GPT-6.1 Sol for teams already running agents

In 60 seconds: GPT-6.1 Sol is an upgrade to GPT-6 Sol for coding, computer use, and professional work. OpenAI positions it near Astra at one-fifth of Astra’s Standard input and output token prices. For an agent already in production, the clearest difference may be repeated context: cached input costs $0.10 per million tokens, half the GPT-6 Sol rate. Do not switch models blind. Rerun the workflow’s evals, check cost per correct outcome, and keep the same barriers around sensitive actions.

What OpenAI released

OpenAI released GPT-6.1 Sol on September 29, 2026. Its official name is GPT-6.1 Sol and its API ID is gpt-6.1-sol. Although some searches call it “ChatGPT 6.1 Sol,” the release spans three separate surfaces: ChatGPT Work, Codex, and the API.

The provider’s claim is specific: near-GPT-6 Astra performance in agentic coding, computer use, and professional work at one-fifth of Astra’s Standard input and output token prices. That does not guarantee one-fifth of the cost for your process. Tool calls, retries, context volume, and human review still belong in the real cost of an AI agent.

Published specification	GPT-6.1 Sol
API ID	gpt-6.1-sol
Standard input	$2 per million tokens
Cached input	$0.10 per million tokens
Cache writes	$2.50 per million tokens
Standard output	$10 per million tokens
Context window	1,050,000 tokens
Maximum output	128,000 tokens
Knowledge cutoff	April 30, 2026

The prices in the table apply to Standard processing and prompts up to 272,000 tokens. The model page lists different rates for longer prompts, Fast mode, regional processing, Batch, and Flex. Check that page when budgeting because the headline rates do not describe every workload.

Cached context changes one part of the bill

GPT-6.1 Sol keeps GPT-6 Sol’s Standard input and output rates while cutting cached input from $0.20 to $0.10 per million tokens. The lower rate applies only to tokens that qualify for cached-input pricing.

For a production agent, this favors long prefixes reused across requests: stable instructions, tool definitions, policies, and shared reference material. It does not make the whole context free. New portions of a request use the regular input rate, cache writes have their own price, and sending more than 272,000 input tokens changes the rates for the full request.

Split four values in your traces before estimating savings:

1. uncached input tokens;
2. cached input tokens;
3. cache-write tokens;
4. output tokens and retries.

Then compare cost per accepted case with cost per call. A model that needs fewer corrections may justify a longer run. One that repeats steps can erase the caching advantage.

Coding and computer use: replay the hard cases

The API model supports computer use, web search, file search, hosted shell, apply patch, and MCP through the Responses API. OpenAI says Chat Completions works without tool calling and recommends Responses API for tools.

For a team with agents in production, the useful question is whether GPT-6.1 Sol completes cases that currently end in a retry, escalation, or manual correction. Build a set from real, anonymized runs:

• code changes that cross several files and require tests;
• browser tasks with ambiguous states, permissions, or confirmations;
• long documents with tables, appendices, and exceptions;
• known tool, session, and credential failures.

Run that set with the same harness, tools, and permissions. Record quality, latency, tokens, tool calls, corrections, and blocked actions. The launch benchmarks can tell you what to test. Your traces decide whether to switch.

Where it is available, and where it is not

As of September 30, 2026, the ChatGPT Work and Codex model documentation lists GPT-6.1 Sol for:

• Plus, Pro, Business, Enterprise, and Edu in Codex desktop and CLI;
• ChatGPT Work on web and mobile;
• the OpenAI API under gpt-6.1-sol.

OpenAI says the model is not available in Chat. Free and Go are also outside the initial rollout. Enterprise and Edu administrators must enable it because it is off by default. Actual access depends on the rollout, client, sign-in method, and workspace settings.

Standard and Fast modes are available at launch. The models page says Ultrafast support is coming later. If an official page does not publish a modality, region, or limit, treat it as unpublished and verify it before committing capacity.

What the system card says about controls

OpenAI published the GPT-6.1 Sol system card addendum on September 29, 2026. Under its Preparedness Framework, OpenAI treats the model as Critical for cybersecurity capability and High for biological and chemical capability. GPT-6.1 Sol therefore uses the same safeguards stack as GPT-6 Astra.

The document reports better results than GPT-6 Sol in five of eight Production Benchmark categories and fewer unintended outcomes in adversarial computer-use tests. It also warns that evaluations ran in a research environment or through the API, where system prompts, tools, and effort can differ from production.

That evidence can support a model evaluation. It does not replace system controls. Keep least-privilege credentials, an explicit tool allowlist, confirmation before sending, paying, deleting, or deploying, and logs that reconstruct every action. Test the workflow against untrusted pages, emails, and documents too. A model change does not change who is accountable for an authorized action.

A short rollout for an existing agent

Choose a workflow with an existing baseline and run GPT-6.1 Sol in shadow mode without write permissions. Use the same cases and rubric as production. Review normal cases, policy boundaries, and tool failures separately.

If it passes the quality and safety gate, send a small share of traffic with a defined rollback. Keep the change when it improves cost per accepted outcome without increasing unauthorized actions, corrections, or escalations. A detailed Sol versus Opus comparison needs a separate source set and evaluation design, so it sits outside this article.

Sources and currency

We verified this article on September 30, 2026 against official sources:

• GPT-6.1 Sol announcement: positioning, relative price, availability, and release date.
• GPT-6.1 Sol API model page: pricing, context, knowledge cutoff, supported tools, and published limits.
• ChatGPT Work and Codex models: plans, surfaces, rollout, and availability outside Chat.
• System card addendum: safety evaluations, Preparedness Framework classification, and safeguards.

Prices, access, and limits can change. Check the current pages and your account settings before moving production traffic.