GPT-6.1 Sol: near-Astra results at a fifth of the price

OpenAI released GPT-6.1 Sol on 29 September 2026 as an upgrade to GPT-6 Sol. Source: OpenAI, checked 2 October 2026.

GPT-6.1 Sol keeps GPT-6 Sol’s API price of $2 per million input tokens and $10 output, halves cached input to $0.10, and in independent testing matches GPT-6 Sol’s best score at a fifth of the cost per task. Here is what changed, what it costs, how it compares with Claude and when GPT-6 Astra is still worth paying for. Prices, availability and independent scores were checked on 2 October 2026.

The answer in brief

GPT-6.1 Sol is OpenAI’s middle GPT-6 model and a direct upgrade to GPT-6 Sol, available in the API as gpt-6.1-sol since 29 September 2026. If you already use GPT-6 Sol, it is worth switching once your own tasks pass on it: it costs the same per token, scores higher on independent tests and finishes them for less.

• Price: $2 per million input tokens, $0.10 cached input and $10 output. Prompts over 272K input tokens are billed at twice the input rate and 1.5 times the output rate for the whole request.
• Independent results: on the Artificial Analysis Intelligence Index v4.3.2 it scores 52 at max effort against 53 for GPT-6 Astra, at $0.72 per task against $3.26.
• Better value at lower effort: at medium effort it scores 48, the same as GPT-6 Sol at max, for $0.21 per task instead of $1.04.
• Against Claude: Claude Opus 5.5 (58) and Sonnet 5.5 (56) still score higher at max effort on the same index. At high effort, GPT-6.1 Sol scores 50 for $0.32 per task, against 47 for Sonnet 5.5 at high for $1.08.
• Where it runs: ChatGPT Work and Codex on Plus, Pro, Business, Enterprise and Edu plans (Enterprise and Edu administrators must switch it on), plus the API. It is not yet in ordinary ChatGPT chat.
• Usage in ChatGPT: OpenAI estimates 15 to 160 local messages per five hours with GPT-6.1 Sol on Plus and Standard Business, against 5 to 45 with Astra.
• When to pay for Astra: the hardest scientific and computer-use work, where OpenAI’s own figures still put Astra ahead.
• The catch: at max effort it is slow to start. Artificial Analysis measured almost five minutes to the first token.

What is GPT-6.1 Sol?

GPT-6.1 Sol is the middle tier of OpenAI’s GPT-6 family, between GPT-6 Astra at the top and GPT-6 Luna for cheap, high-volume work. OpenAI launched it at DevDay on 29 September 2026 as an upgrade to GPT-6 Sol, which had arrived only on 22 September. OpenAI’s own pitch is near-Astra intelligence for coding, computer use and professional work at one-fifth of Astra’s standard token prices.

It is a reasoning model. In the API, reasoning.effort accepts low, medium (the default), high, xhigh and max; the none and minimal settings are not supported, so there is no non-reasoning mode as there was for GPT-6 Sol. It reads text and images and writes text, with a 1,050,000-token context window, up to 128,000 output tokens and a knowledge cutoff of 30 April 2026, according to OpenAI’s model page.

OpenAI’s model-selection guide describes Sol as the choice for complex tasks where time and cost still matter, and suggests medium effort for complex technical work and xhigh for polished deliverables. Its own example is a board presentation built from financial results.

How much does GPT-6.1 Sol cost?

GPT-6.1 Sol costs $2 per million input tokens and $10 per million output tokens, the same as GPT-6 Sol. The change is cached input, which drops from $0.20 to $0.10 per million, 5% of the uncached rate. For agents that resend the same long context on every turn, that is where the saving shows up.

Four pricing rules on OpenAI’s model page change real bills:

• Long prompts cost more. A request with more than 272K input tokens is charged at twice the input and cache rates and 1.5 times the output rate, for the full request.
• Cache writes are not free. Writing to the cache costs $2.50 per million tokens, 1.25 times the input rate.
• Batch and Flex are 50% cheaper than standard. Fast mode costs twice standard.
• Regional processing adds 10% where it is available.

A worked example: an agent turn that reads 200,000 cached tokens, 20,000 new input tokens and writes 5,000 output tokens costs about $0.11 on GPT-6.1 Sol ($0.02 + $0.04 + $0.05). The same turn on GPT-6 Sol costs about $0.13, because the cached part costs twice as much. This is AIWIZ’s arithmetic from the published rates, not an OpenAI figure. It counts tokens only and assumes all 200,000 cached tokens are cache hits; cache writes and paid tool calls would add to the bill.

