GPT-6 Sol and GPT-6 Luna: catalog rows, family prompt preset, and default-model support #2023

Summary

OpenAI's GPT-6 family is Astra, Sol and Luna (https://developers.openai.com/api/docs/guides/latest-model). senpi 2026.9.22-4 supports Astra only: the catalog ships gpt-6-astra rows, the GPT-6 prompt preset matches only Astra ids, and the recommended/default OpenAI model is still gpt-5.6-sol.

Reproduction

1. senpi --provider openai --model gpt-6-sol (or gpt-6-luna, or the same ids under chatgpt-subscription) on 2026.9.22-4.
2. The model is not in the built-in catalog, so it is unselectable through the first-party providers.
3. Through OpenRouter, openai/gpt-6-sol and openai/gpt-6-luna do exist (models.dev regeneration pulled them) but carry no effort ladder, so xhigh / max cannot be selected and the row sits at the raw 922k window.
4. Even with a hand-written models.json row, the session runs on the generic prompt: resolvePresetName returns undefined for gpt-6-sol / gpt-6-luna while gpt-6-astra gets the GPT-6 preset.

Expected (ideal state)

• gpt-6-sol and gpt-6-luna are first-class catalog rows on openai and chatgpt-subscription (plus Azure / OpenCode / OpenRouter / Venice / Vercel passthrough rows), with published pricing (Sol $2/$10, Luna $0.10/$0.50 per 1M tokens, >272k tiers), 128k output, text + image, tool search / additional tools, Priority -fast variants, and the documented ladder none/low/medium/high/xhigh/max (no minimal; unlike Astra, off stays selectable).
• Prompt budgets: Luna ships the full 922k input cap, Sol ships 400k, Astra keeps 600k, stamped on every provider row like Astra already is.
• Map-less custom rows infer xhigh / max from the id.
• The whole GPT-6 family renders the GPT-6 system prompt (one prompting guide for the family).
• gpt-6-sol joins the recommended-model ladder and becomes the implicit default for the OpenAI providers; the high-reasoning warning covers it at xhigh / max.

Actual

See Reproduction: Astra-only catalog, OpenRouter rows without a ladder, generic prompt for Sol/Luna, gpt-5.6-sol defaults.

Evidence

• packages/ai/src/providers/data/openai.json / chatgpt-subscription.json at v2026.9.22-4: no gpt-6-sol / gpt-6-luna keys (control: gpt-6-astra present in 7 shards).
• packages/ai/src/providers/data/openrouter.json at v2026.9.22-4: openai/gpt-6-sol, openai/gpt-6-luna (+-pro, :batch) with thinkingLevelMap absent, contextWindow: 922000.
• packages/coding-agent/src/core/extensions/builtin/prompt-preset/presets.ts: hasGpt6AstraSignal matches gpt-6-astra only.
• packages/coding-agent/src/core/extensions/builtin/recommended-models/index.ts: ladder is astra, gpt-5.6-sol, ...; model-resolver.ts defaultModelPerProvider.openai = gpt-5.6-sol.

Root cause

Every GPT-6 branch in the generator, the preset matcher and the defaults was written against the Astra id alone when Astra was the only GPT-6 model.

Scope / acceptance criteria

• Catalog rows + -fast variants + ladder + budgets as above, regenerated with --strict; packages/ai/test/gpt-6-family-catalog.test.ts green; the input-cap invariant still rejects a 1,050,000 row.
• models.ts id inference for map-less Sol/Luna rows.
• Preset resolution for every Sol/Luna id shape; rendered prompt identical to Astra's.
• Recommended ladder, provider defaults and the high-reasoning warning cover gpt-6-sol.
• Changelog: Added (first-party rows, preset, defaults) and Fixed (OpenRouter ladder/budget).

Related

• feat(coding-agent): recommend GPT-6 Astra ahead of GPT-5.6 Sol #1375 (recommend Astra ahead of Sol), fix(ai): ship GPT-6 Astra at its documented 1,050,000-token context window #1377 (Astra context window), fix(coding-agent): warn for Astra only at max reasoning #1567 (Astra warning only at max), Context overflow is not failure-proof: one persisted flag disables every recovery path, goal re-prompts a deterministic overflow, astra contextWindow overstates the 922k input cap #1422 (input-cap invariant).