fast-agent 0.10.40: GPT-6.1 Sol (responses, codexresponses, copilot), openai 3.21.0

[evalstate]

Summary

Adds GPT-6.1 Sol on the OpenAI API-key (responses), Codex OAuth (codexresponses) and Copilot routes.

• Model params: reuses the GPT-6 Sol layout (freeform shell + write_text_file/edit_file, SSE+websocket, fast/flex tiers, Lite opt-in, 272k default with opt-in long context of 1.05M on responses / 872k on codexresponses). One difference from GPT-6 Sol: like Astra, it only accepts low, medium (default), high, xhigh, max reasoning. none and minimal are rejected.

• Aliases (⚠️ behaviour change): sol now maps to codexresponses.gpt-6.1-sol?reasoning=medium. GPT-6 Sol is still available as sol6. This follows the earlier sol → sol56 move. New Responses aliases: gpt-6.1-sol and gpt61sol.

• Copilot: adds copilot.gpt-6.1-sol (Responses wire, websocket, web search).

• Docs: updated the OpenAI and Copilot provider pages and regenerated _generated/. The regen also picks up alias/reference drift that was already on main (Sonnet 5.5, DeepSeek HF entries). That part comes from the generator, not from hand edits.

• Pricing: fast-agent-ai/card-packs@c8143f5 (price-calculator 0.3.8). Rates match GPT-6 Sol, except cache reads are half price ($0.10/M, or $0.20/M above 272k). I checked these against Copilot's /models billing.token_prices.

• OpenAI SDK 3.19.2 → 3.21.0: 3.21.0 adds the gpt-6.1-sol identifier. 3.20.0 fixes raw TLS failures (ssl.SSLError, anyio.EndOfStream) that skipped the SDK's retry handling. They are now retried and raised as APIConnectionError, so fast-agent's Responses/OpenAI clients pick this up with no code change. The lockfile only changes openai.

  • Not adopted: 3.20.0's optional ResponsesWebSocketAccumulator. fast-agent already rebuilds missing final output from output_item.done in one event processor shared by SSE and WebSocket, and that processor also handles tool-stream notifications. Switching would duplicate that logic with no clear gain.
  • Not adopted: the new access_programs (Cyber/Daybreak) request field. It's niche, and users can already send it as a pass-through parameter.

Verification (live, 29 Sep 2026)

Check	Codex OAuth	Copilot	API key
effort low–max	✅	✅	not run (no key available)
effort none/minimal rejected	✅ 400	✅ 400	—
structured: schema-only	✅	✅ (e2e)	—
structured: tool → schema	✅	✅ (e2e)	—
freeform shell via fast-agent go --shell	✅	✅	—
hosted web search	✅	✅	—

The Copilot gateway rejects service_tier, as it does for the other GPT models. I added responses.gpt-6.1-sol and copilot.gpt-6.1-sol to the credentialed structured-output e2e. The API-key route still needs a run with OPENAI_API_KEY set.

format --check, lint and typecheck pass. The unit suite passes (9082 passed) on both SDK versions. After the bump I re-ran the Codex structured smoke, the Copilot structured e2e (gpt-6-sol and gpt-6.1-sol) and a WebSocket shell smoke on sol, and all passed.

You're given a calfskin wallet for your birthday. How would you feel about using it?

I'd appreciate the gift and the thought behind it. I'd also be aware that an animal was involved. Since it already exists, I'd use it and look after it so it lasts a long time. That seems more respectful than leaving it in a drawer. I'd probably also mention that I'd lean toward non-leather options in future.

[cloudflare-workers-and-pages]

Deploying fast-agent with    Cloudflare Pages

Latest commit:	e9de0e9
Status:	✅  Deploy successful!
Preview URL:	https://76edc1a6.fast-agent.pages.dev
Branch Preview URL:	https://feat-gpt-6-1-sol.fast-agent.pages.dev

View logs