runs Browser Arena and competes in it); raw JSONL re-analyzed by the B3OS fact-check · AWS us-east-1 client (t3.medium), provider in-region; Notte and Lightpanda from us-west-1 (t3.small) · measured 2026-10-22Arena hello-browser: create session via provider SDK, Playwright connectOverCDP, goto example.com (domcontentloaded), release; percentiles over successful runs only

Who	Value	Provenance and conditions
Browser Use	273–295ms	n not statedsourcec10 runs, Oct 16-22
Hyperbrowser	417ms	n not statedsourcec10 runs, Oct 16-22
Notte	559ms	n not statedsourcec10 runs, Oct 16-22
Cloudflare Browser Run	564–588ms	n not statedsourcec10 runs, Oct 16-22
Kernel	2,500–3,000ms	n not statedsourcec10 runs, Oct 16-22
Tilion	12,600ms	n not statedsourcec10 runs, Oct 16-22
Browserbase	54,400ms	n not statedsourcec10 runs, Oct 16-22

session create (API return)

client side · client in the provider's region · ms · ascending (lower is better)

Competitor-runp50 · by Notte (runs Browser Arena and competes in it); raw JSONL re-analyzed by the B3OS fact-check · AWS us-east-1 client (t3.medium), provider in-region; Notte and Lightpanda from us-west-1 (t3.small) · measured 2026-10-22Arena hello-browser: create session via provider SDK, Playwright connectOverCDP, goto example.com (domcontentloaded), release; percentiles over successful runs only

Who	Value	Provenance and conditions
Tilion (waitlist beta)	18ms	n=700sourceOct 16-22 window, sequential
Kernel headless (1 GiB)	20ms	n not statedsourcelatest two days, Oct 21-22, 100% success
Kernel headful (8 GB)	23ms	n not statedsourcelatest two days, Oct 21-22, 100% success
Kernel headless	24ms	n not statedsource30 days, Sep 25-Oct 22, 96.3% success

create + CDP connect

client side · client in the provider's region · ms · ascending (lower is better)

Competitor-runp50 · by Notte (runs Browser Arena and competes in it); raw JSONL re-analyzed by the B3OS fact-check · AWS us-east-1 client (t3.medium), provider in-region; Notte and Lightpanda from us-west-1 (t3.small) · measured 2026-10-22Arena hello-browser: create session via provider SDK, Playwright connectOverCDP, goto example.com (domcontentloaded), release; percentiles over successful runs only

Who	Value	Provenance and conditions
Tilion	79–90ms	n not statedsourcerange over the fact-check windows
Kernel headful	86–111ms	n not statedsourcerange over the fact-check windows

create + connect + goto(example.com)

client side · client in the provider's region · ms · ascending (lower is better)

Competitor-runp50 · by Notte (runs Browser Arena and competes in it); raw JSONL re-analyzed by the B3OS fact-check · AWS us-east-1 client (t3.medium), provider in-region; Notte and Lightpanda from us-west-1 (t3.small) · measured 2026-10-22Arena hello-browser: create session via provider SDK, Playwright connectOverCDP, goto example.com (domcontentloaded), release; percentiles over successful runs only

Who	Value	Provenance and conditions
Tilion	122–137ms	n not statedsourceOct 16-22; p99 13.0 s
Notte	265ms	n not statedsourceOct 16-22; us-west-1 client (t3.small)
Kernel	267–283ms	n not statedsourceSOTA-bar range; the report's 272-336 / 638-684 mixed headless and headful
Browser Use	385–405ms	n not statedsourceOct 16-22
Browserbase	463–483ms	n not statedsourceOct 16-22; Browserbase's p50 fell from ~780-830 ms to ~450-530 ms from Oct 16 (provider-side change)
Cloudflare Browser Run	810ms	n not statedsourceOct 16-22
Steel	1,319ms	n not statedsourceOct 16-22
Anchor Browser	1,950ms	n not statedsourceOct 16-22

average connect / average navigation to example.com

client side · vendor figure, method not stated · ms · ascending (lower is better)

Competitor-runmean_connect · by Browserless (a hosted-browser vendor measuring its competitors) · not stated · published 2026-01none recorded beyond the metric and vantage (fact-check research-vision.md A15)

Who	Value	Provenance and conditions
Hyperbrowser	692ms	n not statedsource
Browserless	936ms	n not statedsource
Browserbase	1,930ms	n not statedsource
Anchor Browser	5,582ms	n not statedsource

browser lifecycle total per session

client side · client, location not stated · ms · ascending (lower is better)

Independenttotal · by ComputeSDK · not disclosed (region, run count and date undisclosed) · retrieved 2026-10-25

Who	Value	Provenance and conditions
Tilion	195ms	n not statedsourcephase breakdown: create 14 / connect 98 / navigate 60 ms
Kernel	242ms	n not statedsourcephase breakdown: create 9 / connect 33 / navigate 139 ms
Browser Use	320ms	n not statedsourcenone recorded beyond the metric and vantage (fact-check research-vision.md A13)
Browserbase	622ms	n not statedsourcenone recorded beyond the metric and vantage (fact-check research-vision.md A13)
Hyperbrowser	701ms	n not statedsourcenone recorded beyond the metric and vantage (fact-check research-vision.md A13)
Steel	1,309ms	n not statedsourcenone recorded beyond the metric and vantage (fact-check research-vision.md A13)
Notte	1,750ms	n not statedsourcenone recorded beyond the metric and vantage (fact-check research-vision.md A13)

browserbench: create / end to end (averages)

client side · vendor figure, method not stated · ms · ascending (lower is better)

Vendor claimmean_create · by Kernel (write-up of a browserbench run) · not stated in the write-up · measured 2026-01none recorded beyond the metric and vantage (fact-check research-vision.md A12)

Who	Value	Provenance and conditions
Kernel	37ms	n=5,000sourceaverages, 5,000 runs
Steel	182ms	n=5,000sourceaverages, 5,000 runs
Browserbase	213ms	n=5,000sourceaverages, 5,000 runs
Anchor Browser	3,797ms	n=5,000sourceaverages, 5,000 runs

live single-URL scrape to markdown (maxAge 0)

client side · client, location not stated · ms · ascending (lower is better)

Who	Value	Provenance and conditions
Firecrawl	2,829ms	Independentp50by techstackups (Ritza; no sponsorship disclosed)client location not statedupdated 2026-10-20n=821sourcenone recorded beyond the metric and vantage (fact-check research-vision.md B24)

create, connect, goto google.com, release: total

client side · client in the provider's region · ms · ascending (lower is better)

Competitor-runp50 · by Steel (steel-dev/browserbench) · AWS us-east-1 client · published 2025-11-07none recorded beyond the metric and vantage (fact-check research-vision.md A12)

Who	Value	Provenance and conditions
Steel	867ms	n=5,000source5,000 runs per provider
Kernel	1,485ms	n=5,000source5,000 runs per provider
Browserbase	1,669ms	n=5,000source5,000 runs per provider
Hyperbrowser	3,665ms	n=5,000source5,000 runs per provider
Anchor Browser	7,919ms	n=5,000source5,000 runs per provider

end-to-end session (average)

client side · client in another region · ms · ascending (lower is better)

