List the GPT-6.1 Sol Bedrock Runtime profiles in model discovery
Refs #3113
Summary
Adds openai.gpt-6.1-sol to the Bedrock Runtime capability catalog, so discovery offers its us. and global. inference profiles. On main, a Bedrock connection lists only the Mantle base id openai.gpt-6.1-sol. global.openai.gpt-6.1-sol and us.openai.gpt-6.1-sol are ACTIVE in AWS, but discovery only offers Runtime profiles of catalogued models, and this one was not catalogued.
The AWS model card lists the Standard tier, short context (up to 272K input tokens), per 1M tokens: Global CRIS $2.00 in / $10.00 out; US CRIS and Mantle regional $2.20 / $11.00. The Global profile is the cheaper option the issue asks for.
What changes
bedrock-runtime-capabilities.ts: one catalog entry (both APIs,max_completion_tokens,rejectsReasoningNone: true, the model card assourceUrl,verifiedAt: 2026-11-05).- Specs: 34 added lines, no existing test touched.
- Changeset (
manifest, patch).
rejectsReasoningNone is set because of what the Runtime does with this model (below). With the Responses route from #3110, Chat Completions requests with function tools to these profiles go to Runtime Responses.
Measured on Bedrock Runtime (global.openai.gpt-6.1-sol, us-east-1, 2026-11-05)
| Request | Result |
|---|---|
| Chat Completions, no tools | 200 |
Chat Completions with max_tokens | 400 (max_completion_tokens only) |
Chat Completions, reasoning_effort low / medium / high / xhigh | 200 |
Chat Completions, reasoning_effort none or minimal | 400 (the error lists low, medium, high, xhigh, max as supported) |
| Chat Completions, function tools, no effort | 400 ("use /v1/responses") |
Chat Completions, function tools, effort high | 400 (same message) |
Chat Completions, function tools, effort none | 400 (none unsupported) |
| Responses, no tools | 200 |
Responses, function tools, with and without reasoning | 200 |
Responses, reasoning.effort none | 400 |
This is the same behaviour as openai.gpt-6-astra, which already has rejectsReasoningNone. For us.openai.gpt-6.1-sol I measured only the no-tools Chat Completions request directly (200); its other cases ran through Manifest (below).
Through Manifest
This branch, local build, fresh database, a Bedrock API key. Discovery lists openai.gpt-6.1-sol, global.openai.gpt-6.1-sol and us.openai.gpt-6.1-sol next to the three gpt-6-sol ones. POST /v1/chat/completions returned 200 for both profiles in all four cases: no tools, and function tools with no effort, high and max. The three tool cases returned a tool call.
Not verified, and limits
- The price in the dashboard. Prices show as
nullon my local instance, forgpt-6-solas well, so I could not see the $2.00 / $10.00 option appear. The price data comes from theamazon-bedrockprovider of models.dev, which listsglobal.openai.gpt-6.1-solat 2 / 10. This PR makes the profile available; the reporter or a maintainer should confirm the price in the UI. - An explicit
reasoning_effort: "none"with function tools still returns 400: the model rejectsnoneon Responses too. - On the Responses route the conversion drops sampling parameters such as
temperature,top_p,stopandseed, as described in #3110. - Requests above 272K input tokens are billed at the long-context rates in the model card; nothing changes in the code for that.
- I did not check whether the two new ids can be picked automatically when tiers are recalculated.
- The same gap exists for
openai.gpt-6.1-terra,openai.gpt-5.6-solandopenai.gpt-5.6-terra(ACTIVE profiles, not catalogued). Each needs its own verification, so they are left out of this PR.
Tests
Backend unit: 523 suites, 9952 tests passing, with bedrock-runtime-capabilities.ts at 100%. tsc, eslint (0 errors), prettier and the changeset check are clean. Removing the catalog entry makes 7 of the new tests fail, and restoring it brings them back to passing.
Summary by cubic
Adds openai.gpt-6.1-sol to the Bedrock Runtime capability catalog so model discovery lists its us. and global. inference profiles. The profiles follow the openai.gpt-6-astra behavior: max_completion_tokens only, no reasoning_effort: none, and Chat Completions with function tools routed to Runtime Responses, which addresses the #3113 request for the cheaper global. profile.
Notes
- Price data (2.00 in / 10.00 out for the global profile) comes from the
amazon-bedrockprovider and was not visible in the local dashboard. openai.gpt-6.1-terra,openai.gpt-5.6-sol, andopenai.gpt-5.6-terrahave the same discovery gap and are left for separate PRs.