Browserbase

Session replays can now be embedded in your product and streamed to end users within seconds of a session ending.

Fetch API now returns markdown and JSON — the Fetch API, launched for sessionless page retrieval at ~$1/1,000 pages, now returns content as markdown or structured JSON. The data cap rose from 1MB to 5MB, and extract pricing landed at $4/1,000 pages ($7 with proxies). Use it when a full browser session isn't warranted.

Session replay streaming — embed an HLS-compatible replay in your own product by fetching a .m3u8 playlist from the API and handing it to any player. Browserbase handles storage, fMP4 encoding, and CDN delivery; end users stream segments directly.

Stagehand agent reliability — Stagehand 3.3.0 introduced verified mode for improved agent identity on sites that gate bot traffic. Claude CUA now scales extended-thinking budget to match task complexity, and extract() uses strict JSON schema enforcement to eliminate malformed responses. Stagehand 3.4 expanded CUA model support to openai/gpt-5.4-mini, openai/gpt-5.5, and anthropic/claude-haiku-4-5.

observe() and extract() noise filtering — ignoreSelectors, available since Stagehand 3.4, lets agents skip ads, navbars, modals, and other irrelevant elements during both observation and extraction passes without re-tuning prompts. Stagehand 3.5 added a screenshot option to extract() and introduced an evaluator backend for recording and assessing agent trajectories.

Downloads API — every file an agent downloads is now individually addressable with its own ID and metadata. Filter by filename, MIME type, size, or timestamp; fetch a single file or just its metadata. Existing session-level zip endpoints continue to work without migration.

Model Gateway and auth — passing a Browserbase API key now routes Stagehand through OpenAI, Anthropic, or Gemini at list price with no extra accounts. Server v3.7.2 added Azure Entra model auth support. projectId was deprecated; only BROWSERBASE_API_KEY is required to start a session.