DX-2986: add Claude Fable 5.1 and GPT-6 Astra, correct spark and nano picker entries

[alitariksahin]

What

Adds Claude Fable 5.1 and GPT-6 Astra to the SDKs and the model picker, and corrects two picker entries that no longer match what the backend charges.

Changeset included: patch bump on @upstash/box and @upstash/box-cli.

Claude Fable 5.1

Available on the Claude API, OpenRouter, Vercel AI Gateway and OpenCode Zen. Verified against each catalog rather than assumed.

ClaudeCode.Fable_5_1                anthropic/claude-fable-5-1
OpenRouterModel.Claude_Fable_5_1    openrouter/anthropic/claude-fable-5.1
VercelModel.Claude_Fable_5_1        vercel/anthropic/claude-fable-5.1
OpenCodeModel.Claude_Fable_5_1      opencode/claude-fable-5-1
OpenCodeModel.Zen_Claude_Fable_5_1  opencode/claude-fable-5-1

Note the spelling split: the Claude API and Zen use claude-fable-5-1, while the OpenRouter and Vercel gateways list it as claude-fable-5.1. Confirmed against OpenRouter's live /api/v1/models and the Vercel gateway catalog.

Eight picker entries, mirroring exactly where Fable 5 already appears. CursorModel is deliberately untouched, for both models: Cursor publishes no catalog I could check, so an entry there would be a guess.

Same members added to the Python SDK, plus the model table in the SDK README.

GPT-6 Astra

Added to every catalog that carries it, each one checked rather than assumed. Unlike Fable 5.1 there is no spelling split: every provider uses gpt-6-astra.

OpenAICodex.GPT_6_Astra      openai/gpt-6-astra              OpenAI model page
OpenRouterModel.GPT_6_Astra  openrouter/openai/gpt-6-astra   OpenRouter live catalog
VercelModel.GPT_6_Astra      vercel/openai/gpt-6-astra       Vercel gateway catalog
OpenCodeModel.GPT_6_Astra    opencode/gpt-6-astra            Zen pricing page

Six picker entries, mirroring where the current OpenAI flagship already appears. Same members in the Python SDK, plus the SDK README table.

Not added: gpt-6-astra-pro, which OpenRouter lists at a price identical to plain astra while appearing on no OpenAI page and not on Zen; and gpt-6-astra-fast, which is a 2x billing mode rather than a model id on the OpenAI API, exposed as one only by the Vercel gateway.

Picker corrections

OpenAICodex.GPT_5_3_Codex_Spark removed from the picker. It is a ChatGPT research preview, not an API model: its docs page 404s and the coordinator no longer prices it, so selecting it produced a run against a model that does not exist and would bill at the unknown-model default.

OpenCodeModel.GPT_5_3_Codex_Spark is deliberately kept. Same name, different product: Zen sells it and publishes a price.

GPT-5 Nano moved from the free group to the paid group and lost its (Free) suffix. Zen now charges $0.05 / $0.40 per MTok, so the picker was advertising free while the backend billed paid. Big Pickle is still genuinely free and stays.

Not a breaking change

Both spark enum members remain exported from the TypeScript and Python SDKs. Removing them would break consumers, so they should be deprecated with JSDoc in a separate PR. Taking the option out of the picker is what stops users selecting it.

Verification

• pnpm ci:lint passes across all four workspace projects
• tsc --noEmit clean on packages/sdk and packages/cli
• packages/sdk unit tests: 30 files, 431 passed
• packages/cli unit tests: 43 files, 479 passed
• Python types.py parses, with every added member carrying the intended value

Note for local checkouts: packages/cli resolves @upstash/box through a workspace symlink to packages/sdk/dist, so the CLI will not typecheck against a stale SDK build. Run pnpm build first. CI does this already.

Related

This is the picker and SDK half of DX-2986. Pricing for both models is in upstash/box-backend#246 and the console display mirror is a separate PR on DX-2986 in upstash-console-v2. The backend PR should land first; this one removes the (Free) label that the backend change would otherwise contradict.

[linear-code]

DX-2986