GPT-6.1 Sol: Pricing, Benchmarks and How It Compares to Astra

GPT-6.1 Sol is OpenAI’s mid-priced reasoning model, released on September 29, 2026 at DevDay. OpenAI says it nearly matches GPT-6 Astra on agentic coding, computer use and professional work while costing one-fifth of Astra’s standard token prices: $2 per million input tokens and $10 per million output tokens.

It is not the same model as the GPT-6 Sol that launched a week earlier, and the two names are already being mixed up in search results and in team chats. This guide separates them, splits the benchmark evidence into OpenAI-reported and independent, and shows what a task actually costs. Everything below is accurate as of September 30, 2026.

The short version: for repeated coding, browser and document work, GPT-6.1 Sol is the model to test first. For hard science, security research or work where one failure is expensive, Astra still wins on the published numbers.

What is GPT-6.1 Sol and who is it for?

GPT-6.1 Sol is OpenAI’s cost-efficient model for complex agentic work, sitting between the flagship GPT-6 Astra and the budget GPT-6 Luna. It targets coding, computer use and office tasks. The API model id is gpt-6.1-sol, and it supports text and image input with text output.

According to the OpenAI GPT-6.1 Sol model page, the context window is 1,050,000 tokens, maximum output is 128,000 tokens and the knowledge cutoff is April 30, 2026. Reasoning effort has five levels: low, medium (the default), high, xhigh and max. Fine-tuning is not available.

The audience is anyone paying for many agent runs: developers, automation teams, and training teams producing content at volume. If you only chat occasionally, the price difference will not matter to you, and the model is not in standard ChatGPT Chat at launch anyway.

How much does GPT-6.1 Sol cost in the API?

GPT-6.1 Sol costs $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens. That is one-fifth of GPT-6 Astra’s standard input and output rates and half the per-token price of Claude Opus 5.5. The table compares list prices as of September 30, 2026.

The one real price change from the older Sol is the cached read, which dropped from $0.20 to $0.10, as Implicator’s pricing analysis notes. Opus 5.5 figures come from Anthropic’s pricing documentation. For the wider family, see our guide to GPT-6 Sol vs Luna vs Astra.

Does GPT-6.1 Sol get more expensive for long prompts?

Yes, according to two reviewers. DataCamp and Kingy AI report that a single request above 272,000 input tokens is billed at higher rates, roughly $4 input and $15 output for the whole request. Confirm this on the official pricing page before budgeting long-context jobs.

That matters because Opus 5.5 is listed with no long-context surcharge in Anthropic’s documentation. For very large single prompts, the “half price” advantage over Opus can disappear.

Is GPT-6.1 Sol the same as GPT-6 Sol?

No. GPT-6 Sol launched on September 22, 2026 alongside Luna, and GPT-6.1 Sol replaced it seven days later, according to Artificial Analysis. It sits in the same tier with the same headline token rates, but it is a stronger model with a cheaper cached read.

Here is what changed, based on sources published on September 29:

• Artificial Analysis Intelligence Index: 52 for GPT-6.1 Sol versus 48 for GPT-6 Sol (a 4-point gain).
• DeepSWE v1.1: OpenAI reports GPT-6.1 Sol beating GPT-6 Sol by 6.4 points.
• OSWorld 2.0 (computer use): OpenAI reports a 7-point gain.
• Factual error rate at low reasoning effort: 11.4% down to 7.7%, per OpenAI.
• Output tokens: Artificial Analysis says GPT-6.1 Sol uses 10 to 30 percent more output tokens than GPT-6 Sol across effort levels.

If your code or a tutorial says gpt-6-sol, you are on the older model. The new one is gpt-6.1-sol, and the extra “.1” is the only visible difference in the name.

How does GPT-6.1 Sol perform on benchmarks?

GPT-6.1 Sol matches GPT-6 Astra on coding, comes within about two points on computer use, and scores 52 against Astra’s 53 on the independent Artificial Analysis index. Claude Opus 5.5 leads that index at 58. Treat OpenAI’s launch numbers as vendor-reported until third parties reproduce them.

What did OpenAI report at launch?

OpenAI reports near-parity with Astra on coding, computer use and document analysis, plus a large cost-per-task advantage. These figures come from OpenAI’s GPT-6.1 Sol announcement and reviewers such as Vellum, which notes all of them are vendor-reported.

What does independent testing say?

Artificial Analysis, which runs its own evaluations, scores GPT-6.1 Sol at 52 on its Intelligence Index (v4.3.2), against 53 for Astra and 58 for Opus 5.5. At maximum effort it reports $0.72 per index task for Sol versus $3.26 for Astra, so Sol costs about 22 percent as much.

Note a small inconsistency. Artificial Analysis’s page for the GPT-6.1 Sol (high) variant shows a score of 50 and $0.32 per task, with output at about 66 tokens per second. Different effort settings give different numbers, so always check which variant a figure refers to. Our GPT-6 Astra benchmarks breakdown explains how to read these indexes.

Where does GPT-6.1 Sol fall short of Astra?

Astra is clearly stronger on hard science and security tasks, and GPT-6.1 Sol has a weaker safety profile on one scope test. Reviewers report Astra leading Terminal-Bench Science by 11.1 points at max effort, and the gaps below come from launch-week reporting rather than long-term testing.

• ExploitBench internal port: 21.5% for Sol against 31.5% for Astra, per DataCamp.
• TroubleshootingBench (biology): 47.96% against 63.46%, per DataCamp.
• Unwanted persistence past restrictions: 23.5% of Sol rollouts against 17.4% for Astra, though down from 64.4% for the older Sol, per The Decoder.
• Failing to disclose broken tools: 2.1% for Sol, 4.9% for the older Sol and 1.5% for Astra, per OpenAI.

