Community model owners can now set eligible Pollinations models as fallbacks when an upstream fails. My Models shows 7 days of health data, groups failures, and lets owners rerun a proxy model's registration test. Less guessing at what broke.

Added Sonilo sound effects to the audio API. Give it a public video URL; it returns a separate MP3 track timed to the video. Live transcription now starts without a client session update. Less session choreography.

Added static-site hosting at <name>.pollinations.app. Publish with `polli sites deploy`, the SDK, MCP tools, or the Sites API. Barbara can publish a site she built and return its URL. 'Works on my machine' needed a domain name.

Model Monitor now shows community-model Trust beside Health, with sortable comparisons to the official models they're named after. Model health also separates Pollen-related 402s from other 4xx responses. A billing error is not a model review.

Enter now has My Apps for signed-in people with catalog listings on their linked GitHub account. Activity charts, totals, navigation, and CSVs now use your browser's timezone. Midnight was doing too much.

merged: Messages and the CLI now show refusal-only replies. Blank non-streaming chat completions get rejected before caching, so the existing fallback or error path handles them. A refusal and an empty answer are different bugs.

This week: static-site hosting at <name>.pollinations.app. Publish via Polli CLI, SDK, MCP, or Sites API. Barbara can publish a built page and return its URL. Sites are public. Kling 3.0 Standard and FLUX 3 Max also joined the video catalog.

Merged support for Barbara to read every message in Telegram groups set to Every message. She's instructed to reply to unmentioned messages only when useful. Restraint is a feature.

merged: Scout now tries pinned DeepInfra before Novita if its primary provider fails, for chat and Responses. Economics keeps the last completed snapshot through failed refreshes and offers Retry on long loads. added Complimentik and FX Trade Setup Studio to the catalog.

this week we added `Prefer: respond-async` for image, video, audio & 3D jobs: fetch status and results later. sandboxes can extend past a day; daily restarts preserve files, memory and processes, but drop open connections. even servers need naps.

Added Kling 3.0 Standard for 720p video from text or a start frame (3–15s, optional audio). FLUX.3 Image Klein can generate or edit images with up to 10 references. The mood board has become an input format.

added JavaScript to Computer MCP. agent code can use persistent files and return a JSON-encoded result. /v1/systemone now aliases TypeSafe decisions on /alpha/decisions. the naming problem remains undefeated.

added async jobs for image, video, audio and 3D: submit a request, get a job ID, then poll for status and fetch the result. retries keep the original owner; cache hits resolve pending jobs. the loading spinner can take a break.

merged SDK support for 3D generation, model status, and speech word timings. React gets generation hooks and a streaming useChat transport. empty chat streams no longer get cached as answers, so the next request can reach a model.

added a sandbox page in Enter: create, manage, and connect from the dashboard. both templates now include ffmpeg, uv, cron, and rclone; the large one gets more disk. a toolbox is better when it contains tools.

added a REST-based Agent Skill so agents can call Pollinations without a CLI install. docs serve Markdown on request; SDK and CLI text streams now reject incomplete output instead of returning half an answer. the other half was important.

this week: added /v1/responses for Chat text models, preserved parallel Gemini tool calls, and rebuilt Play for image/video/audio with shareable inputs + API URLs. model search and filters landed too. fewer detours, same number of tabs.

added sandbox renewals past E2B's 24-hour limit (subject to key budget), a polli CLI timeout option, and an opt-in 4-vCPU/8-GB template. also stopped legacy text caches from storing full prompts. longer jobs; shorter memory.

added MAI-Voice-2.1 and Flash TTS (97 voices each), multilingual transcription with timed segments, live PCM transcription, and GPT-6 Luna Decisions for streaming text/JSON decisions. the audio menu is getting crowded.

added /v1/responses for every Chat text model, including Claude, Gemini, community models and chat-only agents. Community endpoints can now return plain JSON; streaming clients still get valid events. less protocol paperwork.