GPT-6.1 Sol: what changed from GPT-6 Sol and what a task costs

OpenAI released GPT-6.1 Sol on September 29, 2026, seven days after GPT-6 Sol, at the same $2 per million input tokens and $10 per million output tokens. One price moved: cached input, the rate for the part of a prompt the API has already stored, went from $0.20 to $0.10 per million tokens. The announcement calls the model “an upgrade to GPT-6 Sol” that nearly matches GPT-6 Astra, the model OpenAI describes as its most capable, “at one-fifth of Astra’s standard input and output token prices”.

If your code already calls gpt-6-sol, the model ID is not the only thing to change. One reasoning effort level is gone and tool calling moves to a different API, and that list comes first below.

After it come the costs. OpenAI’s page gives cost per task mostly as charts, so we took the points out of the page’s chart data.

They answer two questions. The first is what a task costs on GPT-6.1 Sol next to GPT-6 Sol when reasoning effort, the setting that runs from low to max, stays at its default. The second is where Claude stands in the same charts.

Switching from gpt-6-sol changes one price and two request rules

The table puts the two Sol models side by side, from the model pages of gpt-6-sol and gpt-6.1-sol in OpenAI’s API documentation. Prices are per million tokens, for prompts of up to 272K input tokens. Cache write is the rate for storing a prompt prefix, and cached input is the rate for reading it back.

	GPT-6 Sol	GPT-6.1 Sol
Model ID	gpt-6-sol	gpt-6.1-sol
Input	$2	$2
Cached input	$0.20	$0.10
Cache write	$2.50	$2.50
Output	$10	$10
Reasoning effort	none, low, medium (default), high, xhigh, max	low, medium (default), high, xhigh, max
Tools in Chat Completions	Function calling, only with effort none	Not supported; tool calling goes through the Responses API
Knowledge cutoff	April 20, 2026	April 30, 2026
Context window and maximum output	1,050,000 and 128,000 tokens	1,050,000 and 128,000 tokens

The cached input price halves in every tier of OpenAI’s pricing page where both models appear. For prompts of up to 272K input tokens it goes from $0.20 to $0.10 on Standard, from $0.10 to $0.05 on Batch and Flex, and from $0.40 to $0.20 on Fast.

A longer prompt pays twice the input and cache rates and 1.5 times the output rate on the whole request. For GPT-6.1 Sol that is $4 input, $0.20 cached input and $15 output.

The announcement’s one-fifth compares with GPT-6 Astra’s $10 input and $50 output. On cached input the gap is wider, $0.10 against Astra’s $1.

GPT-6.1 Sol does not accept effort none

The first request rule is reasoning effort. GPT-6 Sol accepts an effort of none and GPT-6.1 Sol does not. OpenAI’s GPT-6 guide says to use low in its place.

By our reading, that makes those calls cost more. OpenAI’s reasoning guide says OpenAI bills reasoning tokens as output tokens. A request that ran at none starts paying for thinking at $10 per million tokens once it moves to low.

The effort rule has a consequence for sampling. The GPT-6 guide says to remove temperature, top_p and top_logprobs whenever effort is not none, so by our reading every GPT-6.1 Sol request goes without them.

Tool calling requires the Responses API

The second rule is about tools. On GPT-6 Sol, the Chat Completions API supports function calling only when effort is none. On GPT-6.1 Sol, Chat Completions takes requests without tools, and the guide says tool calling requires the Responses API.

Code that calls tools through Chat Completions has to move.

Multi-agent is in beta, and the safeguards are GPT-6 Astra’s

Outside the two rules, the API changelog lists multi-agent support in beta, where the model delegates work to subagents inside one Responses request. The system card addendum says OpenAI treats GPT-6.1 Sol as Critical in cybersecurity and High in biological and chemical capability under its Preparedness Framework. The addendum adds that it uses the same safeguards as GPT-6 Astra.

At the default effort, a task costs about the same on three benchmarks and less on two

Six of the ten charts in the announcement plot a result against cost per task at each effort level, and five of those are named benchmarks. The table has the Medium point of each one, from the chart data on the page; Medium is the default effort of both models in the API. The first figure in a cell is the score and the second is the cost of one task, and every figure is as OpenAI reports it.