What does OpenAI claim GPT-6.1 Sol can do?

OpenAI says GPT-6.1 Sol matches or comes close to GPT-6 Astra on coding, documents and computer use, at a fraction of the cost per task. These are OpenAI’s own runs, and competitor scores come from public reports, so treat them as the vendor’s case rather than a neutral test.

OpenAI is candid about the ceiling. On Terminal-Bench Science, Astra still has the top score of the models tested, 68.1%, and OpenAI says Astra should be used for the hardest scientific research.

OpenAI also reports fewer factual errors. On deliberately difficult prompts taken from conversations where users had flagged a mistake, the share of answers with a factual error fell from 11.4% to 7.7% at low effort, a cut of about a third. OpenAI notes these prompts are not typical of everyday use.

Do independent benchmarks back it up?

Largely, yes. Artificial Analysis ran GPT-6.1 Sol through its Intelligence Index v4.3.2, ten evaluations covering agentic work, coding, reasoning and long context. At max effort it scores 52, one point behind GPT-6 Astra at max, at $0.72 per task against $3.26.

The more useful finding sits lower down the effort scale. GPT-6.1 Sol at medium scores 48, the same as GPT-6 Sol at max, for $0.21 per task instead of $1.04. Most teams that ran GPT-6 Sol flat out can drop to medium on 6.1 and pay about a fifth as much for the same index score. An equal index score is an equal aggregate across ten tests, not equal results on every task, so check your own workload before dropping the effort setting.

The gains are not free on every axis. Artificial Analysis measured GPT-6.1 Sol at max producing about 64 tokens a second and taking about 291 seconds to its first token, because of the reasoning it does before answering. GPT-6 Sol at max was faster to stream, at 89 tokens a second. If people are waiting on the answer, medium is the practical setting: about 6 seconds to the first token on the same tests.

All Artificial Analysis figures in this post are from Intelligence Index v4.3.2, checked 2 October 2026. They are not comparable with scores from earlier index versions.

How does GPT-6.1 Sol compare with Claude?

On independent tests Claude still leads at the top, and GPT-6.1 Sol wins on cost at everyday settings. On list price it matches Claude Sonnet 5.5 and costs half as much as Claude Opus 5.5.

Per-token prices only go so far. Anthropic notes that Claude 4.7 and later models use a tokenizer that produces roughly 30% more tokens for the same text, and reasoning models differ widely in how many tokens they spend on a task. Cost per task on your own workload is the comparison that counts.

Artificial Analysis has now run all three on the same index version, which is a fairer comparison than either vendor’s charts.

Read it in two halves. At the top, Claude is ahead: Opus 5.5 at max scores six points more than GPT-6.1 Sol at max, and Sonnet 5.5 four more. In the middle, where most teams actually run models, GPT-6.1 Sol is the better buy. At high effort it outscores Sonnet 5.5 at high for under a third of the cost per task, and comes within a point of Opus 5.5 at medium for about a quarter. Our Claude Sonnet 5.5 and Claude Opus 5.5 guides cover the Claude side.

OpenAI’s own figures point the same way on cost: about a third of the cost of Opus 5.5 on AutomationBench with a slightly higher score, and $5.47 against $23.21 per task on Terminal-Bench Science. An index score is still an average across ten tests, so run both on your own briefs before moving a production workload.

Where can you use GPT-6.1 Sol?

GPT-6.1 Sol is available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, and in the OpenAI API as gpt-6.1-sol. At launch it was not available in ordinary ChatGPT chat, and the API free tier does not support it. On Enterprise and Edu plans it is off by default until an administrator enables it, according to OpenAI’s model guidance.

• Codex: pick it in the model picker in the desktop app, or run codex --model gpt-6.1-sol in the CLI. In ChatGPT Work it is available on the web and mobile.
• Which ChatGPT plans: Plus includes GPT-6.1 Sol and GPT-6 Luna in Work and Codex; the Free plan gets GPT-6 Luna, not Sol.
• API: Responses, Chat Completions and Batch endpoints, with US and EU data residency.
• Microsoft Foundry: Microsoft lists gpt-6.1-sol (version 2026-09-29) among the models it sells directly, with multi-agent orchestration in preview.
• Ultrafast: OpenAI says Ultrafast support for GPT-6.1 Sol is coming later, with up to 8 times faster token generation in Codex. GPT-6 Astra Ultrafast is already live on Pro 500 and Enterprise plans, per the DevDay recap.

What should developers check before switching?

Changing the model ID is not enough on its own. Run your own tasks through GPT-6.1 Sol first, as OpenAI advises, then check these five details from OpenAI’s migration guide and model page.

