POST /v1/chat/completions now accepts reasoning_effort (minimal | low | medium | high | xhigh | max) and include_reasoning, and translates them to whatever the model that serves the request actually takes — output_config.effort for Claude, reasoning_effort for OpenAI reasoning models, and nothing at all for models with no reasoning control, which is reported in the log rather than silently dropped. A level the model does not accept is mapped to the nearest one it does: the mapping never resolves to a cheaper level than the caller asked for, so a request for max cannot quietly become low.

Thinking configuration moved from the protocol to the model, because the models no longer share one shape. Claude Sonnet 5, Opus 4.7/4.8 and the Fable line reject thinking: {type: "enabled", budget_tokens: N} and require thinking: {type: "adaptive"} with an effort. A budget sent to one of those is converted to adaptive with the depth carried across in effort, instead of being forwarded and failing with a 400. display: "summarized" is only sent when reasoning text was actually asked for — sending it otherwise made the provider return an empty thinking block.

Two fixes behind it. The interleaved-thinking beta flag was being put in the request body, where it is ignored; it now travels as an anthropic-beta header. And sampling parameters are withheld per model rather than per request: the API rejects temperature, top_p and top_k when thinking is on — and rejects the first two on the adaptive models whether or not thinking was requested — so they are omitted rather than forwarded to fail.

Users who have purchased credits via Stripe now get elevated limits: 600 RPM (vs 10–60) and a $300/day spend cap (vs $30). The upgrade is automatic — no configuration needed. Personal spend caps set on the profile page still override tier defaults.

Also added IP-level rate limiting to the verify-code endpoint to prevent account creation floods, sharing the same limiter bucket as request-code.

Groq and Cerebras free-tier models have been removed from the economy tier. Free provider rate limits (39% failure rate on Cerebras) were causing cascading 502 errors as user volume grew beyond what free quotas could sustain.

Economy tier now routes to cheap paid models: GPT-4.1 Mini, Gemini 2.5 Flash, Claude Haiku 4.5, Grok 3 Mini, and Bedrock models. New accounts still get $1 in trial credits — enough for millions of tokens.

Zero-balance users receive a 402 Payment Required response with a link to /billing.

The product has no tiers — the correct framing is that some models are free to use (Groq and Cerebras-hosted open-source models) and others are paid. Updated all user-facing copy, docs, and landing page content to reflect this.

Two new endpoints alongside the existing /v1/chat/completions: /v1/completions for legacy text completion workflows, and /v1/responses for the OpenAI Responses API format. Both support auto-routing and the same model parameter values.

Note: model pinning is required for these endpoints — auto-routing with tier aliases (economy, standard, premium) is not available.

Pass model: "auto" and the router analyses your messages to pick the appropriate tier automatically. Simple queries route to economy, complex reasoning to standard or premium. Overridable with explicit prefer hints (fast, coding, reasoning).

Full documentation at /docs/api#auto-routing.

New /try page: send requests to the router directly from your browser, authenticated with your session. Usage is billed to your credit balance at standard rates. Useful for testing routing decisions without writing code.

Added Groq and Cerebras as providers. Models on these platforms (including Llama 3.3 70B and Llama 3.1 8B) are available at no cost — the router covers provider costs. Paid credits are only required for models on Anthropic, OpenAI, Google, and AWS Bedrock.

The economy tier now routes exclusively to free models by default.

API keys no longer have a tier. The tier is specified per-request via the model parameter. This simplifies key management — one key works across all tiers and all endpoints.

Several small improvements to reduce the activation gap:

• New API keys are revealed in full on creation (with a copy button) before being masked
• A curl example using the new key is shown inline
• A banner nudges users who have a key but haven't made their first call
• Welcome email content updated — arrives 1 hour after signup with clearer next steps

claude-sonnet-4-6 and claude-opus-4-6 context windows updated to 1,048,576 tokens (1M) following Anthropic GA announcement on March 13 2026.

Vertex AI provider adapter added (Google's managed API endpoint, separate from the direct Gemini API). Nemotron 3 Nano 8B and 51B models added via AWS Bedrock.

Users can configure an OTLP endpoint in their profile to receive request telemetry (latency, token counts, routing decisions, provider used) as OpenTelemetry spans. Every response includes an X-Request-Id header for correlation.

Added /v1/embeddings with embed-small and embed-large tier aliases. Backed by Amazon Titan Embed Text v2. Supports batch input and returns normalized vectors.

Added prefer: "coding" as a routing hint. Routes to models with strong code benchmark scores within the selected tier.

Token bucket rate limiter applied per API key. Limits are tier-based and visible in response headers (X-RateLimit-*). Concurrent overdraft prevention via atomic credit reservation.

Authentication switched to passwordless email codes. Enter your email, receive a 6-digit code, done. No passwords to manage or forget.

Account system launched: sign up, manage API keys, view usage history, and add payment methods — all from /profile.

Credit-based billing via Stripe. Add credits to your account and they are deducted at cost as you make requests. A 4% platform fee (minimum $0.80) applies to each top-up.

Initial release. One endpoint (/v1/chat/completions), OpenAI-compatible, routes across Anthropic, OpenAI, and Google Gemini. Circuit-breaker failover, streaming support, cost-based tier selection.