Skip to content

feat(catalog)!: refresh lineup for the Fable 5.1 / GPT-6 Astra generation - #15

Merged
dorukozgen merged 30 commits into
mainfrom
feat/catalog-sept-2026-refresh
Sep 21, 2026
Merged

dorukozgen merged 30 commits into
mainfrom
feat/catalog-sept-2026-refresh

Conversation

@dorukozgen

Copy link
Copy Markdown
Contributor

Catalog pass for everything that shipped since the 2026-08-01 refresh, sourced from each provider's own documentation (and, for OpenRouter, NVIDIA NIM and GitHub, from their live model listings). 166 → 184 models.

Added

Provider New
anthropic, claude-code Claude Fable 5.1
codex-cli GPT-6 Astra
gemini-api Gemini 3.8 Flash, Gemini 3.7 Flash
antigravity-sdk Gemini 3.8 Flash and 3.7 Flash, each High / Medium / Low
grok Grok 4.6, Grok Build 0.1
deepseek deepseek-flash (V4.1 Flash, the new canonical id)
mistral Z.ai GLM 5.3
kimi-api Kimi K2.7 Code HighSpeed
openrouter Fable 5.1, GPT-6 Astra, Gemini 3.8 Flash, Grok 4.6, DeepSeek V4.1 Flash, GLM 5.3
github-copilot Fable 5.1, Fable 5, GPT-6 Astra, Gemini 3.8 Flash, Grok 4.6, Kimi K3, MAI-Code-1.1-Flash
perplexity Grok 4.6, Gemini 3.8 / 3.6 Flash, Fable 5, Kimi K3, GLM 5.3, Nemotron 3 Ultra
cloudflare-workers-ai GLM 5.3, GLM 5.3 Flash, DeepSeek V4 Pro 0813, DeepSeek V4 Flash 0731, Qwen 3.8 27B
nvidia-nim Kimi K3, GLM 5.3, GLM 5.3 Flash, Nemotron 3 Ultra, Nemotron 3.5 Lightning

Fixed

  • Claude Sonnet 5 is $2 / $10, not $3 / $15. Anthropic made the launch price permanent on 2026-08-10. Fixed on anthropic, openrouter, github-copilot and perplexity; the expired "introductory pricing" note is gone.
  • Grok 4.5 rejects reasoning_effort: none and defaults to high. Grok 4.3 gained xhigh. Grok 4.20 has a 1M context, not 2M.
  • GPT-5.4 / 5.4 mini / 5.4 nano / 5.5 do not accept minimal; their levels are none … xhigh.
  • GPT-5.6 Sol dropped to $4 / $20 (promotional, guaranteed through 2026-11-21). Gemini 3.6 Flash bills $0.75 / $3.75 through 2026-12-31.
  • Mistral Large pointed at the retired 2411 snapshot (404s); now Large 3. Codestral is 25.08 (128K, closed weights). Small 4 takes images. Medium 3.5 and Small 4 declare reasoning_effort.
  • Kimi K2.7 Code HighSpeed was an alias of K2.7 Code but is priced at 2×; it now has its own entry. Kimi K2.6 / K2.7 / K3 reject custom temperature.
  • DeepSeek V4 Pro GA pricing; both DeepSeek models declare low / high / max effort.
  • Nemotron 3 Super on NIM has a 1M context, not 256K.
  • OpenRouter prices synced with its live API; Cloudflare paid-plan notes.

cost records the price billed today; scheduled changes and peak/off-peak rules live in fleetify.note.

