Skip to content

feat: add GPT Image 2.6, key-aware media schemas and model updates - #436

Merged
ProfSynapse merged 1 commit into
mainfrom
feat/gpt-image-2.6-media-catalogs
Nov 5, 2026
Merged

ProfSynapse merged 1 commit into
mainfrom
feat/gpt-image-2.6-media-catalogs

Conversation

@ProfSynapse

Copy link
Copy Markdown
Owner

Summary

Started as adding GPT Image 2.6 and grew into three connected changes. They touch the same catalogs and files, so they ship together.

GPT Image 2.6

  • Sunburst and Flare on OpenAI and OpenRouter.
  • New complexity setting on both. On OpenAI direct only: SVG and GIF output through outputFormat, and markup-guided edits through markupImage. OpenRouter refused SVG and GIF when tested.
  • SVG gets a 300 s deadline, because a real Sunburst SVG took 144 s against the old 120 s limit.
  • Same token rate as 2.5, but about 3.3x the tokens per image (1,449 vs 439 at 1024x1024 medium). No default moved, and retired image ids still map to 2.5.

Schemas follow the keys you have

  • generateImage, generateVideo, generateAudio and image requests in executePrompts now build their parameters and limits from the catalogs, filtered to providers with a key that are switched on. With only OpenRouter, an agent never sees outputFormat or markupImage.
  • Each image model declares its aspect ratios, checked before any request.
  • numberOfImages is gone from the schema, because no adapter ever made more than one image. JSON callers that still send it get a note saying it was ignored. The CLI form now gets an "Unknown flag" error.
  • A wrong audio file extension now fails before synthesis is paid for.

Model updates

  • Removed for shutdown: Google's three TTS previews and gemini-3.1-flash-live-preview (Nov 17), gpt-realtime-mini (Jan 20), minimax-m2.7 (expired), mimo-v2-omni, mimo-v2-pro, gpt-5.1-codex*, gpt-5.2-codex, Groq's qwen3.6-27b and mistral-saba-latest. Saved speech and realtime selections move to their successors. A saved chat selection naming a removed model has no automatic mapping and needs re-picking.
  • ElevenLabs music and SFX now send music_v2_5 and eleven_text_to_sound_v3. They sent no model before, so they ran on defaults being removed in January and March. Voice gains v4, v4 Turbo, v3, v3 Conversational and Flash v2.5.
  • New image models: MAI-Image-2.6, Grok Imagine Image 2.0, Qwen Image 3. New video: FLUX.3 Video, Hailuo 3, Wan 3.0, plus 2k in the video resolution type.
  • New chat models: gpt-6.1-terra on OpenAI, Codex and OpenRouter. On OpenRouter also Gemini 4 Argon, Kimi K3.1 Pro, DeepSeek V4.1 Flash, MiMo V2.6 and MiniMax M3.1 Flash.
  • Prices refreshed from provider listings. gpt-5.6-sol, the OpenAI default, was listed at $5/$30 and is now $4/$20.

Bugs fixed along the way

  • tts-1 and tts-1-hd offered four voices their API rejects.
  • A settings change removed generateImage for OpenAI-only setups.
  • Changing ElevenLabs in Apps did not update generateAudio.
  • OpenRouter video sent negative prompts in an undocumented field.

Testing

  • Jest: 7,033 passed, 0 failed, 118 skipped. Lint, mobile check, npm run build, schemas:check and the registry gate all pass.
  • New guards were mutation-checked: removing each one fails a test.
Live through shipped code Result
GPT Image 2.6 on OpenAI: generate, edit, markup, SVG, GIF Pass, outputs inspected by eye
GPT Image 2.6 on OpenRouter, including complexity reaching OpenAI Pass
MAI-Image-2.6, Grok Imagine Image 2.0, Qwen Image 3 Pass
FLUX.3 Video, Hailuo 3 at 2K, Wan 3.0 Pass, one short clip each
gpt-6.1-terra on OpenAI, and all 8 new or corrected OpenRouter chat ids Pass
Gemini 3.8 Flash and Flash-Lite TTS, and a removed id mapping to its successor Pass
OpenAI tts-1 and gpt-4o-mini-tts Pass

Not tested live:

  • ElevenLabs voice, music and SFX (no key available).
  • Any model's behaviour in the running Obsidian plugin.
  • The Veo negative prompt passthrough, which follows OpenRouter's docs.
  • gpt-6.1-terra on Codex, which is confirmed only by Codex's model list.
  • Removals from Groq and Mistral, which rest on their docs.

Worth checking in review

  • Some refreshed OpenRouter prices look odd but match the listing, such as glm-5.2 at $0.006 input and kimi-k3 at $0.192 input.
  • Negative prompts on Seedance, Grok and Runway through OpenRouter now fail before billing instead of being sent.
  • schemas/5.21.3/ is regenerated in place, as the catalog protocol directs.

Follow-up

OpenAI's TTS models shut down January 6, 2027, and their replacement needs a new transport. They stay for now, tracked in #435.

🤖 Generated with Claude Code

Add GPT Image 2.6 Sunburst and Flare on OpenAI and OpenRouter, with
complexity, SVG and GIF output, and markup-guided edits.

Build the generateImage, generateVideo, generateAudio and executePrompts
schemas from the model catalogs, filtered to providers with a key, so an
agent only reads rules it can use. Check per-model aspect ratios before a
request and drop numberOfImages, which no adapter honoured.

Remove models past or near shutdown, mapping saved speech and realtime
selections to their successors. Pin ElevenLabs music and sound effects to
current models instead of deprecated API defaults. Add new image, video and
chat models and refresh prices from provider listings.

Co-Authored-By: Claude Opus 5.5 <[email protected]>
@ProfSynapse
ProfSynapse merged commit 9560784 into main Nov 5, 2026
1 check passed
@ProfSynapse
ProfSynapse deleted the feat/gpt-image-2.6-media-catalogs branch November 5, 2026 12:46
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant