Skip to content

🤖 feat(models): move gpt/sol/astra to GPT-6.2 Sol and Astra - #8172

Merged
ThomasK33 merged 2 commits into
mainfrom
gpt-6-2-astra-sol
Nov 17, 2026
Merged

ThomasK33 merged 2 commits into
mainfrom
gpt-6-2-astra-sol

Conversation

@ThomasK33

@ThomasK33 ThomasK33 commented Nov 17, 2026 •

Copy link
Copy Markdown
Member

Why

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 2026-11-17. This moves the gpt/sol aliases to openai:gpt-6.2-sol and astra to openai:gpt-6.2-astra, with the reasoning policy, pricing, service tiers, Codex OAuth routing and docs each model's page documents (Sol, Astra, pricing).

What changed

  • thinking.ts: one OPENAI_GPT_TIERS table (rejectsNoneEffort, ultrafast, flex) keyed by API id replaces the Astra and 6.1 Sol regexes. Both 6.2 models reject none, so off clamps to low; native max and Pro mode apply as on every GPT-6 tier.
  • Service tiers: Ultrafast is offered on both. Flex is now dropped (Standard) for the 6.2 models, which have no Flex row; other models keep the route-only gate.
  • Pricing (models-extra, serviceTierPricing): Standard and long-context rates from the pricing page, Fast 2x, Ultrafast at the 6x list factor.
  • Cyber: the Daybreak map moves from KnownModelDefinition to an exact-id table in cyberMode.ts, so gpt-6-astra/gpt-6.1-sol keep Cyber after the aliases move. The Daybreak guide lists no 6.2 value, so Cyber is hidden for 6.2 and its requests omit the field (standard safeguards).
  • Codex OAuth: both 6.2 ids allowed, with an assumed 272K cap.
  • gpt-6-astra alias → gpt-6.2-astra (as grok-4.6 → grok-4.7). GPT-6 Astra and GPT-6.1 Sol stay usable as custom model strings and are offered as retired built-ins.
  • Docs: generated model table, models/providers pages, Settings help example.

Tradeoffs

  • Ultrafast is 3x through 2027-02-28 and 6x from 2027-03-01. Xum prices 6x (the "never under-count" rule), so cost tracking shows 2x the real 6.2 Ultrafast cost until March. No date-dependent pricing.
  • The Codex catalog (openai/codex@c08843a) has no 6.2 row yet. 272K copies the other GPT-6 rows; a too-low cap only compacts earlier. Without the allowlist entries, OAuth-only users of the moved aliases would fail with api_key_not_found.

Blast Radius

Anyone on gpt, sol or astra switches model on upgrade. A provider-level flex tier no longer reaches those aliases, and Cyber disappears for them (pick openai:gpt-6.1-sol or openai:gpt-6-astra). No fallback chains or seeded migrations reference these aliases. Every external assumption, tagged, is in the evidence comment.

Verification

  • make static-check, make typecheck, make lint-eslint, make crap-check, make dry-check pass.
  • New tests: wire-level Flex/Ultrafast tier per model (providerModelFactory), Flex gate incl. mapped aliases and OpenRouter (providerOptions), doc-fed Standard/long-context and tier pricing through the real cost calculator (displayUsage), policy, Cyber, Codex cap, retired tokenizers.
  • The 85 test files that touch these models pass, apart from failures that origin/main shows in this sandbox too (file-watch, unreadable-file and EIO tests).

Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: xhigh

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 2026-11-17. The `gpt`/`sol`
aliases now resolve to openai:gpt-6.2-sol and `astra` to openai:gpt-6.2-astra
(the `gpt-6-astra` version alias becomes `gpt-6.2-astra`, as grok-4.6 -> 4.7 did).