Independentmean · by techstackups (Ritza) · client in Hetzner Helsinki (transatlantic RTT included) · updated 2026-05-31

Who	Value	Provenance and conditions
Steel	1,441ms	n=10sourceaverage of 10 measured sessions. none recorded beyond the metric and vantage (fact-check research-vision.md A14)
Kernel	1,554ms	n=10sourceaverage of 10 measured sessions. none recorded beyond the metric and vantage (fact-check research-vision.md A14)
Browserless	2,090ms	n=10sourceaverage of 10 measured sessions. none recorded beyond the metric and vantage (fact-check research-vision.md A14)
Anchor Browser	3,730ms	n=10sourceaverage of 10 measured sessions. none recorded beyond the metric and vantage (fact-check research-vision.md A14)
Hyperbrowser	4,012ms	n=10sourceaverage of 10 measured sessions. none recorded beyond the metric and vantage (fact-check research-vision.md A14)
Browserbase	11,933ms	n=10sourceaverage of 10 measured sessions. 9,401 ms of it was create

32 more cited numbers for vision

Published by vendors and third parties, each measured its own way. They are listed for reference and not compared with each other or with B3OS.

Who	Value	What, provenance and conditions
Tavily	1–1.6USD/1k pages	price per 1,000 pagesVendor claimlistby Tavilylist priceclient sideretrieved 2026-10-25n not statedsource
Steel	5USD/1k pages	price per 1,000 pagesVendor claim/scrapeby Steellist priceclient sideretrieved 2026-10-25n not statedsource
Spider	0.1–0.3USD/1k pages	price per 1,000 pagesVendor claimderivedby Spiderlist priceclient sideretrieved 2026-10-25n not statedsource
Hyperbrowser	1USD/1k pages	price per 1,000 pagesVendor claimlistby Hyperbrowserlist priceclient sideretrieved 2026-10-25n not statedsource
Firecrawl	0.75USD/1k pages	price per 1,000 pagesVendor claimScale monthlyby Firecrawllist priceclient sideretrieved 2026-10-25n not statedsource
Exa	1USD/1k pages	price per 1,000 pagesVendor claimper content typeby Exalist priceclient sideretrieved 2026-10-25n not statedsource
Browserbase Fetch	0.5USD/1k pages	price per 1,000 pagesVendor claimStartupby Browserbase Fetchlist priceclient sideretrieved 2026-10-25n not statedsource
Steel	0.1USD/h	browser-hour list priceVendor claimlistby Steellist priceclient sideretrieved 2026-10-25n not statedsource
Notte	0.01USD/h	browser-hour list priceVendor claimlistby Nottelist priceclient sideretrieved 2026-10-25n not statedsource
Kernel	0.06USD/h	browser-hour list priceVendor claimlistby Kernellist priceclient sideretrieved 2026-10-25n not statedsource
Hyperbrowser	0.1USD/h	browser-hour list priceVendor claimlistby Hyperbrowserlist priceclient sideretrieved 2026-10-25n not statedsource
Cloudflare Browser Run	0.09USD/h	browser-hour list priceVendor claimlistby Cloudflare Browser Runlist priceclient sideretrieved 2026-10-25n not statedsource
Browserbase	0.12USD/h	browser-hour list priceVendor claimlistby Browserbaselist priceclient sideretrieved 2026-10-25n not statedsource
Browser Use	0.02USD/h	browser-hour list priceVendor claimlistby Browser Uselist priceclient sideretrieved 2026-10-25n not statedsource
Anchor Browser	0.09USD/h	browser-hour list priceVendor claimlistby Anchor Browserlist priceclient sideretrieved 2026-10-25n not statedsource
Spider	1,538concurrent	concurrency and creation limitsVendor claim$19,994/moby Spiderlist priceclient sideretrieved 2026-10-25n not statedsource
Kernel	150concurrent	concurrency and creation limitsVendor claimtop self-serve planby Kernellist priceclient sideretrieved 2026-10-25n not statedsource
Firecrawl	100concurrent	concurrency and creation limitsVendor claimScaleby Firecrawllist priceclient sideretrieved 2026-10-25n not statedsource
Cloudflare Browser Run	200concurrent	concurrency and creation limitsVendor claimPaidby Cloudflare Browser Runlist priceclient sideretrieved 2026-10-25n not statedsource
Browserbase	150creates/min	concurrency and creation limitsVendor claimScale (Free 5, Developer 25, Startup 50)by Browserbaselist priceclient sideretrieved 2026-10-25n not statedsource
Browser Use	1,000concurrent	concurrency and creation limitsVendor claimafter $25k cumulative spendby Browser Uselist priceclient sideretrieved 2026-10-25n not statedsource
Bright Data (Browser API)	2,000ms	startup time; success at 250 concurrent agentsAffiliatedapprox_startupby AIMultiple (Bright Data is a disclosed subscriber of its benchmarking services)not statedclient sideretrieved 2026-10-25n=20source
Airtop	4,000ms	startup time; successAffiliatedapprox_startupby AIMultiple (has disclosed commercial ties to tested vendors)not statedclient sideretrieved 2026-10-25n=20source
Lightpanda	4.5s	940 pages at 25 tabs: wall time, memory, CPUVendor claimwallby Lightpandamachine not statedserver sidepublished 2026-10-19n not statedsource
Spider	45ms	static time to first result; average pages/s over 1,000 URLsVendor claimttfr_staticby Spider (designed the test)c6i.4xlarge client in us-east-2client sideupdated 2026-09-02n=1,000source
Firecrawl	310ms	static time to first result; average pages/s (in Spider's test)Competitor-runttfr_staticby Spider (designed the test and competes with Firecrawl)c6i.4xlarge client in us-east-2client sideupdated 2026-09-02n=1,000source
Cloudflare (Browser Run: warm Chromium vs Kitesurf)	472ms	HTML extraction wall time (median of 5 runs over 14 URLs)Vendor claimp50_warm_chromiumby Cloudflare (Celso Martinho)Cloudflare internalserver sidepublished 2026-08-06n not statedsource
Browser Use cloud (Firecracker snapshot-resume)	825ms	session create in a 10,000-session stress testVendor claimp50by Browser Usethrough the public API (region not stated)client sidepublished 2026-06-15n not statedsource
rs-trafilatura	0.903F1	main-content extraction F1 on the WCXB held-out test set; time per pageAffiliatedtestby Murrough Foley (WCXB author, also wrote rs-trafilatura)AMD Ryzen, threading not statedserver sidepublished 2026-05n not statedsource
Kernel	30ms	browser create (vendor-reported)Vendor claimp50by Kernelunstated (Kernel browsers run in us-east only)client sidepublished 2026-04-24n not statedsource
Spider Browser	705ms	browser create p50 / p95; end to endVendor claimp50by Spidernot statedclient sidepublished 2026-02-17n=50source
Firecrawl	3,387ms	scrape p95 on 1,000 URLs (firecrawl/scrape-content-dataset-v1)Vendor claimp95by FirecrawlFirecrawl's internal runclient sidepublished 2026-01-13n=1,000source

render

Measuring

Styled screenshot

1600×900 at DPR 1, JPEG or WebP, quiet window included, server side.

Measuringp50, server side

Styled screenshot at DPR 2

The same capture at device pixel ratio 2.

