Play now keeps separate Image, Video, and Audio drafts. Record or import media, tune model controls, and ask yes/no, multiple-choice, or scale questions in Decide. See answer probabilities and copy a curl request. Even uncertainty gets an endpoint.
Added animated GIF requests to GPT Image 2.6 Sunburst, with frame count and duration controls. Converted Sunburst stills now report JPEG correctly. The pictures move; the format labels keep up.
stored audio and video now support single byte-range GETs, including async job results. players can seek without restarting. `zyphra/zonos3-mini` speech also accepts a public reference audio URL for voice cloning. turns out players like seeking.
image requests now support background=transparent on supported models. PNG alpha survives, and GPU responses are checked before we record success. status codes are easily impressed.
This week: MCP gained tools for papers, media, packages and math. Gemini Omni 1.2 Flash added paid video with audio (3-15s); Wan 3.0 supports 2-15s clips; GPT Image 2.6 Flare outputs GIFs. 28 community apps joined the catalog. Busy little toolbox.
Enter: Agents, Models and Apps now sit under My listings; old links redirect. Signed-out visitors can browse sandbox templates and prices. Keys has Secret and App tabs. Non-timeout Enter/gen failures show one outage banner at the top. The drawer stays quiet.
Added PNG, WebP, and animated GIF output to GPT Image 2.6 Flare, with frame and duration controls. Added supported resolution data to /v1/models where available. FLUX Schnell preserves aspect ratios when adjusting sizes. Pixels have opinions.
`/v1/responses` now reports provider failures as failed responses with token usage. Partial output is marked incomplete. Tool media stays with its result; URL citations keep titles, links, and offsets in buffered and streamed output. Even errors need a paper trail.
added ElevenLabs SFX v3 for MP3 sound effects (paid-only), Scribe v3 for batch transcription with word timestamps and speaker labels, and Scribe v3 Live for realtime incremental and final transcripts. Existing sound-effects and Scribe v2 models remain available.
Video controls and durations are now in /v1/models. Wan 3.0 accepts 2β15 seconds instead of a fixed 5. Unsupported durations return 400 before dispatch. The model list knows the rules, so your request doesn't have to guess.
shipped this week: `model` accepts up to 4 compatible IDs or aliases. Retryable errors try the next model, including on audio and `/v1/messages`. Opening text-stream errors now reach fallback; image jobs retry after worker disconnects. failure remains an option.
Legacy image requests now return 503 (Retry-After: 10) if the per-IP limiter fails. Cached images still work. Model catalog cache refreshes every minute; Quest leaderboard rebuilds every 15 minutes, not on visitor misses. Tomadix joined the image app catalog. Totals may lag.
Added model fallback chains: up to 4 IDs or aliases in `model`, tried in order on retryable upstream failures, including audio and `/v1/messages`. Coordinator recovery now returns completed cached results and wakes waiting callers without rerunning the provider.
added `polli acp [agent]` in CLI 0.1.80. Pollinations agents now stream into ACP editors, with permission-gated local tools. Code agents can return calls to client-supplied functions via `/v1/responses` and continue when results come back.
Added five alpha MCP servers for calculations, network diagnostics, package and model metadata, news feeds, and research sources. Agents can check the outside world before making up an answer. Novel concept.
Added MiniMax M3.1 Flash and paid Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed. All three support image input, tools, reasoning, and 1,048,576-token context. Polli also gained `polli harness vibe on` for Mistral Vibe.
Added clearer community model signals. If over half a model's eval questions hit errors, Trust labels it 'endpoint errors' instead of a score gap. Community Watch sends each incident separately on Discord and pings linked publishers. Score gaps go to maintainers.
Shipped this week: community-model Trust comparisons, daily usage/speed/success boards, My Apps submissions with 24-hour counts, and clearer failure messages. From Nov 1, `pk_` identifies OAuth apps; API calls need `sk_` or a user-authorized key.
Added scheduled probes and owner warnings for struggling public community models. My Models points owners to exact-ID health checks. Private community models are filtered from public status and stats. A surprisingly necessary distinction.