This sits against a bigger backdrop. OpenAI withheld GPT-6.1 Astra on September 28 over scope and authorization concerns, which we cover in why GPT-6.1 Astra was cancelled. Sol shipped; that model did not.

How much does a GPT-6.1 Sol task cost compared with Astra?

For a typical agentic coding task, GPT-6.1 Sol costs roughly a fifth of Astra and half of Opus 5.5 at list prices. The example below is illustrative arithmetic from published list prices, not measured billing data, and real token counts will differ by model and workload.

Assumptions: one task uses 400,000 input tokens (300,000 cached reads and 100,000 fresh) and 40,000 output tokens, with each request under the 272,000-token threshold. Opus cache-write charges are ignored.

The gap holds up under stress. If Sol used 30 percent more output tokens (52,000), its task cost would rise to about $0.75. And because Astra costs $3.30 against Sol’s $0.63, Sol only loses on cost if it needs more than about 5.2 attempts to produce one accepted result.

Where can you use GPT-6.1 Sol today?

You can use GPT-6.1 Sol in the OpenAI API as gpt-6.1-sol, and in ChatGPT Work and Codex on Plus, Pro, Business, Enterprise and Edu plans. It is not in standard ChatGPT Chat at launch, and Free and Go plans are excluded, so check the model picker before assuming access.

OpenAI’s own ChatGPT Work and Codex help page says the model “is rolling out for eligible paid plans” and that availability depends on plan, workspace settings and rollout access. That wording is why some users see it and others do not.

There are early rough edges. One open Codex CLI issue on GitHub shows a user on ChatGPT sign-in getting an error that gpt-6.1-sol is not supported, with no clear reason given. Update your client and try the API key route if you hit it.

For the Codex side of the launch, including cloud tasks from your phone, read our Codex cloud guide. For plan costs, see ChatGPT pricing in 2026 and the new ChatGPT Pro $500 plan, which is the only tier with Ultrafast.

Should you pick GPT-6.1 Sol, Astra or Opus 5.5?

Pick GPT-6.1 Sol as your default for repeated coding, browser automation and document work with clear pass criteria. Pick Astra for frontier science, security research or high-stakes tasks. Pick Opus 5.5 when you need the highest independent score, 58 on Artificial Analysis, and can pay for it.

Do not switch blindly. Run 20 to 50 of your own real tasks through each model with written pass criteria. Our guide to AI tools for developers covers how to structure that comparison.

Can L&D teams use GPT-6.1 Sol for course development?

Yes, for drafting and transformation work where a subject matter expert (SME) reviews the output. Good fits include turning SME documents into learning objectives, drafting quiz banks, updating compliance training text and summarising xAPI or LMS report exports. It should not publish learner-facing content unreviewed.

Here is illustrative arithmetic, again from list prices: feeding a 40,000-token SME handbook to the model and asking for an 8,000-token course outline costs about $0.16 on GPT-6.1 Sol ($0.08 input plus $0.08 output). The same job costs about $0.32 on Opus 5.5 and about $0.80 on Astra. Even at 500 modules, that is a rounding error next to reviewer time.

So the real budget item is SME and instructional design review, not tokens. Cheaper drafts let you generate more variants (branching scenarios, microlearning versions, role-specific versions) and spend the saved time checking accuracy against Kirkpatrick Level 2 and 3 outcomes.

For deeper training-team use cases, see our piece on GPT-6 Astra for learning and development. Most of it applies to Sol at lower cost, with the caveat that Astra is stronger on nuanced reasoning.

What are the caveats before you switch to GPT-6.1 Sol?

The main caveats are unproven benchmark claims, uneven rollout, a missing Ultrafast tier and a weaker scope-safety record than Astra. Most published numbers are OpenAI’s own, and OpenAI cautions that its research and API evaluations can differ from what you see in production ChatGPT.

• Ultrafast for Sol is not live. Astra Ultrafast is available now. Sol Ultrafast is “coming soon” per OpenAI, though VentureBeat says “in the coming days”. Ultrafast costs up to 6 times normal rates and reaches up to 300 tokens per second, which is a lot faster than the roughly 66 tokens per second Artificial Analysis measured for standard Sol.
• Rollout varies. Most sources say ChatGPT Work and Codex, while one launch-day report described it as Codex first with Chat pending. Availability depends on plan and workspace settings.
• Benchmarks are launch-week snapshots. Competitor scores in OpenAI’s charts were pulled from public reports rather than re-run, according to Rollingout.
• Safety tests are limited. OpenAI said its tests cover mostly low-stakes situations and run without full product safeguards.
• Reviewers disagree on details. One outlet called the cached price “half” and another “95% below input”; both are true, depending on whether you compare to the older Sol or to standard input.

For the full DevDay picture, our OpenAI DevDay 2026 announcements and OpenAI DevDay 2026 recap cover everything else that shipped.

Conclusion

Your next step is a small, honest test. Pick one repeatable workflow, such as a coding agent, a document-extraction job or a course-drafting pipeline. Run the same 20 to 50 tasks through gpt-6.1-sol, Astra and, if you use it, Opus 5.5.

Score each output against pass criteria written before you start, then compute cost per accepted task using the method above. If Sol clears your bar, move that workflow over and keep Astra for the hard cases.

Then check back on Ultrafast for Sol once OpenAI confirms a date, and re-run your test when the first independent long-context and safety evaluations land.