Measuringp50, server side

Each extra variant

Time each additional encoding of the same capture adds.

Measuringp50, server side

10-second scripted demo

From request to finished MP4 on a HAR-cached app. The first recording pass is reported separately.

Measuringp50, server side

30-second narrated demo

With text-to-speech on the GPU slice; a CPU text-to-speech row is separate.

Measuringp50, server side

Encode throughput

Frames encoded per second across concurrent jobs, x264 veryfast CRF 20, on a named content mix within render's CPU budget.

Measuringthroughput, server side

screenshot total, PNG with a transparent background

Measuringp50, server side

video job total on the first (recording) pass, no HAR cache yet

Measuringp50, server side

narrated video job total, 30 s output, TTS on CPU

Measuringp50, server side

What others publish

Each table holds numbers measured the same way, from the same side and kind of vantage point, best first. B3OS joins a table only with a number measured that way too.

uncached screenshot, average over 7 URLs (DPR 2, full page, ad blocking)

client side · client, location not stated · ms · ascending (lower is better)

mean · by Microlink (a screenshot-API vendor measuring itself and its competitors) · single New York vantage point · measured 2026-0310 runs at different hours, slowest dropped

Who	Value	Provenance and conditions
Microlink	4,111.84ms	Vendor claimn not statedsource
Urlbox	7,334.22ms	Competitor-runn not statedsource
ScreenshotOne	7,711.14ms	Competitor-runn not statedsource
ApiFlash	9,463.2ms	Competitor-runn not statedsource

10 more cited numbers for render

Published by vendors and third parties, each measured its own way. They are listed for reference and not compared with each other or with B3OS.

Who	Value	What, provenance and conditions
Remotion Lambda	720s	2 h video, 1920x1080 at 30 fps (216,000 frames)Vendor claimwallby Remotion200 warm ARM64 Lambdas, us-east-1server sideretrieved 2026-10-25n not statedsource
Shotstack	0.05–0.2USD/min	render price per output minuteVendor claimpay-as-you-go to $499 planby Shotstacklist priceclient sideretrieved 2026-10-25n not statedsource
Remotion	0.01USD/render	Remotion automation licenceVendor claimAutomatorsby Remotionlist priceclient sideretrieved 2026-10-25n not statedsource
HeyGen (HyperFrames)	0.1USD/min	HTML-to-video render price per output minute, 1080p60Vendor claimEnterpriseby HeyGenlist priceclient sideretrieved 2026-10-25n not statedsource
Editframe	0.02USD/processing-min	cloud rendering per minute of server processingVendor claim≤2 MPby Editframelist priceclient sideretrieved 2026-10-25n not statedsource
Creatomate	0.38USD/min	render price per output minuteVendor claimEssential, about 14 credits/min at 720p25by Creatomatelist priceclient sideretrieved 2026-10-25n not statedsource
Microlink	49USD/month	screenshot API planVendor claimPro, 46,000 requests (~$1.07/1k)by Microlinklist priceclient sideretrieved 2026-10-25n not statedsource
Microlink	968ms	uncached screenshot of example.com (average)Vendor claimmeanby MicrolinkNew Yorkclient sidemeasured 2026-03n not statedsource
Remotion Lambda	15.52s	"1 minute" video: 1280x720 at 24 fps (1,440 frames), video-in-videoVendor claimcoldby RemotionAWS us-east-1, default concurrency (~72 Lambdas by the documented formula)server sideversion-date 2025-12-03n not statedsource
Remotion Lambda	56.09s	"10 minute HD": 1280x720 at 24 fps (14,315 frames)Vendor claimwarmby RemotionAWS us-east-1server sideversion-date 2025-12-03n not statedsource

sandboxes

Measuring

Claim to first exec

Create a sandbox and run its first command, end to end on the same host, at idle. Pool depth stated.

Measuringp50, client side

Claim to first exec, 100 at once

The same under a burst of 100 simultaneous claims.

Measuringp50, client side

Exec round trip

Run /bin/true in a running sandbox: b3osd to the guest agent, process spawn, exit status back.

Measuringp50, server side

Pause and resume

Each of pause and resume on a running sandbox.

Measuringp50, server side

Fork

One fork of a sandbox with up to 128 MiB of RAM and 64 MiB dirty. Larger sizes are published as a curve.

Measuringp50, server side

Time to interactive, ComputeSDK harness

Create a sandbox and run node -v, measured by ComputeSDK's own harness from their runners: the shape other sandbox providers are measured in.

Measuringp50, client side

What others publish

Each table holds numbers measured the same way, from the same side and kind of vantage point, best first. B3OS joins a table only with a number measured that way too.

DAX workload (install, typecheck, run) total

client side · client, location not stated · s · ascending (lower is better)

Independenttotal · by ComputeSDK (computesdk/benchmarks) · ComputeSDK runner (region undisclosed); client-timed · measured 2026-10-23TTI = create sandbox + run `node -v` (destroy untimed), via each provider's @computesdk package

Who	Value	Provenance and conditions
promigence	32.4s	n not statedsourcesingle run (2026-10-23)
isorun	34.3s	n not statedsourcesingle run (2026-10-23)
MIOSA	40.7s	n not statedsourcesingle run (2026-10-23)
Blaxel	49.8s	n not statedsourcesingle run (2026-10-23)
Vercel Sandbox	63.8s	n not statedsourcesingle run (2026-10-23)
Daytona	74.4s	n not statedsourcesingle run (2026-10-23)
Cloudflare Sandbox	149.6s	n not statedsourcesingle run (2026-10-23)
declaw	230.3s	n not statedsourcesingle run (2026-10-23)

TTI, 100 concurrent (burst)

client side · client, location not stated · ms · ascending (lower is better)

Independentp50 · by ComputeSDK (computesdk/benchmarks) · ComputeSDK runner (region undisclosed); client-timedTTI = create sandbox + run `node -v` (destroy untimed), via each provider's @computesdk package

