GPT-6.1 Sol rollout: separate API, Codex, and Copilot before changing defaults

Author: Ahab · Date: September 30, 2026

Quick answer

GPT-6.1 Sol is available as gpt-6.1-sol in the OpenAI API, Codex, and ChatGPT Work. GitHub also marks it GA with a gradual Copilot rollout. Treat each surface as a separate contract: endpoint, plan, policy, context, and billing rules differ. Roll out one task class at a time, record the model actually used, and keep the previous route until the candidate passes acceptance and cost gates.

Who this is for

This guide is for teams moving API agents, Codex workflows, or Copilot users from an approved model to GPT-6.1 Sol.

If you are still choosing between GPT-6 Sol and Luna by workload, use the accepted-result cost routing guide. If your model setting is distributed through GitHub enterprise policy, first prove that the policy reaches clients with the managed settings validation checklist.

What changed on September 29, 2026

OpenAI added GPT-6.1 Sol to Codex and ChatGPT Work, describing near-Astra capability at a lower cost. Codex CLI 0.159.1 added it to the bundled catalog. Availability still depends on plan, client, and workspace settings; the announcement does not call it a general ChatGPT model.

The API model page lists a 1,050,000-token context window, 128,000 maximum output, text and image input, and tools through Responses. It does not support none or minimal reasoning effort; Chat Completions supports text but not tool calling.

GitHub announced GPT-6.1 Sol as GA and rolling out to Copilot Pro+, Max, Business, and Enterprise. Business and Enterprise administrators control it through model policy. Default enablement can activate a new model unless the global default is off or that model is disabled.

OpenAI's “near-Astra” positioning and GitHub's token observations are vendor claims, not proof for your workload.

Keep the surface contracts separate

Surface Confirm before testing Cost boundary
OpenAI API Exact model ID, Responses endpoint for tools, reasoning effort, residency, context size Provider token and service-tier pricing
Codex / ChatGPT Work Client version, plan, workspace model setting, selected model Codex or plan usage rules, not an API invoice
GitHub Copilot Plan, client, rollout visibility, organization model policy GitHub AI credits and Copilot pricing

“Visible in one picker” is not evidence that another surface is enabled, billed the same way, or using the same context mode.

A six-step rollout workflow

1. Inventory every selector

List API configuration, Codex defaults, saved agents, scheduled work, Copilot policy, IDE settings, and Auto selection. Record each current model and owner before replacing a shared alias.

2. Define one frozen task corpus

Choose one narrow task. Include happy paths, long context, images, failed tools, ambiguous instructions, and permission-sensitive actions. Freeze prompts, tools, timeouts, retries, and rubric; compare accepted results on the same corpus.

3. Run endpoint and effort canaries

For API tools, use Responses and verify every tool-call/result pair. Test low, medium, and your highest allowed effort; reject unsupported none or minimal.

Record this receipt:

surface: openai-api
requested_model: gpt-6.1-sol
endpoint: responses
reasoning_effort: medium
input_tokens: 184200
cached_input_tokens: 126000
cache_write_tokens: 58200
output_tokens: 7400
tool_calls_expected: 6
tool_calls_completed: 6
accepted: true

For Codex and Copilot, capture client, plan, policy, picker selection, task ID, and available model evidence.

4. Apply the long-context price gate

The standard API list price per million tokens is $2 input, $0.10 cached input, $2.50 cache write, and $10 output. Above 272K input tokens, the full request uses doubled input/cache rates and 1.5× output. Fast mode is 2× Standard; regional processing adds 10%, and Fast mode is unavailable with EU residency.

Calculate cost from the actual surface receipt. Do not copy an API estimate into a Copilot budget, or assume a 1M context option is free because the model supports it.

5. Compare accepted-result cost

Use:

accepted-result cost =
  (model or credit spend + retries + review + failure recovery)
  / accepted outcomes

Gate schema validity, task completion, unauthorized-action rate, review time, p95 latency, retries, and accepted-result cost. A run with fewer tokens can still be worse if it needs more intervention.

6. Stage policy and keep rollback explicit

Start with named testers, then one repository or task class, then a small cohort. Verify disablement, defaults, and client availability. Keep the previous configuration and policy export until the canary covers real work.

Decision tree

Need GPT-6.1 Sol
  -> which surface owns execution and billing?
     API -> Responses for tools; validate effort, context tier, residency
     Codex/Work -> validate client, plan, workspace setting, selected model
     Copilot -> validate plan, client rollout, model policy, AI credits
  -> frozen corpus passes quality and authority gates?
     no: keep current model
     yes: accepted-result cost improves or capability justifies it?
       no: keep current route
       yes: staged canary -> receipt -> promote or roll back

Common mistakes

Make your Mac notch useful with SuperNotch—22 native tools for music, clipboard, focus, screenshots, system controls, and more.

FAQ

Is GPT-6.1 Sol available in regular ChatGPT?

The September 29 OpenAI announcement names Codex and ChatGPT Work and says availability depends on plan, client, and workspace settings. Do not generalize that statement to every ChatGPT surface.

Is GPT-6.1 Sol fully rolled out in GitHub Copilot?

GitHub labels the model GA but also says rollout is gradual. Check the exact plan, client, extension version, and organization model policy before calling an account eligible.

Should GPT-6.1 Sol replace GPT-6 Sol everywhere?

No. Keep routing task-specific. Migrate only where a frozen corpus shows an acceptable quality, authority, latency, and cost tradeoff.

What is the minimum production evidence?

Keep surface, requested and observed model, endpoint, policy, effort, context tier, usage, acceptance, cohort, and rollback trigger in one receipt.

Sources