Removed (breaking)

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

  • github-copilot: 11 models in GitHub's retirement history (Opus 4.5 / 4.6, Sonnet 4.5, Gemini 3.1 Pro / 2.5 Pro / 3 Flash, Raptor mini, GPT-4.1, GPT-5.2, GPT-5.2-Codex) plus GPT-5.4 nano, which was never selectable in the CLI
  • nvidia-nim: 6 ids no longer in /v1/models (gpt-oss-120b, MiniMax M2.7, GLM 5.1, DeepSeek V4 Flash / Pro, Llama 4 Maverick)
  • perplexity: Gemini 2.5 Flash / Pro, nvidia/nemotron-3-super, xai/grok-4-1-fast-non-reasoning
  • kimi-api, kimi-code: kimi-k2.5 (discontinued 2026-08-31, 404s)
  • deepseek: deepseek-v4-flash → deepseek-flash (old slug kept as an alias)
  • codex-cli: gpt-5.4-mini (retired from Codex 2026-08-31)
  • antigravity-sdk: Gemini 3.5 Flash tiers (gone from the Antigravity picker)

The only default that moves is deepseek → deepseek-flash.

Deliberately not done

  • GPT-6 Astra is not on the openai provider. OpenAI's model guide: "GPT-6 Astra supports Chat Completions, but tool calling requires Responses." The app's OpenAI bridge speaks Chat Completions, so listing it would offer an agent that cannot call tools.
  • Perplexity Sonar stays even though Sonar Chat Completions is supported only until 2026-09-27 and sonar-pro is the provider default. Picking a new default is a separate change.
  • kimi-code and antigravity-cli are structurally out of date (the Kimi Code CLI now uses k3 / kimi-for-coding ids; agy now has agy models and --model). Both depend on how the app's bridges pass model ids, so they are left for a follow-up.

Verification

bun run validate   ✓ 17 providers, 184 models
bun run build      ✓ dist/ regenerated
bun test           ✓ 31 pass, 0 fail

Fable 5.1 (`claude-fable-5-1`) shipped 2026-09-01 and is the target of Claude
Code's `fable` / `best` aliases from v2.1.257 on. Same $10 / $50 list price and
1M / 128K limits as Fable 5, but cache reads bill at 0.025x base input ($0.25)
instead of the usual 0.1x.

Opus 5 stays the default on both providers, matching Anthropic's own account
defaults; Fable 5.1 slots directly beneath it. Fable 5 is still served (legacy,
retirement not before 2027-06-09) and keeps its entry.

Source: platform.claude.com models overview, pricing and Fable 5.1 pages;
code.claude.com model-config.
Anthropic made the launch price permanent on 2026-08-10: the increase to
$3 / $15 scheduled for 2026-09-01 was cancelled. Record $2 / $10 with the
matching cache rates ($0.20 read, $2.50 write) and drop the "introductory
pricing through 2026-08-31" note, which now says the opposite of what is true.
claude-code inherits both through `extends`.

Also fill in Fable 5's knowledge cutoff (2026-01), which the entry lacked.
Grok 4.6 went live on the xAI API on 2026-08-12: 500K context, $2 / $6 with
$0.50 cached input, knowledge cutoff 2026-02-01. It always reasons, so its
effort levels run low -> xhigh (default high) with no `none`. xAI documents no
text output limit for it, so `limit.output` is left unset rather than guessed.

Grok Build 0.1 is xAI's coding model (256K, $1 / $2) and the redirect target
of the retired `grok-code-fast-1` slugs, recorded here as aliases. It reasons
but exposes no effort control, hence `effort.kind = "none"`.

Grok 4.3 stays the default. Source: docs.x.ai models page and release notes.
- Grok 4.5 does not accept `reasoning_effort: none` and defaults to high, not
  low. The old level list let the picker offer an effort the API rejects.
  Levels are now low / medium / high / xhigh. Release date is 2026-07-08.
- Grok 4.3 gained `xhigh`.
- Grok 4.20 (both variants) has a 1M max prompt length, not 2M.
- Grok 4.3 and 4.20 take text and jpg/png images only; drop the `pdf` input
  modality, which xAI does not list for any text model.
- Record the `-latest` aliases for 4.3 and 4.5.
…ash`

DeepSeek released V4.1 Flash on 2026-09-10 under a new canonical id,
`deepseek-flash`, and retired V4 Flash and the short-lived V4 Flash Vision Exp
the same day. The old slugs still resolve upstream, so they are kept as
aliases rather than as separate entries. V4.1 Flash adds native image input
and a `reasoning_effort` control (low / high / max, default high).