Who	Value	Provenance and conditions
isorun	33.25ms	measured 2026-09-11n=100sourcebest full run ever (2026-09-11)
MIOSA	57.1ms	measured 2026-10-23n=100sourcebest run (2026-10-23)
MIOSA	57.1ms	measured 2026-10-23n=100sourcelatest run (2026-10-23)
isorun	74.9ms	measured 2026-10-23n=100sourcelatest run (2026-10-23)
isorun	74.9ms	measured 2026-10-23n=100sourcemedian of per-run medians (44 runs, 2026-02-21 to 2026-10-23)
promigence	91.8ms	measured 2026-10-23n=100sourcelatest run (2026-10-23)
declaw	139.1ms	measured 2026-06-17n=100sourcebest run (2026-06-17)
Northflank	158.3ms	measured 2026-07-17n=100sourcebest run (2026-07-17); typical 297.5 ms over 46 runs
MIOSA	192.7ms	measured 2026-10-23n=100sourcemedian of per-run medians (10 runs, 2026-02-21 to 2026-10-23)
Daytona	231.9ms	measured 2026-04-11n=100sourcebest run (2026-04-11)
Blaxel	335.1ms	measured 2026-07-31n=100sourcebest run (2026-07-31)
Vercel Sandbox	403.2ms	measured 2026-07-17n=100sourcebest run (2026-07-17)
declaw	430ms	measured 2026-10-23n=100sourcemedian of per-run medians (97 runs, 2026-02-21 to 2026-10-23)
Blaxel	441.8ms	measured 2026-10-23n=100sourcelatest run (2026-10-23)
E2B	442.7ms	measured 2026-07-11n=100sourcebest run (2026-07-11)
Daytona	453ms	measured 2026-10-23n=100sourcemedian of per-run medians (134 runs, 2026-02-21 to 2026-10-23)
Cloudflare Sandbox	536.3ms	measured 2026-10-02n=100sourcebest run (2026-10-02)
Modal	552.3ms	measured 2026-07-08n=100sourcebest run (2026-07-08)
E2B	639.7ms	measured 2026-10-23n=100sourcemedian of per-run medians (143 runs, 2026-02-21 to 2026-10-23)
Cloudflare Sandbox	647.84ms	measured 2026-09-25n=100sourcethe 2026-09-25 run Cloudflare cited after its rebuild
Cloudflare Sandbox	682.1ms	measured 2026-10-23n=100sourcelatest run (2026-10-23)
Vercel Sandbox	718.4ms	measured 2026-10-23n=100sourcelatest run (2026-10-23)
Vercel Sandbox	752.3ms	measured 2026-10-23n=100sourcemedian of per-run medians (139 runs, 2026-02-21 to 2026-10-23)
Freestyle	855.9ms	measured 2026-09-10n=100sourcebest run (2026-09-10)
declaw	989.9ms	measured 2026-10-23n=100sourcelatest run (2026-10-23)
Modal	1,070.6ms	measured 2026-10-23n=100sourcelatest run (2026-10-23)
Daytona	1,094.6ms	measured 2026-10-23n=100sourcelatest run (2026-10-23)
Blaxel	1,247.9ms	measured 2026-10-23n=100sourcemedian of per-run medians (118 runs, 2026-02-21 to 2026-10-23)
Freestyle	1,437.6ms	measured 2026-10-23n=100sourcemedian of per-run medians (11 runs, 2026-02-21 to 2026-10-23)
E2B	1,439.4ms	measured 2026-10-23n=100sourcelatest run (2026-10-23)
Modal	1,668.4ms	measured 2026-10-23n=100sourcemedian of per-run medians (142 runs, 2026-02-21 to 2026-10-23)
Cloudflare Sandbox	2,066.4ms	measured 2026-10-23n=100sourcemedian of per-run medians (124 runs, 2026-02-21 to 2026-10-23)
Freestyle	3,237.5ms	measured 2026-10-23n=100sourcelatest run (2026-10-23)
CodeSandbox	33,772ms	measured 2026-10-23n=100sourcelatest run (2026-10-23)

TTI, sequential

client side · client, location not stated · ms · ascending (lower is better)

Who	Value	Provenance and conditions
isorun	12.04ms	Supersededmeasured 2026-07-17n=100sourcebest run (2026-07-17)
declaw	13.2ms	Supersededmeasured 2026-08-07n=100sourcebest run (2026-08-07)
declaw	13.2ms	Supersededmeasured 2026-08-07n=100sourcelast sequential run (2026-08-07; mode retired)
isorun	16.8ms	Supersededmeasured 2026-08-07n=100sourcelast sequential run (2026-08-07; mode retired)
isorun	25ms	Supersededmeasured 2026-08-07n=100sourcemedian of per-run medians (30 runs)
declaw	45.7ms	Supersededmeasured 2026-08-07n=100sourcemedian of per-run medians (83 runs)
Northflank	64.5ms	Supersededmeasured 2026-07-17n=100sourcebest run (2026-07-17)
Northflank	82.8ms	Supersededmeasured 2026-08-07n=100sourcelast sequential run (2026-08-07; mode retired)
Daytona	94.2ms	Supersededmeasured 2026-03-14n=100sourcebest run (2026-03-14)
Northflank	152.7ms	Supersededmeasured 2026-08-07n=100sourcemedian of per-run medians (45 runs)
Blaxel	196.9ms	Supersededmeasured 2026-07-17n=100sourcebest run (2026-07-17)
Daytona	197ms	Supersededmeasured 2026-08-07n=100sourcemedian of per-run medians (130 runs)
E2B	233.3ms	Supersededmeasured 2026-03-11n=100sourcebest run (2026-03-11)
Blaxel	251.4ms	Supersededmeasured 2026-08-07n=100sourcelast sequential run (2026-08-07; mode retired)
Vercel Sandbox	308.8ms	Supersededmeasured 2026-07-17n=100sourcebest run (2026-07-17)
Vercel Sandbox	420.5ms	Supersededmeasured 2026-08-07n=100sourcelast sequential run (2026-08-07; mode retired)
E2B	452.7ms	Supersededmeasured 2026-08-07n=100sourcemedian of per-run medians (130 runs)
Modal	457.6ms	Supersededmeasured 2026-07-17n=100sourcebest run (2026-07-17)
Blaxel	536.2ms	Supersededmeasured 2026-08-07n=100sourcemedian of per-run medians (130 runs)
Modal	548.9ms	Supersededmeasured 2026-08-07n=100sourcelast sequential run (2026-08-07; mode retired)
E2B	575.5ms	Supersededmeasured 2026-08-07n=100sourcelast sequential run (2026-08-07; mode retired)
Vercel Sandbox	577.3ms	Supersededmeasured 2026-08-07n=100sourcemedian of per-run medians (128 runs)
Daytona	988.1ms	Supersededmeasured 2026-08-07n=100sourcelast sequential run (2026-08-07; mode retired)
Cloudflare Sandbox	1,364.3ms	Supersededmeasured 2026-05-06n=100sourcebest run (2026-05-06)
Modal	1,467.2ms	Supersededmeasured 2026-08-07n=100sourcemedian of per-run medians (126 runs)
Cloudflare Sandbox	1,728ms	Supersededmeasured 2026-08-07n=100sourcemedian of per-run medians (115 runs)
Cloudflare Sandbox	2,000.5ms	Supersededmeasured 2026-08-07n=100sourcelast sequential run (2026-08-07; mode retired)

sandbox start (300 starts)

client side · client, location not stated · ms · ascending (lower is better)

Competitor-runp50 · by Lizard (sells a competing sandbox) · one Railway container in us-east4 · measured 2026-10-12none recorded beyond the metric and vantage (fact-check research-sandbox.md F.2 row 12)

Who	Value	Provenance and conditions
Lizard	228ms	n=300source
E2B	264ms	n=300source
Daytona	819ms	n=300source
Modal	929ms	n=300source

create; resume to first command

client side · client, location not stated · ms · ascending (lower is better)

Independentcreate · by LogRocket (Ikeh Akinyemi) · not disclosed · published 2026-08-04none recorded beyond the metric and vantage (fact-check research-sandbox.md F.2 row 11)

Who	Value	Provenance and conditions
E2B	717ms	n not statedsource
Daytona	742ms	n not statedsource
Vercel Sandbox	1,852ms	n not statedsource
Modal	2,437ms	n not statedsource
Blaxel	2,824ms	n not statedsource

34 more cited numbers for sandboxes

Published by vendors and third parties, each measured its own way. They are listed for reference and not compared with each other or with B3OS.