Benchmark	What it tests	GPT-6 Sol at Medium	GPT-6.1 Sol at Medium
DeepSWE v1.1	Software engineering tasks in real codebases	56.6%, $0.38	73.0%, $0.42
GDP.pdf	Questions about complex PDF documents	25.4%, $0.34	30.0%, $0.34
AutomationBench	Multi-step business workflows	26.9%, $0.21	31.7%, $0.19
OSWorld 2.0, offline set	Computer-use workflows	54.0%, $1.38	66.8%, $0.77
Terminal-Bench Science 0.1	Scientific workflows such as data analysis and simulation	14.5%, $4.41	47.6%, $2.34

Read across each row. The score is higher on all five. The cost moves by four cents or less on the first three: $0.42 against $0.38 on DeepSWE, $0.34 on both sides of GDP.pdf, $0.19 against $0.21 on AutomationBench.

On the last two, a task costs a little more than half of what it did: $0.77 against $1.38 on OSWorld, and $2.34 against $4.41 on Terminal-Bench Science.

OpenAI’s text does not say how much of the lower cost comes from the cheaper cache and how much from the model using fewer tokens. GitHub’s changelog points at the second: it says that in early testing the model “reliably completed tasks while using noticeably fewer tokens and steps than earlier models in the GPT-6 and GPT-5.6 families”.

On DeepSWE the GPT-6.1 Sol score peaks at High effort

More effort does not always buy a higher score. OpenAI’s text says GPT-6.1 Sol passes GPT-6 Sol’s best DeepSWE score “by 6.4 percentage points at a lower reasoning effort and cost”. In the chart data that is High effort, 75.2% for $0.65, against GPT-6 Sol’s 68.8% at Max for $2.74.

Above High, the GPT-6.1 Sol score goes down: Xhigh and Max both score 71.9%, for $0.79 and $1.57. GDP.pdf has the same shape, with its top score at High, and on the other three benchmarks the top score is at Max.

By our reading of OpenAI’s own points, Xhigh and Max add nothing on DeepSWE except a larger bill, and the default already scores 73.0% for $0.42. GPT-6 Astra’s best point on the same chart is 74.1% at Xhigh, for $4.43.

OpenAI’s charts have Opus 5.5 points and no Sonnet 5.5 point

Claude Opus 5.5, the model we compared with GPT-6 Sol in our News section, appears in three of the ten charts, labelled “Opus 5.5 w/ fallbacks”. A note at the end of the announcement says the competitor numbers “were taken from publicly available reports”.

On AutomationBench, Zapier’s benchmark of business workflows, OpenAI’s text cites one effort level: GPT-6.1 Sol “scores 2.2 percentage points above Opus 5.5 at medium reasoning effort, at roughly a third of the cost”. The chart has all five. GPT-6.1 Sol is ahead at Low, Medium and High, and behind at Xhigh, 35.5% against 35.8%, and at Max, 36.1% against 42.5%.

GPT-6.1 Sol is cheaper at every level, $0.16 to $0.30 a task where Opus 5.5 costs $0.51 to $1.44. The Max point of Opus 5.5, 42.5% for $1.44, is the highest on OpenAI’s chart, above GPT-6 Astra at Max, 41.4% for $1.73.

Three Opus 5.5 points on AutomationBench are rows of Zapier’s leaderboard

Those Opus 5.5 points trace to Zapier’s public leaderboard. Its page shows ten rows before a button that loads the rest, and those ten are all we could read. We read them on the live page on October 1, 2026 and in a capture from September 26, 2026.

Three of the five Opus 5.5 points are among them, as rows marked “default fallbacks”: 42.47% for $1.44 at Max, 35.77% for $0.89 at Xhigh and, in the capture, 33.03% for $0.71 at High. Zapier’s note explains the label: a step the model refuses goes to “the provider’s own default fallback model”.

OpenAI’s own models are on that leaderboard with the numbers from OpenAI’s chart. The capture lists GPT-6 Sol at 33.2% for $0.27 at Xhigh and 32.0% for $0.34 at Max, the same points the chart plots. The live page has four of the GPT-6 Astra points.

The leaderboard lists Sonnet 5.5, and the charts do not

The live page has models the charts do not. Its first two rows are Gemini 4 Argon.

It also lists Claude Sonnet 5.5, which launched on September 28, 2026 at $2 and $10 per million tokens. Zapier has it at 36.83% for $0.48 at Xhigh and 44.75% for $1.14 at Max.

None of the ten charts has a Sonnet 5.5 point. The ten rows of the September 26 capture have no Sonnet 5.5 or Gemini 4 Argon row either, and the Wayback Machine has no later capture. We cannot tell whether the leaderboard listed either when OpenAI published.

