feat(shared,host-service): add GPT-6 Astra to the Codex model picker and pricing table

Merged
AV
2mo ago
+46-4

Summary

  • Add GPT-6 Astra (gpt-6-astra, launched 2026-09-03) to the Codex model picker and the Superset chat catalog, and let it take Codex's max reasoning effort.
  • Price it in the usage history table at $10 / $50 per MTok (cached input $1, which is the default 0.1x), and bump the "pricing table updated" date shown on Settings → Usage.

Why / Context

GPT-6 Astra is OpenAI's new flagship and the pickers only offer curated ids. Without the pricing row, Codex sessions run on Astra would be costed at the cheapest Codex rate and flagged approximate.

How It Works

  • gpt-6-astra is the only GPT-6 id in OpenAI's API docs and the slug Codex's model docs publish. It sits first in the picker's Current group.
  • Effort: OpenAI documents low through max for reasoning.effort, so Astra joins the max allowlist. Codex's ultra (max plus subagent delegation) stays off until a live catalog entry confirms it, because a wrong allowlist entry breaks the launch instead of hiding an option.
  • Pricing uses longest-prefix match, so openai-codex/gpt-6-astra from OMP or opencode logs also resolves through the slash segment (covered by a test).

How the id was verified

  • OpenAI API model page and the Codex models doc.
  • Live per-account catalog: codex app-server → model/list with this login on Codex 0.152.1 and 0.153.0 still lists only the GPT-5.6 family, GPT-5.5, and the retiring GPT-5.4 pair. OpenAI's rollout is phased (Trusted Access Program today, Plus/Pro/Business/Enterprise "in the coming days").
  • models.dev and the bundled OMP catalog have no gpt-6 entry; cursor-agent's --list-models needs a login.

Manual QA

  • Not launched end to end: this account's Codex catalog does not include Astra yet, so a launch fails on OpenAI's side today. Picker rendering, grouping, and the effort filter are covered by the unit tests. Once the rollout lands, choosing Astra + Max should start codex --model gpt-6-astra -c model_reasoning_effort=max.

Testing

  • bun test packages/shared/src/agent-models.test.ts (52 pass)
  • bun test packages/host-service/src/trpc/router/usage/history/pricing.test.ts (6 pass)
  • bunx biome check on the four changed files
  • bun run typecheck in packages/shared and packages/host-service

Known Limitations

  • Astra shows in the picker before every account can use it; until OpenAI enables the account, the launch errors server-side.
  • Not added to the opencode, OMP, cursor-agent, or Copilot pickers: their catalogs do not carry the id yet and they reject unknown ids at launch.

Follow-ups

  • Add ultra for Astra once model/list shows it.
  • The Codex picker's "Retiring 2026-08-31" group is past its date and gpt-5.3-codex no longer appears in the live catalog; drop them in a separate PR.

https://claude.ai/code/session_011FkRSxPP9wCUU1psiLfvn5

Summary by cubic

Adds GPT-6 Astra (gpt-6-astra) to the Codex model picker and Superset chat catalog, priced at $10/$50 per MTok in the usage history table. Without the pricing row, Astra sessions on Codex were costed at the cheapest rate and flagged approximate.

  • Supports Codex's max reasoning effort; ultra stays off until the live catalog confirms it.
  • The picker shows Astra before every account has access, so launching errors server-side until OpenAI enables the account.
  • Bumps the usage history pricing table's updated date to 2026-09-03.

Summary by CodeRabbit

  • New Features

    • Added GPT-6 Astra to the available Superset Chat and Codex model catalogs.
    • GPT-6 Astra supports low through maximum reasoning effort levels in Codex.

  • Billing

    • Added GPT-6 Astra usage pricing at $10 per million input tokens and $50 per million output tokens.
    • Updated the pricing information date to September 3, 2026.