Who	Value	What, provenance and conditions
zeroboot	0.79ms	template fork primitiveVendor claimp50by zeroboot (prototype)same hostserver sideretrieved 2026-10-25n not statedsource
Unikraft	10ms	advertised start time (marketing, no method)Vendor claimmaxby Unikraftnone statedclient sideretrieved 2026-10-25n not statedsource
Vercel Sandbox	0.3408USD/h	price of 2 vCPU / 4 GiB for one hourVendor claim100% CPUby Vercel Sandboxlist priceclient sideretrieved 2026-10-25n not statedsource
Northflank	0.0667USD/h	price of 2 vCPU / 4 GiB for one hourVendor claim100% CPUby Northflanklist priceclient sideretrieved 2026-10-25n not statedsource
Modal	0.2379USD/h	price of 2 vCPU / 4 GiB for one hourVendor claim100% CPUby Modallist priceclient sideretrieved 2026-10-25n not statedsource
Lizard	0.0162USD/h	price of 2 vCPU / 4 GiB for one hourVendor claim100% CPUby Lizardlist priceclient sideretrieved 2026-10-25n not statedsource
E2B	0.1656USD/h	price of 2 vCPU / 4 GiB for one hourVendor claim100% CPUby E2Blist priceclient sideretrieved 2026-10-25n not statedsource
Daytona	0.1656USD/h	price of 2 vCPU / 4 GiB for one hourVendor claim100% CPUby Daytonalist priceclient sideretrieved 2026-10-25n not statedsource
Cloudflare Containers	0.198–0.22USD/h	price of 2 vCPU / 4 GiB for one hourVendor claim100% CPUby Cloudflare Containerslist priceclient sideretrieved 2026-10-25n not statedsource
boxd	0.039EUR/vCPU-h	boxd EU rate cardVendor claimlistby boxdlist priceclient sideretrieved 2026-10-25n not statedsource
Blaxel	0.1656USD/h	price of 2 vCPU / 4 GiB for one hourVendor claim100% CPUby Blaxellist priceclient sideretrieved 2026-10-25n not statedsource
AWS Bedrock AgentCore Code Interpreter	0.056USD/h	price of 2 vCPU / 4 GiB for one hourVendor claim10% CPUby AWS Bedrock AgentCore Code Interpreterlist priceclient sideretrieved 2026-10-25n not statedsource
OpenAI (containers)	0.03USD/20 min	container price per 20-minute sessionVendor claim1 GBby OpenAIlist priceclient sideretrieved 2026-10-25n not statedsource
Anthropic (code execution tool)	0.05USD/container-hour	code execution container priceVendor claimafter 1,550 free hours/monthby Anthropiclist priceclient sideretrieved 2026-10-25n not statedsource
Perplexity SPACE	60ms	sandbox creationVendor claimp50by Perplexityproduction, server-side, identical traffic vs an unnamed previous providerserver sideretrieved 2026-10-25n not statedsource
Vercel Sandbox	10,000concurrent	concurrency and creation limitsVendor claimProby Vercel Sandboxdocumented limitsclient sideretrieved 2026-10-25n not statedsource
Modal	5,000containers	concurrency and creation limitsVendor claimTeamby Modaldocumented limitsclient sideretrieved 2026-10-25n not statedsource
E2B	1,100concurrent	concurrency and creation limitsVendor claimPro maxby E2Bdocumented limitsclient sideretrieved 2026-10-25n not statedsource
Cloudflare Containers	1,500vCPU	concurrency and creation limitsVendor claimaccount defaultby Cloudflare Containersdocumented limitsclient sideretrieved 2026-10-25n not statedsource
boxd	50machines	concurrency and creation limitsVendor claimper org defaultby boxddocumented limitsclient sideretrieved 2026-10-25n not statedsource
Blaxel	100,000concurrent	concurrency and creation limitsVendor claimTier 9by Blaxeldocumented limitsclient sideretrieved 2026-10-25n not statedsource
Freestyle	65ms	advertised start time (marketing, no method)Vendor claimmaxby Freestylenone statedclient sideretrieved 2026-10-25n not statedsource
Firecracker	125ms	boot to /sbin/init; VMM start; VMM memoryVendor claimboot maxby Firecracker projectm5d.metal (HT off), m6g.metalserver sideretrieved 2026-10-25n not statedsource
Daytona	90ms	advertised start time (marketing, no method)Vendor claimmaxby Daytonanone statedclient sideretrieved 2026-10-25n not statedsource
Cloudflare	5.387s	containers started by one accountVendor claim100,000 containersby Cloudflare (preliminary test)six locationsserver sideretrieved 2026-10-25n not statedsource
boxd	5ms	creation to running, median (279,577 machines)Vendor claimp50by boxdserver-side (production, Moyai)server sideretrieved 2026-10-25n not statedsource
Blaxel	25ms	resume from standbyVendor claimmaxby Blaxelserver-sideserver sideretrieved 2026-10-25n not statedsource
AWS Lambda MicroVMs	4.04s	launch to RUNNING; first response; suspended-VM task; warm taskIndependentlaunchby AWS Hero practitioner who built the tooling he evaluatesAWS us-east-1client sidemeasured 2026-10-08n=20source
DeltaBox	10.83ms	checkpoint work (hidden under inference); rollback; forkVendor claimcheckpointby DeltaBox authors (SJTU, Huawei)authors' own evaluationserver sidepublished 2026-06-08n not statedsource
CubeSandbox	21.5–25.7MB	host memory per VM, amortizedVendor claim100 to 1,000 VMsby Tencent (CubeSandbox 2026-06-01 perf report)same hostserver sidepublished 2026-06-01n not statedsource
CubeSandbox	47.8ms	create, serial (2 vCPU / 2 GiB)Vendor claimmeanby Tencent (CubeSandbox 2026-06-01 perf report)same host, bare metal BMI5 (Xeon 8255C, 96 threads), POST /sandboxes to runningclient sidepublished 2026-06-01n not statedsource
CubeSandbox	276.1ms	create at 50 concurrentVendor claimmeanby Tencent (CubeSandbox 2026-06-01 perf report)same host, bare metalclient sidepublished 2026-06-01n not statedsource
CubeSandbox	98.1ms	create at 20 concurrentVendor claimmeanby Tencent (CubeSandbox 2026-06-01 perf report)same host, bare metalclient sidepublished 2026-06-01n not statedsource
CubeSandbox	88.7ms	create at 10 concurrentVendor claimmeanby Tencent (CubeSandbox 2026-06-01 perf report)same host, bare metalclient sidepublished 2026-06-01n not statedsource

routines

Measuring

Pure step dispatch

Dispatch of steps with no side effects in an ephemeral run, 1 KB payloads.

Measuringp50, server side

Pure step dispatch, p99

The same at the 99th percentile.

Measuringp99, server side

Action call from a step

A routine step calling a pack action in-process, upstream excluded, 1 KB payloads.

Measuringp50, server side

Durable step, standard

An effect step with standard durability, journal commit included.

Measuringp50, server side

Durable step, strict

An effect step with strict durability: an intent commit and a result commit.

Measuringp50, server side

Webhook to first step

From a hook delivery received to the first step started.

Measuringp50, server side

Cron lateness

How late a per-second cron starts its runs: start time minus scheduled time.

