Skip to content

feat(catalog)!: drop retired models and move defaults off closed ones - #18

Merged
dorukozgen merged 1 commit into
mainfrom
fix/catalog-nov-2026-retirements
Nov 15, 2026
Merged

dorukozgen merged 1 commit into
mainfrom
fix/catalog-nov-2026-retirements

Conversation

@dorukozgen

Copy link
Copy Markdown
Contributor

Follow-up to #17: remove the models each vendor has retired, or retires within the next five days, and move the defaults that pointed at deprecated or closed models. Every removal is sourced from the vendor's own retirement notice or live model list. 269 → 218 models.

Removed (breaking)

Sessions pinned to a removed id heal to the provider default.

Provider Removed Why
github-copilot (9) claude-opus-4.7, gemini-3.5-flash, gemini-3.6-flash Retired 2026-10-02
gpt-5-mini, gpt-5.4, gpt-5.4-mini, gpt-5.5 Retired 2026-10-19
claude-opus-5, gpt-5.6-terra Retire 2026-11-16
cloudflare-workers-ai (12) deepseek-r1-distill-qwen-32b, deepseek-v4-flash-0731, deepseek-v4-pro-0813, llama-3.1-8b-instruct-fp8, llama-3.2-1b-instruct, llama-3.2-3b-instruct, kimi-k2.6, kimi-k2.7-code, nemotron-3-120b-a12b, qwen3-30b-a3b-fp8, qwq-32b, glm-5.3-flash Return 410 Gone from 2026-11-20
nvidia-nim (4) nvidia/nemotron-3-super-120b-a12b Deprecated 2026-10-19 ("no longer supported after 10/19/2026")
moonshotai/kimi-k3, mistralai/mistral-nemotron, z-ai/glm-5.3 Gone from /v1/models
antigravity-sdk (3) claude-opus-4-6-thinking, claude-sonnet-4-6-thinking, gpt-oss-120b-medium Removed 2026-11-18
perplexity (21) openai/gpt-5.4, gpt-5.2, gpt-5.1, gpt-5-mini Retired 2026-10-24, no longer accepted
anthropic/claude-{fable-5, fable-5-1, opus-5, sonnet-5, opus-4-8, opus-4-7, opus-4-6, opus-4-5, sonnet-4-6, sonnet-4-5, haiku-4-5}, google/gemini-{3.8-flash, 3.6-flash, 3.5-flash, 3.1-pro-preview, 3-flash-preview}, perplexity/nemotron-3-ultra-550b-a55b No longer on the Agent API model list
gemini-api (2) gemini-3.7-flash, gemini-3.5-flash Deprecated; Google routes them to 3.8 / 3.6 Flash, which now list them as aliases

Copilot's Opus 4.8 and Sonnet 4.6 stay: GitHub keeps both for individual subscribers on annual plans (#15 kept Sonnet 4.6 for the same reason). Their notes now say so.

Defaults

Provider Before After Why
nvidia-nim nvidia/nemotron-3-super-120b-a12b nvidia/nemotron-3-ultra-550b-a55b NVIDIA's own flagship and GA (Nemotron 3.5 Super VL is still early access). Moves to the top of the list.
antigravity-sdk gemini-3.1-pro-high gemini-4-hydrogen-high Gemini 3.1 Pro is Enterprise-only now. Hydrogen runs on every plan and with a GEMINI_API_KEY, and High is its default thinking level.
kimi-api, kimi-code kimi-k2.6 kimi-k3.1-pro K2.6 is closed to accounts created after 2026-10-31, and Kimi Code no longer lists it. kimi-code inherits the default from kimi-api.

Every other default is unchanged. Perplexity keeps sonar-pro: synchronous Sonar requests still work, reformulated as Agent API requests.

Fixed

  • mistral: Mistral Large 4 is GA (version 26.10, 2026-11-05) at $0.70 / $2.10 with $0.07 cache reads, and its weights are published. Its effort levels are now low / medium / high (it rejects the catalog's old none), default high. mistral-large-latest resolves to it, so the alias moves off Large 3.
  • codex-cli: Codex offers no ultra tier for GPT-5.6 Luna (models.json lists low … max).