V4 Pro is still served as the 0813 GA snapshot: DeepSeek announced it would
reroute to V4.1 Flash on 2026-09-14, then reversed that in the changelog. Its
pricing moved with GA, so refresh it and declare the same effort levels.

DeepSeek now bills peak and off-peak rates. `cost` records the peak list price
and a note carries the off-peak halving, since the schema has one price tier.

The default moves to `deepseek-flash`: it is the model DeepSeek now recommends
and V4 Pro is being phased out ahead of a V4.1 Pro.

BREAKING CHANGE: the `deepseek-v4-flash` model id is gone from the deepseek
provider. Sessions pinned to it heal to the provider default.

Source: api-docs.deepseek.com pricing, news260910, updates, thinking_mode.
OpenAI released GPT-6 Astra on 2026-09-03: 1.05M context, 128K output, $10 /
$50 with $1 cached input and $12.50 cache writes, knowledge cutoff 2026-04-30.
It rejects the `none` effort and custom `temperature`.

The Codex docs list it for the CLI (`codex -m gpt-6-astra`) but not for Codex
cloud, and gate its Enterprise rollout, so GPT-5.6 Sol stays the default and
Astra sits directly beneath it.

Not added to the `openai` API provider yet: OpenAI's model guide states that
"GPT-6 Astra supports Chat Completions, but tool calling requires Responses",
and the app's OpenAI bridge speaks Chat Completions. Listing it there would
offer a coding agent that cannot call tools.

There is no `gpt-6-astra-pro` API id (the model page 404s); "Pro" is the
`reasoning.mode = "pro"` setting on the same model.

Source: developers.openai.com model page, model guide, changelog; Codex models.
- GPT-5.6 Sol dropped to $4 / $20 ($0.40 cached) on 2026-08-21. OpenAI calls
  it promotional pricing guaranteed through 2026-11-21, so a note carries the
  expiry and the previous $5 / $30 rate. Applies to openai and codex-cli.
- GPT-5.4, 5.4 mini, 5.4 nano and 5.5 do not accept `reasoning.effort:
  minimal`. Their documented levels are none / low / medium / high / xhigh.
  The picker default stays `medium`, which all four accept.
- codex-cli's GPT-5.5 has the same 1.05M context window as the API model; the
  two entries disagreed (400K vs 1.05M).
OpenAI retired GPT-5.4 and GPT-5.4 mini from Codex on 2026-08-31; GPT-5.6 Luna
is the documented replacement and is already in the lineup. GPT-5.4 itself left
this provider in the previous refresh. The model is still served on the API, so
the `openai` entry is untouched.

GPT-5.5 is scheduled to leave Codex on 2026-10-14 and stays until then.

BREAKING CHANGE: `gpt-5.4-mini` is gone from the codex-cli provider. Sessions
pinned to it heal to the provider default.
…ntigravity-sdk

Both went GA on the Gemini API (3.7 on 2026-08-13, 3.8 on 2026-09-02) with a
1,048,576 / 65,536 token window and a March 2026 knowledge cutoff. They accept
`thinking_level` low / medium / high, default medium; `minimal` is rejected.

Antigravity's model picker offers each at Low, Medium and High, so
antigravity-sdk gets the usual three tier-suffixed slugs per model. The
`-high` and `-medium` slugs appear in the documented `agy models` output; the
`-low` slugs follow the same pattern as the existing 3.6 Flash entries.

`cost` records the $0.75 / $3.75 introductory rate that is billed today, with a
note for the scheduled move to $1.50 / $7.50 on 2027-01-01.

There is still no Pro model newer than 3.1 Pro Preview, which stays the
default on both providers.

Source: ai.google.dev models, pricing, changelog and thinking docs; DeepMind
model cards; antigravity.google docs/models and CLI docs.
The catalog carried $1.50 / $7.50 / $0.15, which is the rate Google has
scheduled for 2027-01-01. The pricing page bills half that until the end of
the year. Record today's rate on gemini-api and the three antigravity-sdk
tiers, with a note for the scheduled increase.