- thinking.ts: one OPENAI_GPT_TIERS table (rejectsNoneEffort, ultrafast, flex)
  replaces the per-model Astra / 6.1 Sol regexes.
- Service tiers: Ultrafast is offered on both 6.2 models; Flex is dropped for
  them (no Flex row on the pricing page). Pricing: Fast 2x, Ultrafast at the 6x
  list factor (over-counts 2x during the 3x promo through 2027-02-28).
- Cyber: the Daybreak program map moves to an exact-id table in cyberMode.ts;
  6.2 is not in the guide, so Cyber stays on gpt-6-astra and gpt-6.1-sol only.
- Codex OAuth: both 6.2 ids allowed, 272K cap ASSUMED (no catalog row yet).
- GPT-6 Astra and GPT-6.1 Sol stay usable as custom model strings with their
  pricing, tiers (Flex included), reasoning policy and Codex routing.

---

_Generated with [`xum`](https://github.com/coder/xum) • Model: `anthropic:claude-opus-5-5` • Thinking: `xhigh`_
@mintlify

mintlify Bot commented Nov 17, 2026 •

Copy link
Copy Markdown

Preview deployment for your docs. Learn more about Mintlify Previews.

Project Status Preview Updated
Mux 🟢 Ready View Preview Nov 17, 2026, 6:41 PM

💡 Tip: Enable Automations to automatically generate PRs for you.

@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 17, 2026 •

Copy link
Copy Markdown
Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-11-17T19:12:02.028656Z b2faad7 Manual request
🔒 Security Review ✅ Completed 2026-11-17T19:13:55.133377Z b2faad7 Manual request
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@ThomasK33

Copy link
Copy Markdown
Member Author

External assumptions this PR encodes

Tags: observed (I checked it on 2026-11-17), documented (OpenAI docs, fetched 2026-11-17), assumed (no source yet).

Fact Tag Source
Model ids gpt-6.2-astra, gpt-6.2-sol; no dated snapshots listed documented Astra, Sol
Both reject reasoning.effort: "none"; migrate none/minimal to low documented model pages, latest-model guide
Effort ladder low..max (native max) assumed family behavior (gpt-6-astra, gpt-6.1-sol); 6.2 pages do not list efforts
Pro reasoning mode on the same id documented reasoning guide (gpt-6.2-sol mode: "pro" example)
Tool calling needs Responses; Chat Completions has none documented latest-model guide
1,050,000 window, 922,000 max input, 128,000 max output; cutoffs Astra 2026-05-18, Sol 2026-06-01 documented model pages
Standard: Astra $8 / $40 / cached $0.40 / 30m write $10; Sol $2 / $10 / $0.10 / $2.50. Above 272K input: input and cache 2x, output 1.5x documented model pages, pricing
Fast 2x both bands; Batch 0.5x (Xum does not use Batch) documented model pages
No Flex for either 6.2 model (absent from the Flex table; gpt-6-astra and gpt-6.1-sol are in it) documented pricing
Ultrafast on both: 3x through 2027-02-28, 6x from 2027-03-01 (Pacific). Priced at 6x documented (rate) / decision (6x) model pages
Explicit prompt-cache breakpoints ("GPT-5.6 and later") documented prompt caching
Daybreak table has no 6.2 row; models from gpt-6.1-terra on default to standard when access_programs.cyber is omitted documented Daybreak guide
Both models served in Codex for ChatGPT sign-in documented (announcement) OpenAI release-day announcement
Codex catalog has no 6.2 slug at main = c08843a5bd16de5524129b2f4270490447991984 (2026-11-17T17:29Z) observed codex-rs/models-manager/models.json
Codex OAuth context cap 272K for both assumed copied from the other GPT-6 catalog rows
@ai-sdk/openai 4.0.71 parses gpt-6.2-* as GPT-6+ (reasoning model, efforts low..max, no sampling params) observed getOpenAILanguageModelCapabilities in node_modules/@ai-sdk/openai/dist/index.js
GPT-6.2 drops reasoning items on prefix edits, model downgrade, end-user or org change documented (no code change) latest-model guide

Test runs

Per-file bun test over the 85 files that touch these models, branch vs origin/main in the same sandbox: failures that differ are in config.test.ts (file watch), aiService.test.ts (unreadable inheritance file), workspaceService.goalContinuation.test.ts (EIO between tests) and streamManager.usage.test.ts (aborted-stream sidecar). Those files fail on origin/main here too and reference none of the changed models; CI is the judge for them.