1. Tool calling needs the Responses API. Chat Completions works, but without tool calling.
2. No non-reasoning mode. Low is the lightest setting. If your GPT-6 Sol code sent none or minimal, change it.
3. Drop the sampling parameters. The migration guide says to remove temperature, top_p and top_logprobs when reasoning is on, and on GPT-6.1 Sol it always is.
4. Fast mode and EU residency do not mix. Fast mode is unavailable with EU data residency.
5. Rate limits start modestly. Tier 1 accounts get 500 requests and 500,000 tokens a minute; Tier 5 gets 15,000 requests and 40 million tokens.

The tools available through the Responses API include web search, file search, code interpreter, hosted shell, apply patch, computer use, MCP and tool search. Multi-agent support, which lets the model hand work to subagents within one Responses API request, is in beta. Fine-tuning is not supported.

Is GPT-6.1 Sol safer than GPT-6 Sol?

On OpenAI’s evaluations, yes. OpenAI reports lower failure rates than GPT-6 Sol on admitting a broken search tool, respecting explicit restrictions and avoiding unauthorised outcomes in agentic tasks.

The clearest figure is the broken-search-tool test, which checks whether the model tells the user its search tool has failed rather than guessing. GPT-6.1 Sol failed to say so in 2.1% of cases, against 4.9% for GPT-6 Sol and 1.5% for Astra. The test is built to provoke failures, so the rates are not what you would see in normal use. Full results are in OpenAI’s GPT-6.1 Sol system card addendum.

What does GPT-6.1 Sol mean for a marketing team?

More Sol-level work for the same subscription, and a smaller model bill for the same quality of draft. It does not change who checks the work.

In ChatGPT Work and Codex, OpenAI’s pricing page estimates 15 to 160 local messages per five hours with GPT-6.1 Sol on Plus and Standard Business seats, against 15 to 150 with GPT-6 Sol and 5 to 45 with Astra. These are OpenAI’s ranges, not fixed limits, but a Plus user can run roughly three times as many Sol tasks as Astra tasks before reaching the cap. On credit-based Business and Enterprise plans, GPT-6.1 Sol costs 50 credits per million input tokens, 2.5 for cached input and 250 for output, half the cached rate of GPT-6 Sol. Fast mode draws on included usage at 2.5 times the Standard rate and on purchased credits at twice it.

In the API, a single marketing job costs the same as on GPT-6 Sol, so the illustrative costs in our GPT-6 Sol guide still apply: report commentary or a set of ad variants costs pennies in model charges. The saving shows up on repeated work. Ten ad-variant runs that reuse the same 30,000-token brand pack cost about $0.10 for the pack on GPT-6.1 Sol with one cache write and nine cache hits, against about $0.13 on GPT-6 Sol and $0.60 with no caching (AIWIZ’s arithmetic from published rates).

The bigger lever is the effort setting. Dropping from max to medium cut the Artificial Analysis cost per task from $0.72 to $0.21 for a four-point drop in score. For drafts a person will edit anyway, start at medium.

Judge the result on cost per approved deliverable: model charges plus review time, rework and failed runs. A good first test is one recurring job, such as a monthly Google Ads or Meta performance commentary from exported data, run through GPT-6.1 Sol and your current method against the same written checklist.

Should you switch to GPT-6.1 Sol?

For most GPT-6 Sol users, yes: the token price is unchanged, cached input is cheaper and the independent scores are higher at every effort setting Artificial Analysis tested. Run your own tasks first and make the API changes listed above, then start at medium, which matched the best result of GPT-6 Sol for a fifth of the cost per task, and move up only where your own tests show a gain.

If you use GPT-6 Astra, try GPT-6.1 Sol on the same tasks before renewing that spend. One index point separates them at max effort, and Astra costs about four and a half times as much per task on that test. Keep Astra for the hardest scientific, computer-use and open-ended work, which is where OpenAI’s own results still put it ahead.

If you use Claude Sonnet 5.5 at high effort, run a side-by-side test. On the independent index GPT-6.1 Sol scores higher at that setting for under a third of the cost per task, though Claude still leads when both run flat out.

If you mainly use ChatGPT for questions and writing, nothing changes yet. GPT-6.1 Sol is in ChatGPT Work and Codex, not in ordinary chat.

AIWIZ’s AI adoption services cover AI strategy, workflow automation and integration. If you want to test where GPT-6.1 Sol belongs in your marketing work, talk to AIWIZ. Bring one process, and we will help you design a trial that measures approved work before you scale it.