Also fill in Gemini 3.5 Flash-Lite's knowledge cutoff (2026-03, per its model
card), which the entry lacked.
Antigravity's model picker now lists Gemini 3.8, 3.7 and 3.6 Flash, Gemini 3.1
Pro, the two Claude 4.6 thinking models and GPT-OSS 120B. Gemini 3.5 Flash is
absent from both the plan-availability table and the selector, so its three
tier slugs go. The model is still served on the Gemini API and keeps its
gemini-api entry.

BREAKING CHANGE: `gemini-3.5-flash-high`, `-medium` and `-low` are gone from
the antigravity-sdk provider. Sessions pinned to them heal to the default.
Mistral started hosting Z.ai's GLM 5.3 on 2026-09-15 as a third-party,
open-weights model served without modification: 1M context, 128K output,
$1.40 / $4.40 with $0.14 cached input, text only, function calling and
structured outputs. It is in public preview, so the entry uses the pinned id
`zai-glm-5-3` and records `zai-glm-5` / `zai-glm-latest` as aliases.

Mistral documents no reasoning-effort control for it, hence
`effort.kind = "none"`.

Mistral has shipped no first-party frontier model since Medium 3.5
(2026-04-28), and the Devstral and Magistral families are fully retired.

Source: docs.mistral.ai model pages, changelog and lifecycle policy.
- `mistral-large-latest` resolves to Mistral Large 3 (`mistral-large-2512`,
  released 2025-12-02). The entry still named the 2411 snapshot as an alias,
  which Mistral retired on 2026-05-31 and which now 404s. Drop the knowledge
  cutoff, which described 2411 and is not published for Large 3.
- `codestral-latest` is Codestral 25.08 (2025-07-30): 128K context rather than
  256K, and a Premier licence, so `open_weights = false`.
- Mistral Small 4 takes image input.
- `mistral-medium-3-5-26-04` is a docs-page slug, not an API id. The real ids
  are `mistral-medium-3-5` and `mistral-medium-3`.
- Medium 3.5 and Small 4 accept `reasoning_effort` (none ... xhigh). Default to
  `high`, which Mistral recommends for agentic and code use.
- Cached prompt tokens bill at 10% of input on every model; add `cache_read`.
`kimi-k2.7-code-highspeed` was recorded as an alias of `kimi-k2.7-code`, but it
is a separately listed and separately priced model: the same weights served
faster at $1.90 / $8.00 ($0.38 cached), twice the base rate. As an alias it
would have been billed at half its real price. Give it its own entry and drop
the alias.

While in `kimi-k2.7-code`: record its release date (2026-06-12) and that the
model rejects any `temperature` other than its fixed 1.0.

Source: platform.kimi.ai pricing, models and models-overview pages.
Moonshot fixes `temperature` per model (K3 at 1.0; K2.6 at 1.0 thinking / 0.6
non-thinking) and returns an error for any other value, so `temperature` is
false. kimi-code inherits this through `extends`.

Also record K3's documented 1,048,576 max output (131,072 when unset).
Moonshot discontinued `kimi-k2.5` on 2026-08-31 together with the
`moonshot-v1-*` family; calls now return 404 "model not found". K2.6 remains
the default on both providers.

BREAKING CHANGE: `kimi-k2.5` is gone from kimi-api and kimi-code. Sessions
pinned to it heal to the provider default.
Claude Fable 5.1, GPT-6 Astra, Gemini 3.8 Flash, Grok 4.6, DeepSeek V4.1 Flash
and GLM 5.3, all listed on OpenRouter's public /api/v1/models today. Ids,
prices, context and max output come from that endpoint; release dates and
knowledge cutoffs come from each vendor's own docs where they publish them.

The list stays curated to one or two current models per vendor, so Gemini 3.7
Flash (superseded by 3.8 three weeks later at the same price), the GPT-6 Astra
Pro routing id and the dated DeepSeek snapshots are left out.

