fix(models): GPT-5.6 (Sol/Terra/Luna) + refresh OpenRouter's catalogue against the live API

[ericleepi314]

Summary

Refreshes OpenRouter's curated catalogue against the live API and adds OpenAI's current frontier generation, GPT-5.6.

GPT-5.6 (Sol / Terra / Luna). Three durable capability tiers on one generation rather than a size ladder: Sol is the flagship, Terra balances capability against cost, Luna is the cheap high-volume tier, and gpt-5.6 is OpenAI's alias for Sol. Offered by the direct provider and by OpenRouter, which also carries a -pro variant of each tier.

Ids came from OpenAI's own API docs rather than stripping OpenRouter's openai/ prefix — the direct API takes the bare id, and that equivalence needed confirming rather than assuming.

Model configs are the non-obvious half. Adding ids to a provider list isn't enough: without a ModelConfig these resolved through get_model_config's prefix fallback to the 272K catch-all, firing auto-compact roughly three quarters of a 1.05M window early.

Placement there is load-bearing. The fallback matches on key.rsplit("-", 1)[0], so the bare gpt-5.6 alias has base gpt and would become the catch-all for every unknown gpt id if it preceded gpt-5.5, handing them a 1.05M window. Over-estimating overflows the context; under-estimating only compacts early. So the catch-all stays on the 272K entry and the alias is registered after it, while the tier keys (base gpt-5.6) go first — which is what lets an unlisted variant like gpt-5.6-sol-pro inherit the right window. Exact lookups are unaffected by position.

OpenRouter's list had gone stale, and two entries were dead. The OpenAI section still led with gpt-5 / gpt-4o / o1, and openai/o1-mini had been delisted upstream entirely — a row the picker offers that fails at request time, the same "offers something you cannot select" shape as the picker bug in #772.

Validating the rest against https://openrouter.ai/api/v1/models caught five more dead ids: anthropic/claude-3.5-sonnet, anthropic/claude-3.5-haiku, google/gemini-2.0-flash, meta-llama/llama-3.1-405b-instruct, deepseek/deepseek-v3.2-speciale, x-ai/grok-2. All replaced with live successors. Every one of the 46 ids in the catalogue now resolves, including the default model.

Batch (:batch), image, audio and embedding variants stay out — none serves an interactive agent turn, and free-text ids are accepted regardless of this list.

Two duplicate lists removed. PROVIDER_INFO["openrouter"] was duplicated verbatim in OpenRouterProvider._curated_models(), and PROVIDER_INFO["openai"] in OpenAIProvider.get_available_models(). Each pair feeds different surfaces — the registry drives login and the /model picker, the copies drive discovery — so editing one shows up as a model one surface offers and the other drops. Both now read the registry. The subscription branch keeps its own list, since a ChatGPT plan serves a smaller set.

Deliberately unchanged

• default_model stays gpt-5.4. The default is a cost-sensible tier, not the frontier — Anthropic likewise defaults to claude-sonnet-4-6 while listing claude-fable-5 first. Moving it would silently change cost and behaviour for everyone without an explicit --model.
• SUBSCRIPTION_MODELS untouched. Which models the ChatGPT-subscription backend serves is a wire fact sourced from Codex CLI's cache. Guessing there breaks subscription sessions rather than showing a stale row.

Test Coverage

New tests/providers/test_openrouter_catalog.py (107 tests): the single-sourcing of both lists, ids vendor-qualified for OpenRouter and bare for the direct API, no :batch variants, the flagship-first tier ordering, the 1.05M windows, and the catch-all invariant.

Mutation-verified — each fix reverted and the intended test confirmed to fail:

Mutation	Test that catches it
Move the gpt-5.6 alias above gpt-5.5	test_an_unknown_gpt_id_still_gets_the_conservative_catch_all
Restore either duplicate list	test_provider_reads_the_registry_rather_than_a_second_copy
Order Luna ahead of Sol	test_the_5_6_tiers_are_ordered_flagship_first
Re-add the delisted openai/o1-mini	test_openai_section_leads_with_the_current_frontier

The tests deliberately do not hit the network: CI would otherwise fail whenever a vendor retires a model, which is a fact about the world rather than a defect in the diff. The file documents the revalidation command instead.

Test plan

• Python suite: 9373 passed, 3 skipped, 0 failed
• Every curated OpenRouter id verified resolvable against the live API
• gpt-5.6, -sol, -terra, -luna resolve to 1.05M; gpt-5.6-sol-pro inherits it; an unknown gpt id still gets 272K

🤖 Generated with Claude Code