NEW/search/lite is live at $0.25 / 1k reqs, beating Parallel turbo 91/100See the battle →

Release · September 1, 2026

Keirolabs V4 is here

Cloud browsers that render anything. Google results, direct and parsed. An autonomous WebAgent that now leads its benchmark class across four datasets. And infrastructure rebuilt from the substrate up to hold 99.95% — not promised, engineered.

99.95%Uptime SLA
<300ms/search/lite P50
3xProvider redundancy
6,109Benchmark queries

Cloud Browsers

Every agent eventually hits the wall: a page that loads fine in curl and renders nothing in your pipeline. V4 removes the wall. Managed headless Chromium sits behind the API — spin up a browser, render JavaScript, interact with pages, take screenshots, and extract content from SPAs. No Puppeteer to configure, no browser pools to babysit, no anti-bot systems to fight. You send a URL; you get the rendered web.

Session managementJS renderingScreenshotsForm interactionAnti-detect

Google SERP

For the first time, Keirolabs speaks to Google directly. One call returns the full results page as clean, structured JSON: organic results, knowledge panels, featured snippets, local packs, and “People also ask.” Fifty billion indexed pages behind a single endpoint — no parsing HTML, no proxies, no SERP scraping infrastructure of your own.

Organic resultsKnowledge graphFeatured snippetsLocal packs

Stability, engineered — not promised

Every failure mode from our July incident has been architecturally eliminated. V4 runs on three independent compute providers, none carrying more than 45% of traffic, with per-region circuit breakers and active health probes every five seconds. Pre-warmed standby capacity activates in under twenty seconds. When something upstream breaks, your requests fail over before you notice — and the 99.95% uptime SLA holds because the system was built to hold it.

3 compute providersPer-region circuit breakers5s health probes20s failoverQueue backpressure503 on timeout

WebAgents, autonomous by default

The headline capability of V4 isn't an endpoint — it's an agent. The KeiroLabs WebAgent doesn't query a search API and hand you URLs. It operates a self-correcting ReAct loop over the live web: it plans its own sub-queries, scrapes and chunks what it finds, grades every chunk against a quality gate, and re-searches when the evidence is thin — up to six iterations — before synthesizing a cited answer.

01

Intent

Classifies the query and picks a depth: light, medium, or deep.

02

Plan

Generates 3–5 hyper-specific sub-queries aimed past SEO spam, at primary sources.

03

Execute

Runs the queries across the web; payloads are scraped, converted to Markdown, and chunked.

04

Quality gate

An internal judge grades every chunk. Insufficient or contradictory data triggers a new iteration — up to six.

05

Synthesize

Formats the final answer with strict, inline citations.

WebAgents in Slack

You don't need a pipeline to put the WebAgent to work. V4 ships it inside Slack — drop a question in a channel, and the same autonomous loop that tops the benchmarks runs it live: planning, searching, verifying, and answering with citations, right in the thread. Deep research where your team already is.

Benchmarked, not marketed

We put the WebAgent through four industry-standard datasets — over 6,000 queries in total — and published everything: methodology, prompts, judge configuration, and raw scores. Accuracy is graded by an independent LLM judge against official ground truth. Read the full study on GitHub

vs. frontier LLMs (zero-shot)

GAIA · complex
Frames · multi-hop
FreshQA · temporal
SimpleQA · factoid
KeiroLabs WebAgent
77.25
72.94
90.60
90.42
Claude Opus 5
52.30
61
65
55
GPT-5.6 Sol
50.50
58.50
63
52
Qwen3.8 Max
49.50
59.50
61
51.50
Gemini 3.1 Pro
49
63
60
50
DeepSeek-V4-Pro
48.80
57
62
50.50
Kimi K3
47
56.50
58
49

vs. search APIs & web agents

GAIA · complex
Frames · multi-hop
FreshQA · temporal
SimpleQA · factoid
KeiroLabs WebAgent
77.25
72.94
90.60
90.42
Perplexity Pro
62
65
80
75
You.com Pro
58
62
75
70
Tavily API
50
55
54
65
Exa API
48
52
58
68
Firecrawl API
47
51
53
63
Browser Use
46.50
49
52
47
AutoGPT 2.0
45.10
48
50
44.50

What was tested

DatasetQuestionsCorrectAccuracy
SimpleQA4,3213,90790.42%
FreshQA50045390.60%
GAIA46636077.25%
Frames82460172.94%

Everything else in V4

New

/search/lite

Sub-300ms searches at $0.25/1k requests, built for high-volume RAG pipelines that need speed over depth.

New

Streaming responses

SSE-based streaming on /answer and /agentic. Intermediate results arrive as they happen — no waiting on the full completion.

New

Regional pinning

Pin requests to EU, US, or APAC with the X-Region header. Data sovereignty with zero configuration.

Improved

Batch v2

10x throughput, live progress webhooks, auto-retry on transient failures. 100k queries in a single job.

Improved

SDKs, rebuilt

Fully typed Python, Node, and Go SDKs with async/await, streaming, and automatic retries. Zero config to start.

Improved

Adaptive rate limits

Per-endpoint token buckets that burst when you need them, with clear Retry-After headers and usage visibility in the dashboard.

Changelog

Added

Cloud Browsers — managed headless Chromium behind the API, with session management, screenshots, and JS rendering.

Added

Google SERP endpoint — structured Google results with organic, knowledge graph, and local pack parsing.

Added

/search/lite — ultra-fast search at $0.25/1k for high-volume RAG pipelines.

Added

SSE streaming on /answer and /agentic for real-time intermediate results.

Improved

Stability overhaul — multi-provider compute, circuit breakers, 5s active health probes, pre-warmed standby.

Improved

Batch v2 — 10x throughput, live progress webhooks, 100k query jobs.

Improved

All SDKs rebuilt with streaming, full type safety, and automatic retries.

Improved

Regional pinning is now GA — X-Region header for EU/US/APAC data sovereignty.

Fixed

HTTP 200 on timeout — upstream timeouts now correctly return 503 with Retry-After.

Fixed

Empty result ambiguity — genuine empty results now include an empty_reason field.

V4 is live. Start building.

All existing API keys work — no migration required. New features are available immediately on all paid tiers.