Features, plan changes, improvements and fixes, newest first. Entries are dated records: older ones describe the product as it was then.
Latest22 October 2026·Free API week
The free API week is 25 calls a day
•The 7-day free API key now gives 25 calls a day instead of 100: enough to wire it into your tool and see it answer. Most free-week traffic was a few accounts calling all day, and every free call comes out of the same allowance that serves paying customers
•Annual (2,000 a day) and Lifetime (5,000 a day) are unchanged
Release9 October 2026·Own keys
Annual and Lifetime: carry on with your own Groq or OpenRouter key
•Add a Groq and/or OpenRouter key in Integrations. Once your daily API calls run out, or if our free-model quota fails, requests carry on on your own key — same endpoint, ranking and failover, no limit of ours
•Calls on your own key don't count against your daily calls, and responses say so with X-ZeroLimit-Keys: own
•Keys are checked with the provider when you save them and stored encrypted. Groq or OpenRouter keys saved on the old Integrations page before October are never used: add them again to opt in
•Plans are unchanged: Annual keeps 2,000 calls a day and Lifetime 5,000
Release3 October 2026·Annual plan
An Annual plan, a 7-day free API key, and a cleaner chat
•Annual: $49 a year ($29 in Latin America) — the same app as Lifetime, with 2,000 API calls a day. Cancel any time. Lifetime stays at $99 once with 10,000 a day
•The free API key is now a 7-day evaluation at 100 calls a day, then it needs Annual or Lifetime (HTTP 402 upgrade_required). We tried a permanent free tier for ten days: all free inference comes out of one shared allowance, and every free call was one a paying customer could not make. Accounts that got their key under the old promise keep 50 a day until 3 November, and we emailed them
•Over a daily quota you get HTTP 429, on every plan — never an automatic charge
•Chat: Stop keeps what was already written, image generation from the chat works, regenerate no longer duplicates the conversation, and code previews run fully sandboxed
Release23 September 2026·Free tier
A free API tier, and we stopped sending your prompts to trainers
•The free API tier became permanent: 100 calls/day for the first week, then 50 calls/day (replaced on 3 October by a 7-day free key — see above)
•We never route to a provider that may train on your prompt. OpenRouter publishes each provider's policy; we refuse the four it lists as training, on every single request. It cost us five otherwise-good models, including the second-ranked one
•A model name we do not serve is now routed instead of refused. Point an app at us without touching its hardcoded gpt-4o-mini and it works — the response says which model actually answered, and X-ZeroOptimize-Substituted-From names what you asked for
•Function calling: send tools in the OpenAI format and the request goes to the best free model that supports them, with failover
•Sampling parameters — temperature, top_p, stop, seed, penalties — are validated and forwarded; max_tokens caps the visible answer
•The leaderboard names the company that actually serves each model, not the one that trained it, and shows a retirement date when the provider has published one
•Community votes on the leaderboard: no login, one per model per browser. They sit beside the measured score and never feed the routing
•The first screen after signing up asks one question and then delivers the first success: an API key created and tested right there, or the chat
•Public support assistant that answers technical questions before you sign up, without spending your quota
•Fixed: free accounts could not see the Developer menu at all — the entry holding their own API key was hidden by a sidebar rule that contradicted the setting shown in our admin
•Fixed: a response cache keyed on too little of the request could replay another conversation's answer
•Fixed: empty replies. A model has not answered until you can see a word, so a silent stream now fails over instead of arriving blank
•New brand: the mark is a route — the request enters one node, crosses, and the answer leaves the other
ReleaseSeptember 2026·API-first
One plan, a real developer API, live routing
•One plan: Lifetime, $99 once — developer API (10,000 calls/day), unlimited chat, images, video and agents. Existing Lifetime Core accounts keep everything they bought
•Free API key for every account, no card
•API v1 is OpenAI-shaped end to end: errors as { error: { message, type, code } }, X-RateLimit-* headers on every response, Retry-After on 429, a proper final stream chunk, and an OpenAPI spec at /api/v1/openapi.json
•API reference rewritten against the real handlers (/docs/api), plus /llms.txt for coding agents
•Use cases page: six workloads with working snippets and what they would cost at list price (/use-cases)
•Live routing chain on the homepage, /developers and /status — the models answering right now, not a sample
•/status now measures the actual routing chain instead of a fixed provider list
•Developer dashboard: expired trial keys are labelled and no longer block creating a new one; a fresh key is filled into the code snippets
•Telegram bots: every chat has its own memory and history
•Fair-use terms spelled out: Lifetime is for you and your team; serving third parties goes through the API
•Language follows your browser before your country
•Every lifecycle email carries a one-click unsubscribe
•Homepage rewritten for developers: what it costs elsewhere, when we are the wrong choice, and seven straight answers
FeatureApril 2026·Sprint 15–16
Canvas, PWA & AppSumo Launch
•Live Code Canvas — render HTML, JSX/React, JS, SVG and CSS directly in chat (like Claude Artifacts)
•Slash commands — /image, /search, /remember, /help, /clear in the chat input