@ThomasK33

Copy link
Copy Markdown
Member Author

@codex review

@chatgpt-codex-connector

Copy link
Copy Markdown

🛡️ Codex Security Review · Automatically triggered

Security review completed. No security issues were found in this pull request.

Reviewed commit: 58b1c076f3

View security finding report

Only the user who started this review can view the report in Codex.

ℹ️ About Codex security reviews in GitHub

This is an experimental Codex feature. Security reviews are triggered when:

  • You comment "@codex security review"
  • A regular code review gets triggered (for example, "@codex review" or when a PR is opened), and you’re opted in so security review runs alongside code review

Once complete, Codex will leave suggestions, or a comment if no findings are found.

@github-actions

Copy link
Copy Markdown
Contributor

No job left output: make test-stories stopped before its first job. See the run log.

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 58b1c076f3

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Comment thread src/common/types/thinking.ts Resolved
Comment thread docs/config/providers.mdx Resolved
@ThomasK33

Copy link
Copy Markdown
Member Author

Independent verification: PASS+NOTES

Head 58b1c076f34be5111b3428ee9013924d9465fb2e (base origin/main 31c14d50). I did not write this change.

Lane Result
make typecheck pass
make static-check pass (0 lint errors, formatting clean)
bun test on the 12 touched src/ test files 1181 pass, 1 fail. The fail (discoverAvailableSubagentsForToolContext > fails instead of publishing unverified metadata…, ssh: Connection refused) fails the same way on origin/main in this environment, so it is not new.
jest tests/ui/agents/thinkingSelector.test.ts 1/1 pass
Alias resolution (resolveModelAlias), main → head sol/gpt: gpt-6.1-sol → gpt-6.2-sol; astra: gpt-6-astra → gpt-6.2-astra; gpt-6.2-astra → openai:gpt-6.2-astra; gpt-6-astra no longer resolves (see note 1)
Request body (real ProviderModelFactory + buildProviderOptions, stubbed fetch, 24 cases per SHA) 6.2 Sol/Astra: all on /v1/responses; off → effort: low (never none); max → max; flex → no service_tier; ultrafast and priority sent; Cyber sends no access_programs. gpt-6.1-sol and gpt-6-astra bodies are identical to main (flex, ultrafast, daybreak_blue). On main, 6.2 ids sent flex, dropped ultrafast and sent max as high.
Cost (real createDisplayUsage, 100K and 300K input with cache reads and 30m writes) 48/48 totals match OpenAI's published pricing tables for 6.2 Sol/Astra at Standard, Fast, Ultrafast-list and Flex→Standard, plus 6.1 Sol and 6 Astra on both SHAs
Live xum run -m sol / -m astra (scratch XUM_ROOT, -t LOW, workspace OpenAI gateway) Both accepted; replied OK; Xum model openai:gpt-6.2-sol / openai:gpt-6.2-astra; priced at $0.0229 / $0.0916 (matches the 30m cache-write and output rates). The NDJSON does not carry the API's echoed model id.

Docs: I checked the PR's claims against the OpenAI pages fetched today (model pages, pricing, Daybreak guide, Ultrafast guide, latest-model guide). Pricing, the 922K input / 128K output limits, knowledge cutoffs, the rejection of none, the missing Flex row, the 3x-then-6x Ultrafast promo, the lack of a Daybreak row for 6.2 and "tool calling requires Responses" for 6.2 all match.

