GPT-6 Sol vs Claude Opus 5.5: the official numbers and when to use which

GPT-6 Sol and Claude Opus 5.5 they came out on the same day, September 22, 2026, and precisely for this reason neither producer compared them. OpenAI measures Sol against Claude Opus 5, Anthropic measures Opus 5.5 against GPT-6 Astra and GPT-5.6 Sol. The honest answer to "what is better" is that with the official data you can not establish: you can put next to what each published and understand where the numbers speak and where not.

The practical framework, for an SME, is this. Sol costs half per token, $2 and $10 versus $4 and $20 per million, and works in ChatGPT and Codex. Opus 5.5 has the highest number on the only test with a common reference at the two tables, AutomationBench, and holds the same price up to 1 million context tokens. Verified data on September 23, 2026 on the launch pages of OpenAI and of Anthropic: before deciding always double-check the official pages.

Published issues, side by side

The table only reports values written on the two launch pages, each with the conditions stated by the manufacturer. Where a box says "unpublished," that manufacturer didn't exactly report it: it doesn't mean zero.

Eleven lines, and only three with the same name on the two sides. It is already a first answer: the two manufacturers have chosen almost entirely different tests to tell their model. To read tables like this there is a guide in how to read AI benchmarks.

Why These Numbers Don't Compare 1:1

They went out the same day

OpenAI measured Sol against Claude Opus 5 and Claude Fable 5.1, using Fable 5 where the data was missing, with scores of competitors taken from public reports. Anthropic put Opus 5.5 next to GPT-6 Astra and the old GPT-5.6 Sol. The two new models never appear in the same table.

Same name, different proof: the OSWorld 2.0 case

OpenAI states that it reports the partial score on the offline set of release v2026.08.08. Anthropic writes only "partial". A model present in both tables shows that these are not the same measurement: for OpenAI Claude Opus 5 at medium effort is 60.3%, for Anthropic Opus 5 itself is 74.0%. Between test version and effort level almost 14 points dance: Opus 5.5's 81.8 and Sol's 60.5 stand on two different scales.

The effort: written on one side, implied on the other

OpenAI puts the reasoning level next to every figure. Anthropic uses a general rule, maximum effort for Opus 5.5 unless otherwise stated, but for AutomationBench it refers to Zapier’s test without stating the effort used. How much that matters is shown by Anthropic itself: on CursorBench Opus 5.5 is 52.5% at the default effort and 57.8% at the most, five points difference without changing model. The theme is deepened in reasoning effort: what changes with regard to.

FrontierCode: Sol's number isn't there

For Sol OpenAI only writes that it equals Claude Fable 5.1 at xhigh effort, at a much lower cost. In the table of Anthropic Fable 5.1 it is worth 50.3, but to derive a score from there for Sol means to jump between two tables with different conditions. The data, simply, does not exist.

The only bridge between the two tables: Claude Opus 5 on AutomationBench

There is a line where the two tables talk. On AutomationBench, Zapier's test of workflows between different apps, Claude Opus 5 is worth 26.9% in both OpenAI table, where it is indicated at max effort, both in that of Anthropic. With a common reference the two new numbers can at least be put in line: Opus 5.5 to 40.0, Sol to 33.2 with xhigh effort.

An indicative comparison remains. Anthropic does not indicate version of the test and effort of the Zapier test, and specifies that it was done without backup models: each intervention of the protections counts as failure, and according to Anthropic this lowers the score compared to actual use. OpenAI instead declares for Sol a cost of 0.27 $ per activity, while Anthropic for this line does not write a cost per activity.

Also on this line the effort weighs. GPT-6 Astra appears with 30.3% low effort in OpenAI table and with 41.4% in that of Anthropic, taken from the public ranking of Zapier: eleven points for the same model. The comparison between Astra and Opus 5 we had done in Claude Opus 5 vs GPT-6 Astra.

Prices and context: where Sol costs half and where not

The "half-price" claim applies to standard requests. Above 272,000 input tokens, OpenAI charges double for input and one and a half times the usual output rate across the entire request, while Anthropic applies the same Opus 5.5 pricing throughout its 1-million-token window. In that range, input costs the same, $4 versus $4; Sol output remains cheaper, $15 versus $20; and Opus 5.5 caching costs half as much. The complete Sol pricing, including Flex and other modes, is in the guide to GPT-6 Sol and Luna prices.

Examples are built to show the trend, with assumed tokens and not counting the first cached writing. The reasoning is paid in both cases as output, for OpenAI how for Anthropic: the real account depends on how much the model reasons about your tasks. On how the management of such large contexts changes there is context window: how much it really matters.

An operational detail weighs on agents alternating simple and difficult steps. OpenAI that on GPT-6 change the effort between calls no longer breaks the cache. In Anthropic documentation an example with Opus 5.5 shows that, with the cache point in the messages, switching from medium to low effort restarts the cache from scratch. Whoever designs the agent must know this before choosing.

Which one to try first, work for work

An e-commerce of spare parts that connects orders, warehouse and courier. Think of an online store where every order has to update the stocks, generate the shipping label and alert the customer. It is the territory of AutomationBench, and the published data suggests starting from Opus 5.5. But Sol costs half per token: the right proof is thirty real orders, the same effort for both, counting how many activities come to the bottom without human intervention and how much each costs.

A technical study that reads chapters with dozens of attachments. Here many requests exceed 272,000 tokens, and the price difference almost disappears. It is advisable to choose on the quality of the reading: how many inconsistencies between specifications and attachments finds each model on three races already closed, of which you know the right answers. On the risk of invented data the two producers state different things: in Anthropic internal test 16 out of 18 reports from Opus 5.5 passed a verification where only one invented number failed them, while OpenAI claims for Sol about half of GPT-5.6 Sol's factual errors.

Four bad readings going around these days

• "On OSWorld Opus 5.5 it detaches Sol by 21 points." The two numbers come from different versions and conditions: the same Opus 5 changes by almost 14 points between tables.
• "GPT-6 Sol makes 37.3 on Terminal-Bench." That value in the Anthropic table is GPT-5.6 Sol, the previous model.
• "On FrontierCode Sol it's worth as much as Fable 5.1, so about 50." It is a number obtained by crossing two different tables: OpenAI did not publish it.
• "Sol always costs half." Only below 272,000 input tokens. Above, the input costs the same, the output of Sol remains cheaper and the cache of Opus 5.5 costs half.

The verdict, case by case

There is no single winner, and the announcer is comparing numbers that the producers themselves have not put on the same scale. There is instead a map: Sol where the price per token, the volumes of code and the ChatGPT and Codex ecosystem count; Opus 5.5 where the flows between multiple apps and very long documents count. For everything else, evidence about your cases decides more than any table. The reason for the limitations of these comparisons is told in because vendor benchmarks are not comparable.

If you miss the starting frame, the two OpenAI models are explained in What are GPT-6 Sol and Luna, what changes with Opus 5.5 in Claude Opus 5.5: what it is and to whom it suits, its complete table with Fable 5.1 and GPT-6 Astra in benchmarks and prices by Claude Opus 5.5, and the choice within the OpenAI family in GPT-6 Sol vs Luna. When the agent you have in mind has to choose between the two, we can set the test with you on the same tasks and read together the cost per task finished: we start from Custom AI Agents.