Measuringp99, server side

Pure step throughput

Pure steps completed per second across concurrent ephemeral runs.

Measuringthroughput, server side

Durable record throughput

Standard-durability effect records journaled per second, records up to 256 B.

Measuringthroughput, server side

What others publish

Each table holds numbers measured the same way, from the same side and kind of vantage point, best first. B3OS joins a table only with a number measured that way too.

delay of scheduled invocations

client side · client, location not stated · s · ascending (lower is better)

Who	Value	Provenance and conditions
AWS EventBridge Scheduler	0–50s	IndependentSupersededspreadby Serverless By Theodo (consultancy independent of AWS)10,000 samplespublished 2022-12-01n=10,000sourcenone recorded beyond the metric and vantage (fact-check research-routines.md F.2 row 14)

40 lightweight tasks, total

client side · vendor figure, method not stated · s · ascending (lower is better)

Who	Value	Provenance and conditions
workflow engines	1.222s	Competitor-runHatchetby Windmill (benchmarking competitors)m4.large, docker-composemeasured 2025n not statedsourcenone recorded beyond the metric and vantage (fact-check research-routines.md A11)

single durable step, median

client side · vendor figure, method not stated · ms · ascending (lower is better)

Who	Value	Provenance and conditions
Restate	3ms	Vendor claimp50 low loadby Restate3-node, 3-way replicated c6id.8xlarge clusterpublished 2025-02-20n not statedsource9-step workflow p50 31 ms (low) / 116 ms (high), p99 163 ms
Diagrid Catalyst (managed Dapr Workflows)	25ms	Vendor claimp50 5 activitiesby Diagridclient-measured, schedule call to terminal state; 3 × m7i.2xlarge + Postgresupdated 2026-10-14n not statedsource10 workflows/s warm; ~2.6 ms per activity amortized ((25 − 12) / 5, a model)
Hatchet	25ms	Vendor claimp95by Hatchethot workersretrieved 2026-10-25n not statedsourcenone recorded beyond the metric and vantage (fact-check research-routines.md F.1 item 8, F.2 row 8)
AWS Lambda durable functions	41ms	Vendor claimp50 Pythonby AWS (durable-execution SDK maintainers)Lambda us-east-1, 1024 MB, warm, single runspublished 2026-09-15n=1,000sourcenone recorded beyond the metric and vantage (fact-check research-routines.md F.1 item 4)
Temporal	50ms	Vendor claimper activity (approx)by TemporalTemporal Cloud (approximate)retrieved 2026-10-25n not statedsourcenone recorded beyond the metric and vantage (fact-check research-routines.md F.2 row 5)
AWS Step Functions (Standard)	201ms	Competitor-runSupersededsingleby DBOS (measuring a competitor)not statedpublished 2024-06-10n not statedsourcenone recorded beyond the metric and vantage (fact-check research-routines.md F.2 row 9)

0-step workflow

client side · vendor figure, method not stated · ms · ascending (lower is better)

Who	Value	Provenance and conditions
Restate	5ms	Vendor claimp50 low loadby Restate3-node clusterpublished 2025-02-20n not statedsourcenone recorded beyond the metric and vantage (fact-check research-routines.md F.3)

StartWorkflowExecution API latency

server side · server side, as the provider reports it · ms · ascending (lower is better)

Who	Value	Provenance and conditions
Temporal Cloud	20ms	Vendor claimp50by TemporalTemporal Cloud API request latency, week-long windowmeasured 2026-08n not statedsourceJan 2026: 14 / 21 / 69 ms; StartActivityExecution 13 / 19 / 45; Signal 19 / 42 / 91; p99 SLO 200 ms

eager workflow start (illustrative)

client side · client in the provider's region · ms · ascending (lower is better)

Who	Value	Provenance and conditions
Temporal	16.7ms	Vendor claimp50by Temporalstarter and worker in the namespace's AWS regionpublished 2024-04-29n not statedsourcevs 29.3 / 51.6 ms without eager start

workflow executions per second, one instance

server side · vendor figure, method not stated · ascending (lower is better)

Vendor claim

Who	Value	Provenance and conditions
n8n	220executions/s	single instanceby n8nc5a.large, webhook + Edit Fields noderetrieved 2026-10-25n not statedsourcenone recorded beyond the metric and vantage (fact-check research-routines.md A15)
Diagrid Catalyst	530workflows/s	sustainedby Diagrid3 × m7i.2xlarge + one m7i.2xlarge Postgresupdated 2026-10-14n not statedsourcenone recorded beyond the metric and vantage (fact-check research-routines.md F.1 item 5)
DBOS	43,000workflows/s	directby DBOSone db.m7i.24xlarge (96 vCPU, 120K IOPS io2)published 2026-04-23n not statedsourceno latency figures
Restate	94,286actions/s	clusterby Restate3-node, 3-way replicated cluster (3 × 32 vCPU)published 2025-02-20n not statedsourcenone recorded beyond the metric and vantage (fact-check research-routines.md F.1 item 2)

webhook-triggered 2-task flow at 1,000 executions/min

client side · vendor figure, method not stated · ms · ascending (lower is better)

Vendor claimretrieved 2026-10-25

Who	Value	Provenance and conditions
Kestra 2.0 EE	50ms	meanby Kestran2-standard-4; Kestra 2.0.0 EE (RabbitMQ + Postgres)n not statedsourcep99 ≤107 ms up to 7,000/min; OSS: 115 ms mean, 151 ms p99 at 1,000/min, collapses at 5,000/min
Activepieces	163ms	p50by Activepiecesnot stated by the sourcen not statedsourcesync webhook ~0.2 s warm / ~2 s cold; ~0.5 s with every worker busy; 160 workers: p50 180 / p99 652 ms; the page gives no hardware or date

10 more cited numbers for routines

Published by vendors and third parties, each measured its own way. They are listed for reference and not compared with each other or with B3OS.

Who	Value	What, provenance and conditions
Vercel Workflows	60USD/1M steps	price per 1M workflow stepsVendor claimapprox (3 events per step at $0.02/1K events)by Vercel Workflowslist priceclient sideretrieved 2026-10-25n not statedsource
Temporal Cloud	25–50USD/1M actions	price per 1M workflow stepsVendor claimself-serveby Temporal Cloudlist priceclient sideretrieved 2026-10-25n not statedsource
AWS Step Functions	25USD/1M transitions	price per 1M workflow stepsVendor claimStandardby AWS Step Functionslist priceclient sideretrieved 2026-10-25n not statedsource
Restate Cloud	10–25USD/1M actions	price per 1M workflow stepsVendor claimtieredby Restate Cloudlist priceclient sideretrieved 2026-10-25n not statedsource
AWS Lambda durable functions	8USD/1M operations	price per 1M workflow stepsVendor claimlistby AWS Lambda durable functionslist priceclient sideretrieved 2026-10-25n not statedsource
Inngest	15–50USD/1M executions	price per 1M workflow stepsVendor claimtieredby Inngestlist priceclient sideretrieved 2026-10-25n not statedsource
Hatchet	30USD/1M runs	price per 1M workflow stepsVendor claimafter 1M freeby Hatchetlist priceclient sideretrieved 2026-10-25n not statedsource
Google Cloud Workflows	10USD/1M steps	price per 1M workflow stepsVendor claiminternalby Google Cloud Workflowslist priceclient sideretrieved 2026-10-25n not statedsource
DBOS	40–50USD/1M checkpoints	price per 1M workflow stepsVendor claimTeams / Pro overageby DBOSlist priceclient sideretrieved 2026-10-25n not statedsource
Cloudflare Workflows	8USD/1M steps	price per 1M workflow stepsVendor claimafter 500Kby Cloudflare Workflowslist priceclient sideretrieved 2026-10-25n not statedsource

actions

Measuring

Declarative action overhead

Time added by the action runtime, upstream excluded, for responses up to 16 KB.