None of the ten live rows is GPT-6.1 Sol, although the 36.1% and 35.5% that OpenAI reports would rank inside them. OpenAI does not say which task set its GPT-6.1 Sol evaluation used. The README of Zapier’s repository says Zapier scores the leaderboard on a private task set that is harder than the public one.

So the pair closest in score that we can assemble is GPT-6.1 Sol at Max, 36.1% for $0.30 on OpenAI’s chart, and Sonnet 5.5 at Xhigh, 36.83% for $0.48 on Zapier’s leaderboard. Our read is that it puts the two models in the same range on this benchmark, with GPT-6.1 Sol cheaper. By the same reading, the comparison stays provisional until GPT-6.1 Sol has a row of its own on the leaderboard.

A Pac-Man test from Hacker News has Opus 5.5 as the judge

The Hacker News thread about the launch has tests that are in neither OpenAI’s charts nor Zapier’s leaderboard, and one of them publishes its run data. A comment there posts results from PacBench, which asks a model for a Pac-Man game from a single prompt. The results are 91 points for $0.51 on GPT-6.1 Sol, 99 for $2.00 on Opus 5.5 and 87 for $2.42 on GPT-6 Astra.

The test’s files say all three ran at High effort, the two GPT models in Codex and Opus 5.5 in Claude Code. They also say Opus 5.5 gave the points, by reviewing the finished games. One of the three contestants is the judge, so by our reading the cost column is the one to use.

The same files show what the cached input price acts on: in the GPT-6.1 Sol run, 237,922 of the 283,790 input tokens were cached input.

Against Sonnet 5.5, the list price differs on one line

Sonnet 5.5 is the Claude model at GPT-6.1 Sol’s price. For prompts of up to 272K input tokens, input, output and cache write match: $2, $10 and $2.50 on Anthropic’s pricing page as in OpenAI’s documentation. The cache read is $0.10 on GPT-6.1 Sol against $0.20 on Sonnet 5.5.

The cache write buys a different lifetime on each side. Anthropic’s page says its caching prices reflect a 5-minute lifetime, and OpenAI’s prompt caching guide gives at least 30 minutes after the latest write or reuse.

Anthropic’s launch page for Sonnet 5.5 compares it with the older Sol. Published the day before GPT-6.1 Sol, it says that at High effort Sonnet 5.5 “matches GPT-6 Sol’s best score for about a fifth of the cost per task” on FrontierCode. None of the pages we read has that comparison for GPT-6.1 Sol.

GPT-6.1 Sol runs in five places from launch day, and two AWS pages disagree on caching

GPT-6.1 Sol was available in five places on launch day, each with its own identifier or plan rule:

• In the OpenAI API the ID is gpt-6.1-sol, on the Responses, Chat Completions and Batch endpoints, with US and EU data residency. Fast mode is unavailable with EU residency.
• In ChatGPT it is in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu, by OpenAI’s rollout notes. It is off on Enterprise and Edu until an administrator enables it. It is not in Chat, and the Free and Go plans do not get it.
• In GitHub Copilot it is in the model picker for Pro+, Max, Business and Enterprise, in a gradual rollout, billed at OpenAI’s list price. GitHub’s pricing table shows the $0.10 cached input.
• On Amazon Bedrock it is generally available as openai.gpt-6.1-sol, with us.openai.gpt-6.1-sol and global.openai.gpt-6.1-sol for cross-Region inference. The model card prices the global. ID at OpenAI’s rates and adds 10% on the other two: $2.20 input, $0.11 cache read and $11.00 output.
• On Vercel’s AI Gateway the name is openai/gpt-6.1-sol, by Vercel’s changelog, and the model page lists three providers for it: OpenAI, Bedrock and Azure.

Before you design a Bedrock integration around explicit prompt caching, check which AWS page is right. The AWS announcement says the model “supports explicit prompt caching on Amazon Bedrock”. The model card lists explicit prompt caching under Not Supported and states that “Explicit prompt caching is not supported for this Bedrock model”.

We read both pages on October 1, 2026.

Still open: Ultrafast pricing, an end date for gpt-6-sol, a leaderboard row

OpenAI says an Ultrafast mode for GPT-6.1 Sol, “with up to 8x faster token generation” than its standard speed in Codex, is coming, and its pricing page has that tier only for GPT-6 Astra, at $60.00 input and $300.00 output. The announcement gives no end date for gpt-6-sol; its model page is still up and points to “the newer Sol model”. The comparison with Sonnet 5.5 firms up when Zapier’s leaderboard shows a GPT-6.1 Sol row next to the Sonnet 5.5 ones.