GPT-6 Astra is fine here even though it is held back from the `openai`
provider: OpenRouter lists `tools` among its supported parameters and
translates to the upstream API itself.
Six entries had drifted from what OpenRouter's /api/v1/models reports on
2026-09-21:

- anthropic/claude-sonnet-5   $3 / $15        -> $2 / $10 (now the standard rate)
- google/gemini-3.6-flash     $1.50 / $7.50   -> $0.75 / $3.75
- openai/gpt-5.6-sol          $5 / $30        -> $2 / $10
- moonshotai/kimi-k3          $3 / $15        -> $1.70 / $8.50
- moonshotai/kimi-k2.6        $0.684 / $3.42  -> $0.95 / $4.00
- deepseek/deepseek-v4-pro    $0.435 / $0.87  -> $0.9553 / $1.9105

Cached-input rates move with them. The other ten entries already match.
Claude Fable 5.1, Claude Fable 5, GPT-6 Astra, Gemini 3.8 Flash, Grok 4.6,
Kimi K3 and MAI-Code-1.1-Flash. Each is GA in GitHub's supported-models table
and marked "Included" in its Copilot CLI column. Ids follow the CLI's
lowercased-display-name convention; `claude-fable-5.1` and `kimi-k3` are
spelled out in the copilot-cli release notes.

Prices are GitHub's own per-token rates. GitHub publishes no context or output
limits, so Claude, GPT and Gemini reuse the windows already recorded for their
siblings here, while Grok 4.6 and Kimi K3 carry their vendor's window.
MAI-Code-1.1-Flash has pricing only.

Skipped on purpose: Gemini 3.7 Flash and Grok 4.5 (GitHub retires both on
2026-10-19) and Kimi K2.7 Code (retires 2026-10-02).

Source: docs.github.com supported-models and models-and-pricing; GitHub
changelog; github/copilot-cli releases.
… 3.6 Flash

GitHub's per-token pricing table now lists Claude Sonnet 5 at $2 / $10,
GPT-5.6 Sol at $4 / $20 and Gemini 3.6 Flash at $0.75 / $3.75, following the
upstream vendors. Cached-input and cache-write rates move with them.
GitHub's model retirement history lists these as already retired:

- 2026-09-01: Claude Opus 4.5, Claude Opus 4.6, Claude Sonnet 4.5,
  Gemini 3.1 Pro, Raptor mini
- 2026-07-31: Gemini 2.5 Pro, Gemini 3 Flash
- 2026-06-01: GPT-4.1 (now a non-selectable utility model), GPT-5.2,
  GPT-5.2-Codex

GPT-5.4 nano was never selectable in Copilot CLI: every client column reads
"Not included" and GitHub lists it as a utility model.

Claude Sonnet 4.6 stays. It carries the same 2026-09-01 date but GitHub keeps
it available to annual Pro and Pro+ subscribers and it is still in the
supported table. Models with a future retirement date also stay for now:
Claude Opus 4.7, Gemini 3.5 and 3.6 Flash (2026-10-02); GPT-5.5, GPT-5.4,
GPT-5.4 mini and GPT-5 mini (2026-10-19).

BREAKING CHANGE: eleven model ids are gone from the github-copilot provider.
Sessions pinned to them heal to the provider default (`auto`).
Grok 4.6, Gemini 3.8 Flash, Gemini 3.6 Flash, Claude Fable 5, and three
open-weights models Perplexity hosts itself under a `perplexity/` prefix:
Kimi K3, GLM 5.3 and Nemotron 3 Ultra. Ids and per-token prices are copied
from Perplexity's Agent API models page.

Perplexity does not publish context or output limits, so entries carry the
upstream vendor's window as the rest of this provider already does. Nemotron 3
Ultra has pricing only. Note that Perplexity bills Gemini 3.6 Flash at $1.50 /
$7.50, without the promotion Google and OpenRouter apply.