Measuringp50, server side

Input validation

JSON Schema validation of an action's input.

Measuringp50, server side

JavaScript action overhead

Warm overhead of a first-party JavaScript action.

Measuringp50, server side

T0 call total minus upstream, per MiB of upstream response, for responses over 16 KB

Measuringp50, server side

3 more cited numbers for actions

Published by vendors and third parties, each measured its own way. They are listed for reference and not compared with each other or with B3OS.

Who	Value	What, provenance and conditions
Zapier MCP	3.4–5.1USD/1K calls	tool-call price per 1,000 callsVendor claim2M-task tier, annual to monthlyby Zapier MCPlist priceclient sideretrieved 2026-10-25n not statedsource
Composio	0.3USD/1K calls	tool-call price per 1,000 callsVendor claimProby Composiolist priceclient sideretrieved 2026-10-25n not statedsource
Arcade	10USD/1K calls	tool-call price per 1,000 callsVendor claimTeam ($0.01/call)by Arcadelist priceclient sideretrieved 2026-10-25n not statedsource

infer

Measuring

Time to first token

Streamed chat completion on the co-located model, with B3OS idle and at budgeted load.

Measuringp50, client side

Time per output token

The same stream, per output token.

Measuringp50, client side

What others publish

Each table holds numbers measured the same way, from the same side and kind of vantage point, best first. B3OS joins a table only with a number measured that way too.

LLM time to first token (median / P95) as reported by the guide

client side · vendor figure, method not stated · ms · ascending (lower is better)

Vendor claimp50 · by Daily (voiceaiandvoiceagents.com guide; Daily sells voice infrastructure) · not stated · updated 2026-06Rule of thumb in the guide: "600 ms or less is fast enough".

Who	Value	Provenance and conditions
OpenAI (GPT-4.1)	536ms	n not statedsource
NVIDIA (self-hosted) (Nemotron 3 Ultra)	541ms	n not statedsource
Google (Gemini 2.5 Flash)	597ms	n not statedsource
Anthropic (Claude 4.5 Haiku)	637ms	n not statedsource
OpenAI (GPT-5.1)	739ms	n not statedsource

time to first audio (vendor)

server side · server side, as the provider reports it · ms · ascending (lower is better)

Vendor claimserver-side (vendor) · retrieved 2026-10-25

Who	Value	Provenance and conditions
Inworld (TTS-2 Flash)	20ms	p90by Inworldn not statedsourceP90 time to first audio byte, server-side, excludes network
ElevenLabs (Flash v2.5)	75ms	approxby ElevenLabsn not statedsourceexcluding application and network latency
Cartesia (Sonic 3.6)	90ms	maxby Cartesian not statedsource"first audio in under 90ms", no conditions stated

time to first audio, median of 50 sequential live streaming trials

client side · client, location not stated · ms · ascending (lower is better)

Independentp50 · by Vapi Humanness Index (Vapi is a voice platform, not a TTS vendor) · client, network RTT includednone recorded beyond the metric and vantage (fact-check research-render.md A17, B9)

Who	Value	Provenance and conditions
Cartesia (Sonic 3.5)	128ms	retrieved 2026-10-25n=50source
Fish Audio (S2.1-Pro)	141ms	measured 2026-07n=50source
Cartesia (Sonic 2)	159ms	measured 2026-06n=50source
Cartesia (Sonic 3)	166ms	measured 2026-06n=50source
ElevenLabs (Flash v2.5)	197ms	measured 2026-06n=50source

10 more cited numbers for infer

Published by vendors and third parties, each measured its own way. They are listed for reference and not compared with each other or with B3OS.

Who	Value	What, provenance and conditions
Inworld	12.5–25USD/1M chars	TTS price per 1M charactersVendor claimTTS-2, list price by planby Inworldlist priceclient sideretrieved 2026-10-25n not statedsource
ElevenLabs	0.04USD/1k chars	TTS price per 1,000 charactersVendor claimFlash/Turbo v2.5, v4 Turboby ElevenLabslist priceclient sideretrieved 2026-10-25n not statedsource
RunPod	5.93USD/h	H200 per hourVendor claimserverlessby RunPodlist priceclient sideretrieved 2026-10-25n not statedsource
RunPod	3.59USD/h	H200 per hourVendor claimon-demand pod (Community); Secure $5.29by RunPodlist priceclient sideretrieved 2026-10-25n not statedsource
Modal	4.54USD/h	H200 per hourVendor claimserverless ($0.001261/s)by Modallist priceclient sideretrieved 2026-10-25n not statedsource
Daytona	2.61USD/h	H200 per hourVendor claimpreemptibleby Daytonalist priceclient sideretrieved 2026-10-25n not statedsource
Fish Audio (Drama 3 (open weights))	88.9ms	time to first audio on one H200Vendor claimp50 c=1by Fish Audioone H200, SGLang-Omni, voice-clone streamingserver sideretrieved 2026-10-25n not statedsource
TTS models	1,335Elo	TTS arena EloIndependentEleven v4 Turbo (CI 1318-1352)by Artificial Analysis (crowd votes)not applicable (quality score)server sideretrieved 2026-10-25n not statedsource
Kyutai (TTS 1.6B)	150ms	time to first audio at batch 1; realtime factor at batch 64Vendor claimbatch 1 (computed)by Kyutaione H100, computed from per-step compute timeserver sidepublished 2025-09n not statedsource
Kokoro (hosted by Replicate, DeepInfra)	0.65–0.99USD/1M chars	hosted Kokoro price per 1M charactersVendor claimlistby Kokoro (hosted by Replicate, DeepInfra)list priceclient sidepublished 2025-04n not statedsource

Platform

Under every product

The HTTP floor is a round trip to a scaffold route on b3osd as the harness sees it: the error bar under every client-side number on this page.

