What's new.

pin:byo:@vision, @text and @any send a call to whichever of your subscription lanes has the most capacity to spare, expiring capacity first, alternating providers so one family's limit falls through to another. Pins now work on /v1/messages, streamed Claude Code turns get 600 s with keepalives, in-flight caps are per plan, raised keys get a matching per-second burst, and long-output responses always end in JSON. The docs have one table of every limit, its error and what to do.

A 5-10K-token answer without streaming used to hit the request budget and return 503 after the first lane was cut off. Requests with max_tokens >= 4096 (or long:true) now run under a 600 s budget with whitespace keepalives on the connection and one JSON body at the end; the throughput guard no longer drops slower lanes for them. Speech: an all-lanes rate limit now returns 429 with Retry-After, and speech-2.8-hd falls through to Turbo (same voices) when rate-limited.

POST /v1/video/generations renders clips asynchronously. The Hailuo 2.3 clips included in connected MiniMax Token Plans are spent first at $0, rotating across plans; MiniMax-H3 (up to 15 s, 2K, reference images) runs pay-as-you-go when you pin it or allow it. Poll the job, download the url. Speech now rotates across every connected MiniMax plan too.

POST /v1/audio/speech now routes over MiniMax Speech 2.8 (HD/Turbo, 38 Mandarin and Cantonese voices), Alibaba Qwen3-TTS (Beijing and Sichuan accents, voice design, tone instructions), Qwen-Audio-3.0-TTS, Xiaomi MiMo-TTS, CosyVoice2, IndexTTS-2, Fish Speech, Gemini 3.8 TTS, OpenAI and Kokoro. model=auto detects the language of the text and picks the best lane, falling through on provider errors; auto-hd and auto-fast re-rank by quality or price. GET /v1/audio/models and /v1/audio/voices list what is available. Same per-character metering as before.

Every agent gets a durable address (agent://team/name) and a crash-safe FIFO mailbox, across Claude Code, Codex, Grok, Gemini, Kimi or curl. Free, no registration: the first agent signs itself up; teammates join with an invite code or a minted key. A $0 doorbell daemon wakes a terminal session when mail lands; services get verified, signed webhooks. 19 MCP tools (a bus-only server at /mcp/bus with resources and prompts), lease renewal, poison backoff, idempotent retries, history and replay, key rotation, account health — and one error envelope with a remedy on every refusal. Docs at /docs#bus; the agent's manual at api.flowaiapi.com/v1/bus/onboard.

expect_reply now defaults to false. Ask for an answer explicitly with verb:"ask"; announce with verb:"notice". Four independent reviews called the old default the product's footgun — they were right.

Add verify:true to any (non-streaming) completion and the answer returns with a verdict: pass/fail, confidence, and concrete issues, from free heuristics plus a cheap strict-judge pass. verify:"council" has diverse models re-judge with a majority vote. Fail-open, never rewrites your answer, billed at pass-through. Paired with the MCP convene_council tool for full second-opinion workflows.

Flow AI is now installable as a Model Context Protocol server: claude mcp add --transport http flow-ai https://api.flowaiapi.com/mcp. Your agent can search the live catalog, read real clearing prices, and list the currently-promoted free models from inside its own session — no API key needed for discovery.

Canary-verified free-tier models are part of the default routing pool: each passes a live tool-calling and output-quality canary before serving, and routes only to tasks it's proven on. When a free model can do your job, the tokens cost $0. Live list at /free-models.

The docs now cover the Cortex read endpoints (overview, routing, caching, context) with app/agent/run_id request tagging, the /v1/audio/speech text-to-speech endpoint, and the MCP server.

Strict, opt-in model pinning for judge ensembles, cross-model evals, and reproducible research: pin any model exactly (pin:gpt-4o-mini), or fan one prompt across up to 10 models in a single /v1/panel call. Truthful served-model echo, fail-loud (never a silent fallback), billed at each model's real rate. Enable per key in the dashboard — routing stays the default.

A full visual refresh across the site — new mark, gold/red/obsidian palette, Sora + Inter type, and redesigned landing, Market, Pricing, and dashboard.

Icon-led dashboard sections, a clearer connect flow, and explicit success/error confirmation when you connect a provider key.

Routing that measures whether each model actually completes the task — per agent, content-free — and settles on the cheapest one that finishes the job.

Real cross-provider clearing prices, updated continuously — the published list price is the ceiling, the market only makes it cheaper.

Connect your provider subscriptions, and share spare capacity into the community Hive — capacity, never your data.

Save a routing config as @name for your whole fleet, and opt into measured exact/semantic caching when it would actually help your workload.

One base URL, drop-in for OpenAI, Anthropic Messages, and OpenAI Responses (Codex) clients — with DeepSeek and MiniMax live among many models.