Skip to content

Add GPT-6.2 Sol and GPT-6.2 Astra - #193

Merged
wch merged 1 commit into
mainfrom
add-gpt-6-2-sol-astra
Nov 18, 2026
Merged

wch merged 1 commit into
mainfrom
add-gpt-6-2-sol-astra

Conversation

@wch

@wch wch commented Nov 18, 2026 •

Copy link
Copy Markdown
Collaborator

OpenAI's gpt-6.2-sol and gpt-6.2-astra, and their Amazon Bedrock Mantle equivalents, are not known to ai-lib. They show up as raw IDs, and because the generic GPT-6 rule keeps "off", consumers offer a thinking level the models reject: OpenAI's model pages say reasoning effort none is not supported, and Bedrock Mantle sends "off" as the wire value none. This adds curated names, the correct reasoning levels (low through max) and Bedrock limits for both models.

The companion posit-dev/assistant PR adds pricing and picker status and bumps this submodule pin. Merge this one first.

Why this approach

The GPT-6 capability profile already matches any gpt-6* ID and is shared by direct OpenAI and Bedrock, so this extends the existing no-"off" regex to gpt-6.2-sol and gpt-6.2-astra instead of adding a new rule. Both providers stay in sync by construction. I did not flip the rule to "every GPT-6 model except an allow-list": unknown future IDs, including bare gpt-6.2, keep "off" on purpose until someone documents them.

The OpenAI offline fallback list now stores only IDs, and buildOpenAIModel gets every name from getOpenAIModelName, so each display name is written in one place and neither caller passes one.

What changed

  • Reasoning levels: GPT6_NO_OFF_MODEL in gpt6-model-profile.ts also matches the two 6.2 IDs, still only at a model-name boundary.
  • Bedrock limits: both models are added to GPT6_PUBLISHED_LIMITS (1.05M window, 128K output), taken from their AWS model cards.
  • Names and fallback: "GPT-6.2 Sol" and "GPT-6.2 Astra" in openai-model-names.ts, and both IDs in the OpenAI fallback list.
  • Docs: the GPT-6 paragraphs in memory-bank/aiConfig.md.

What to watch for

  • OpenAI's overview lists efforts low through max, while its "Using GPT-6.2" guide names only none/minimal/low/medium. Live requests settle it: on OpenAI and on Bedrock Mantle (us-east-1), both models accept low through max and reject none and minimal.
  • On 2026-11-18, Mantle's model listing in us-east-1 did not include either model, though inference worked. Discovery skips unlisted IDs, so the new Bedrock limits take effect only once AWS lists them. I accepted this rather than adding a static Mantle list.
  • As with the other no-"off" models, other consumers of the shared OpenAI capability table also stop offering "off" for these IDs.

Testing

  • npm test in ai-config (802 tests) and ai-provider-bridge (1111 tests) passes. The new infer.test.ts case, which checks that both models get no "off" on OpenAI and Bedrock, failed before the change. Per the table-row rule, there are no per-row tests for the new names, fallback IDs or limits.
  • Live checks against OpenAI and Bedrock Mantle (us-east-1), including a tool round trip and raw-HTTP capture through Posit Assistant, are described in the companion PR.

@wch
wch merged commit dec91af into main Nov 18, 2026
4 checks passed
@wch
wch deleted the add-gpt-6-2-sol-astra branch November 18, 2026 22:58
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