Perplexity does not offer GPT-6 Astra or Claude Fable 5.1.

Source: docs.perplexity.ai agent-api/models and changelog.
Perplexity's Agent API table lists $2 / $10 / $0.20, matching Anthropic's now
permanent rate. Drop the expired introductory-pricing note with it.
- google/gemini-2.5-flash: removed from the API on 2026-03-20
- google/gemini-2.5-pro: removed on 2026-04-01
- nvidia/nemotron-3-super-120b-a12b: gone from the Agent API list; the only
  Nemotron there now is perplexity/nemotron-3-ultra-550b-a55b
- xai/grok-4-1-fast-non-reasoning: gone from the list, and xAI retired the
  model itself on 2026-05-15

The Sonar family is untouched even though Perplexity will stop supporting
Sonar Chat Completions on 2026-09-27: it still works today, and replacing the
provider default is a separate decision.

BREAKING CHANGE: four model ids are gone from the perplexity provider.
Sessions pinned to them heal to the provider default.
GLM 5.3 and GLM 5.3 Flash (1.31M context), DeepSeek V4 Pro 0813 and V4 Flash
0731 (Cloudflare's first 1M-context models) and Qwen 3.8 27B, all added to
Workers AI between 2026-08-14 and 2026-08-28 with function calling and
reasoning. GLM 5.3 Flash and Qwen 3.8 27B also take images.

Ids, context windows and unit prices come from Cloudflare's model pages and
pricing table. Four of the five need the Workers Paid plan, which a note
records; Qwen 3.8 27B does not. Cloudflare publishes no output limit or
JSON-mode row for these models, so both are left unset.

Llama 3.3 70B stays the default. GLM 5.2 is skipped: 5.3 replaced it at the
same price.

Source: developers.cloudflare.com workers-ai models, pricing and changelog.
…ions

- Kimi K2.6 and Kimi K2.7 Code have required the Workers Paid plan since
  2026-07-28; on the Free plan they return 403. Note it on both.
- Kimi K2.7 Code reached Workers AI on 2026-06-12; the entry had no dates.
- GLM-4.7-Flash input is $0.060 per MTok, not $0.0605.

All 19 existing ids are still in Cloudflare's live model index.
Kimi K3, GLM 5.3, GLM 5.3 Flash, Nemotron 3 Ultra 550B and Nemotron 3.5
Lightning 30B. All five are in NVIDIA's public /v1/models listing today, with
1,048,576-token context windows per their build.nvidia.com cards. Kimi K3's
card documents low / high / max reasoning effort and a 65,536 `max_tokens`
cap.

NVIDIA publishes no per-token pricing for the hosted API, so `cost` follows
the 0.00 convention the other free-tier entries here already use.

Nemotron 3 Super stays the default.

Source: integrate.api.nvidia.com/v1/models; build.nvidia.com model cards.
build.nvidia.com reports `contextLength: 1048576` for
nvidia/nemotron-3-super-120b-a12b. The entry said 262,144, and it is the
provider default. Cloudflare's copy of the model legitimately stays at 256K,
which is Cloudflare's own serving limit.
None of these ids is in NVIDIA's public /v1/models listing any more.

Deprecated with a dated notice on build.nvidia.com:
- openai/gpt-oss-120b      2026-09-02
- minimaxai/minimax-m2.7   2026-07-27
- z-ai/glm-5.1             2026-07-02

Absent without a notice:
- deepseek-ai/deepseek-v4-flash, deepseek-ai/deepseek-v4-pro (no DeepSeek V4
  id of any spelling is served)
- meta/llama-4-maverick-17b-128e-instruct

The provider uses live discovery, so anything NVIDIA brings back will still
surface in the picker.

BREAKING CHANGE: six model ids are gone from the nvidia-nim provider. Sessions
pinned to them heal to the provider default.
Regenerated with `bun run build` after the September 2026 catalog refresh: 17 providers, 184 models.
@dorukozgen
dorukozgen merged commit 9fdac01 into main Sep 21, 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