HTTP floor

A round trip to a scaffold route on b3osd as the harness sees it: the floor under every client-side number.

p50150.729us

Least favorable of 12 counted runsMeasuredp50, median of 5 runsbest run 111.949 ussame-host, same NUMA node (harness 20-23, b3osd 16-19,80-83)client sideOct 26, 2026n=100,000raw data result fileConfiguration: tcp + node-token, same-node (NUMA), back-to-back, concurrency 1; path=/v1/vision, auth=true, response=scaffold JSON (~50 B), max_sockets=1024Not measured yet: mixed load

Back to back4 runs

HTTP floor, back to back
Run	Value	Stat, sample and data
UDS, same NUMA nodeuds, same-node (NUMA), back-to-back, concurrency 1; path=/v1/vision, auth=true, response=scaffold JSON (~50 B), max_sockets=1024	49.25us	Measuredp50, median of 5 runsbest run 41.019 (p95 51.079, p99 56.438) usclient sideOct 26, 2026n=100,000raw data result fileSame run: Server-Timing total 7 us · client round trip minus server total 42.109 us
TCP, same NUMA nodetcp without a credential, same-node (NUMA), back-to-back, concurrency 1; path=/v1/vision, auth=true, response=scaffold JSON (~50 B), max_sockets=1024	47.81us	Measuredp50, median of 5 runsbest run 47.05 (p95 50.69, p99 53.87) usclient sideOct 26, 2026n=100,000raw data result fileSame run: Server-Timing total 7 us · client round trip minus server total 40.74 us
TCP + node token, same NUMA nodeHeadlinetcp + node-token, same-node (NUMA), back-to-back, concurrency 1; path=/v1/vision, auth=true, response=scaffold JSON (~50 B), max_sockets=1024	150.729us	Measuredp50, median of 5 runsbest run 111.949 (p95 118.479, p99 123.249) usclient sideOct 26, 2026n=100,000raw data result fileSame run: Server-Timing total 79 us · client round trip minus server total 71.588 us
UDS, cross NUMA nodeuds, cross-node (NUMA), back-to-back, concurrency 1; path=/v1/vision, auth=true, response=scaffold JSON (~50 B), max_sockets=1024	44.819us	Measuredp50, median of 5 runsbest run 44.619 (p95 48.659, p99 57.16) usclient sideOct 26, 2026n=100,000raw data result fileSame run: Server-Timing total 7 us · client round trip minus server total 37.719 us

Idle gaps of 100 ms3 runs

HTTP floor, idle gaps of 100 ms
Run	Value	Stat, sample and data
UDS, same NUMA nodeuds, same-node (NUMA), idle gaps of 100 ms, one at a time; path=/v1/vision, auth=true, response=scaffold JSON (~50 B), max_sockets=1024	87.5us	Measuredp50, median of 5 runsbest run 87.024 (p95 131.723, p99 282.051) usclient sideOct 26, 2026n=1,000raw data result fileSame run: Server-Timing total 7 us · client round trip minus server total 78.218 us
TCP, same NUMA nodetcp without a credential, same-node (NUMA), idle gaps of 100 ms, one at a time; path=/v1/vision, auth=true, response=scaffold JSON (~50 B), max_sockets=1024	109.385us	Measuredp50, median of 5 runsbest run 108.291 (p95 145.883, p99 189.499) usclient sideOct 26, 2026n=1,000raw data result fileSame run: Server-Timing total 7 us · client round trip minus server total 100.02 us
TCP + node token, same NUMA nodetcp + node-token, same-node (NUMA), idle gaps of 100 ms, one at a time; path=/v1/vision, auth=true, response=scaffold JSON (~50 B), max_sockets=1024	189.732us	Measuredp50, median of 5 runsbest run 186.16 (p95 230.662, p99 290.364) usclient sideOct 26, 2026n=1,000raw data result fileSame run: Server-Timing total 79 us · client round trip minus server total 107.298 us

Idle gaps of 3 s3 runs

HTTP floor, idle gaps of 3 s
Run	Value	Stat, sample and data
UDS, same NUMA nodeuds, same-node (NUMA), idle gaps of 3 s, one at a time; path=/v1/vision, auth=true, response=scaffold JSON (~50 B), max_sockets=1024	394.238us	Measuredp50, median of 5 runsbest run 385.17 (p95 722.088, p99 722.088) usclient sideOct 26, 2026n=50raw data result fileSame run: Server-Timing total 20 us · client round trip minus server total 350.472 us
TCP, same NUMA nodetcp without a credential, same-node (NUMA), idle gaps of 3 s, one at a time; path=/v1/vision, auth=true, response=scaffold JSON (~50 B), max_sockets=1024	430.844us	Measuredp50, median of 5 runsbest run 380.632 (p95 574.787, p99 574.787) usclient sideOct 26, 2026n=50raw data result fileSame run: Server-Timing total 22 us · client round trip minus server total 403.107 us
TCP + node token, same NUMA nodetcp + node-token, same-node (NUMA), idle gaps of 3 s, one at a time; path=/v1/vision, auth=true, response=scaffold JSON (~50 B), max_sockets=1024	509.83us	Measuredp50, median of 5 runsbest run 341.967 (p95 1,149.926, p99 1,149.926) usclient sideOct 26, 2026n=50raw data result fileSame run: Server-Timing total 81 us · client round trip minus server total 422.194 us

Open loop2 runs

HTTP floor, open loop
Run	Value	Stat, sample and data
UDS, same NUMA nodeuds, same-node (NUMA), open loop, constant arrivals at 2000/s; path=/v1/vision, auth=true, response=scaffold JSON (~50 B), max_sockets=1024	40.868us	Measuredp50, median of 5 runsbest run 40.615 (p95 43.174, p99 51.409) usclient sideOct 26, 2026n=100,000raw data result fileSame run: Server-Timing total 7 us · client round trip minus server total 33.11 us
TCP, same NUMA nodetcp without a credential, same-node (NUMA), open loop, constant arrivals at 2000/s; path=/v1/vision, auth=true, response=scaffold JSON (~50 B), max_sockets=1024	49.276us	Measuredp50, median of 5 runsbest run 49.21 (p95 52.355, p99 64.777) usclient sideOct 26, 2026n=100,000raw data result fileSame run: Server-Timing total 7 us · client round trip minus server total 41.62 us

connections (vault)

3 more cited numbers for connections (vault)

Published by vendors and third parties, each measured its own way. They are listed for reference and not compared with each other or with B3OS.

Who	Value	What, provenance and conditions
Vercel Connect	3USD/1K token requests	token vault price per 1,000 requestsVendor claimProby Vercel Connectlist priceclient sideretrieved 2026-10-25n not statedsource
Infisical	2USD/1K brokered requests	token vault price per 1,000 requestsVendor claimafter 5,000 free/monthby Infisicallist priceclient sideretrieved 2026-10-25n not statedsource
AWS Bedrock AgentCore Identity	0.01USD/1K token or key requests	token vault price per 1,000 requestsVendor claimlistby AWS Bedrock AgentCore Identitylist priceclient sideretrieved 2026-10-25n not statedsource

Method

How we measure

Provenance on every number

Measured (ours, with raw data), vendor claim, independent, competitor-run or affiliated. Each number also carries its vantage, side, date, sample count and a raw-data link.

Server side means the whole request

Server-side numbers run from the first request byte to the last response byte on the node. Client-side numbers are labeled client side.

Open-loop load

Load generators send at a constant arrival rate and time each request from when it should have been sent, so queueing shows up in the results.

Three regimes

Back to back, with idle gaps of 100 ms and 3 s, and under mixed load with B3IQ inference running.

Locked runs only

Published runs hold the bench lock and record the active services, load average, versions and commit. Anything else is a dev run and stays off this site.

Five runs or more

We publish the median of run medians and the best run, with the sample count.

Like for like

Competitors are measured the way their own benchmarks measure them. Same-host numbers never sit in a row beside client-measured competitor numbers.

Baselines are not results

Before any product code existed we measured the node's floors: microVM restore, browser spawn, disk sync, frame capture. Those runs did not hold the bench lock, so they shaped the design and stay off this page.