Not done (next pass)

  • github-copilot: claude-fable-5, claude-haiku-4.5 (annual plans keep it), gemini-3.8-flash, gemini-4-argon, grok-4.6 and kimi-k3 retire on 2026-12-14. Copilot Fusion (fusion + --fusion-tier) does not map onto the auto entry's effort levels.
  • gemini-api: gemini-2.5-pro and gemini-2.5-flash route to 3.8 Flash / 3.5 Flash-Lite from 2026-12-01.
  • openai: gpt-5 and gpt-5-mini shut down on 2026-12-11. GPT-5.6 Sol's promotional price now runs until 2026-11-30.
  • mistral: the default mistral-large (Large 3) and mistral-medium (Medium 3.5) are deprecated and retire on 2027-04-30. zai-glm-5-3 is GA and takes low / high / max effort.
  • kimi-api: K2.6 and K2.7 Code stay for accounts created before 2026-11-01 (retirement expected in late December). Kimi K3's API default effort is now max, and it bills $3 cache writes.
  • kimi-code: subscription ids (k3.1-pro, k3, kimi-for-coding) still differ from the catalog's API-key ids.
  • antigravity-sdk: the agy models sample shows no -low tiers for Gemini 4 Argon and the 3.x models. That is unverified, so they stay.
  • claude-code: the claude-opus-5 200K override and the claude-opus-5[1m] twin are redundant now that Opus 4.7 and later run 1M natively.
  • codex-cli: Codex defaults GPT-6 Astra and GPT-5.6 Sol to low effort; the catalog says medium.

Verification

bun run validate   ✓ 17 providers, 218 models
bun run build      ✓ dist/ regenerated
bun test           ✓ 31 pass, 0 fail

Remove the models each vendor has retired, or retires within the next five
days, and move the defaults that pointed at deprecated or closed models.
269 -> 218 models. Sessions pinned to a removed id heal to the provider
default.

Removed:
- github-copilot (9): claude-opus-4.7, gemini-3.5-flash, gemini-3.6-flash,
  gpt-5-mini, gpt-5.4, gpt-5.4-mini, gpt-5.5 (retired 2026-10-02 and
  2026-10-19); claude-opus-5, gpt-5.6-terra (retire 2026-11-16)
- cloudflare-workers-ai (12): the catalog models Cloudflare turns off on
  2026-11-20 (410 Gone)
- nvidia-nim (4): nemotron-3-super-120b-a12b (deprecated 2026-10-19), and
  kimi-k3, mistral-nemotron and glm-5.3 (gone from /v1/models)
- antigravity-sdk (3): Claude Opus 4.6 and Sonnet 4.6 (Thinking), GPT-OSS
  120B (removed 2026-11-18)
- perplexity (21): GPT-5.4, 5.2, 5.1 and 5 Mini (retired 2026-10-24), and
  17 ids that are no longer on the Agent API model list
- gemini-api (2): gemini-3.7-flash and gemini-3.5-flash. Google routes them
  to 3.8 and 3.6 Flash, which now carry them as aliases.

Defaults:
- nvidia-nim: nemotron-3-super-120b-a12b -> nemotron-3-ultra-550b-a55b
- antigravity-sdk: gemini-3.1-pro-high (now Enterprise-only) ->
  gemini-4-hydrogen-high
- kimi-api, kimi-code: kimi-k2.6 (closed to new accounts) -> kimi-k3.1-pro

Fixed:
- mistral: Large 4 is GA at $0.70 / $2.10 with published weights. Its
  effort levels are low / medium / high (`none` is rejected), and
  `mistral-large-latest` moves to it from Large 3.
- codex-cli: GPT-5.6 Luna has no `ultra` tier.
- github-copilot: Opus 4.8 and Sonnet 4.6 stay for subscribers on annual
  plans, and their notes now say so.
@dorukozgen
dorukozgen merged commit fb580a2 into main Nov 15, 2026
2 checks passed
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