Notes (not blocking)

  1. src/common/constants/knownModels.ts:152: the bare alias gpt-6-astra is gone. xum run -m gpt-6-astra and a task or workflow model: gpt-6-astra now fail with invalid-format. An agent's frontmatter model: gpt-6-astra is dropped without a warning (coerceModel in resolveAgentAiSettings.ts:201) and the inherited model runs instead. This follows the grok-4.7 precedent, and nothing built in uses that alias.
  2. src/common/constants/codexOAuth.ts:142-143: the 272K Codex OAuth caps for 6.2 are assumed, as the code comment says. Replace them once the Codex catalog has rows for 6.2.
  3. src/common/utils/tokens/serviceTierPricing.ts:157-158: Ultrafast on 6.2 is priced at the 6x list rate, so Xum shows twice the billed cost until 2027-02-28. This is intentional and documented.
  4. Cyber: when sol, gpt or astra resolve to 6.2, Cyber stops applying. A chat with Cyber on that picks one of those aliases now sends no access_programs field, so it runs with standard safeguards. Chats that already saved openai:gpt-6.1-sol or openai:gpt-6-astra keep daybreak_blue. The docs say this.

Evidence (diff, logs, probe JSON, OpenAI pages) is kept locally in the verifier workspace.


Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: high

The test relied on GPT-6 Astra's 1.05M input window to give OpenAI a
model larger than the default Opus 5.5 (1M). GPT-6.2 Astra caps input at
922K, so with only OpenAI configured no built-in qualifies and the UI shows
the manual /compact button instead. Seed Google (1,048,576) instead.

---

_Generated with [`xum`](https://github.com/coder/xum) • Model: `anthropic:claude-opus-5-5` • Thinking: `xhigh`_
@ThomasK33

Copy link
Copy Markdown
Member Author

@codex review

Round 1: both findings were answered with OpenAI doc citations and resolved, with no code change. Flex is still not listed for GPT-6.2 on the pricing page or the model pages, and the GPT-6.2 migration guide says tool calling requires Responses. New commit b2faad7 fixes the auto-compaction UI test: with GPT-6.2 Astra's 922K input cap, no OpenAI built-in exceeds Opus 5.5's 1M, so the fixture now seeds Google.

@chatgpt-codex-connector

Copy link
Copy Markdown

Codex Review: Didn't find any major issues. 👍

Reviewed commit: b2faad7d08

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

@chatgpt-codex-connector

Copy link
Copy Markdown

🛡️ Codex Security Review · Automatically triggered

Security review completed. No security issues were found in this pull request.

Reviewed commit: b2faad7d08

View security finding report

Only the user who started this review can view the report in Codex.

ℹ️ About Codex security reviews in GitHub

This is an experimental Codex feature. Security reviews are triggered when:

  • You comment "@codex security review"
  • A regular code review gets triggered (for example, "@codex review" or when a PR is opened), and you’re opted in so security review runs alongside code review

Once complete, Codex will leave suggestions, or a comment if no findings are found.

@ThomasK33

Copy link
Copy Markdown
Member Author

Pausing the PR loop: the only remaining blocker needs a human (AGENTS.md early-stop case 4).

On head b2faad7d08, Codex approved with no findings, every review thread is resolved, and every check passes except two jobs. Required fails only because of them:

  • Test / Integration (1/3, 2/3, 3/3): all 18 failing files call Anthropic models and fail with invalid x-api-key (HTTP 401). No OpenAI case fails, so the live gpt-6.2-sol calls in sendMessage.basic and sendMessage.context pass.
  • Test / Stories (Haiku): sandbox refused: ANTHROPIC_API_KEY rejected by api.anthropic.com (HTTP 401).

The same 401 fails these jobs on #8171, #8169 and #8165. The last green integration run I found is on #8170 (2026-11-16 08:41 UTC). Someone with repo admin needs to rotate the ANTHROPIC_API_KEY Actions secret, then rerun the failed jobs on this head.


Generated with xum • Model: anthropic:claude-opus-5-5 • Thinking: xhigh

@ThomasK33
ThomasK33 merged commit 70ab64f into main Nov 17, 2026
35 of 41 checks passed
@ThomasK33
ThomasK33 deleted the gpt-6-2-astra-sol branch November 17, 2026 19:46
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