26 SEP 2026 — Two model releases landed on 22 September. Anthropic put Claude Opus 5.5 on sale at US$4 per million input tokens and US$20 per million output, a fifth below Opus 5 and the first time an Opus release has arrived cheaper than the one before it.
Hours later OpenAI priced GPT-6 Sol at US$2 and US$10, exactly half Anthropic's new rate.
What the new rate card says
Opus 5.5 charges US$4 per million input tokens and US$20 per million output. Opus 5 charged US$5 and US$25, so both halves fall by 20 per cent.
The deeper cut is in caching. Reading from cache drops from US$0.50 per million tokens to US$0.20, a reduction of 60 per cent, and five-minute cache writes fall from US$6.25 to US$5. For an application that replays a long system prompt on every call, that line decides the bill.
Anthropic makes a second and larger claim. At default settings, Opus 5.5 will cost 40 per cent less than Opus 5 on typical workloads. That is not the per-token cut. The saving depends on how many tokens the model spends, and it comes from Anthropic's own testing.
One setting sits behind that claim. Opus 5.5's default effort level is medium; on Opus 5 it was high. Anthropic does not attribute any of the saving to the change, and a developer who never set the value is comparing two different defaults.
A year of held Opus prices
The 20 per cent cut matters because the Opus line had not moved before. When Opus 5 arrived on 24 July, this desk recorded that it shipped on the same rate card as Opus 4.8, US$5 and US$25, and read that as a capability claim at a held price rather than a price cut.
Sonnet 5 came close to a rise and did not take it: a scheduled increase to US$3 and US$15 was withdrawn on 10 August, leaving US$2 and US$10 as the permanent rate. Opus stayed where it was throughout, and Opus 5.5 is the first break in that pattern.
Opus 5.5 batch processing costs US$2 and US$10, exactly GPT-6 Sol's standard price.
The scorecard names the wrong rival
Anthropic's announcement includes a benchmark table. On Terminal-Bench 4.0 it puts Opus 5.5 at 66.4 per cent against 57.9 for GPT-6 Astra, 55.8 for Claude Fable 5.1, 52.3 for Opus 5 and 37.3 for GPT-5.6 Sol. CursorBench 4.0 reads 57.8 per cent and FrontierCode v1.1 reads 54.4.
The comparison column says GPT-5.6 Sol. GPT-6 Sol shipped the same morning and Anthropic's table does not contain it. OpenAI's own figures do not contain Opus 5.5 either. Each company measured itself against the other's previous model, and both tables were out of date by the afternoon.
Opus 5.5 does not lead everywhere even on its own scorecard. GPT-6 Astra scores higher on Terminal-Bench-Science 0.1, 64.6 per cent against 58.7, and on AutomationBench, 41.4 against 40.0.
Every Opus 5.5 figure in that table uses adaptive thinking at maximum effort, which is not the default setting the cost claim is measured at. Anthropic says the model was tested before release by external evaluators including Frontier Design and METR, on safety rather than on these coding scores.
OpenAI answered the same morning
GPT-6 Sol costs US$2 per million input tokens and US$10 per million output, exactly half GPT-5.6 Sol's US$4 and US$20. GPT-6 Luna costs US$0.10 and US$0.50, half its predecessor on input and 58.3 per cent less on output.
Luna's input price has now fallen twice in two months. It launched at US$1 per million tokens, dropped to US$0.20 on 30 July when OpenAI cut its cheapest tier by 80 per cent, and now sits at US$0.10. That is a tenth of the launch price in two months. OpenAI attributed the July cut to efficiency rather than to competitive pressure.
Sol's price matches Claude Sonnet 5 to the cent and sits at half Opus 5.5. VentureBeat reported that an OpenAI spokesperson described the new rates as "permanent prices, not promotional or introductory pricing".
That wording matters more than the figures. A promotional rate can be withdrawn once developers have built against it; a permanent one is a standing commitment, and OpenAI was asked and answered on the point.
OpenAI reports Sol at 68.8 per cent on DeepSWE 1.1, 60.5 on OSWorld 2.0, 56.4 on Agents' Last Exam and 33.2 on AutomationBench 1.0.6 at US$0.27 per task. Both new models remain below GPT-6 Astra, the company's flagship.
Independent scores, and a crossover
Artificial Analysis runs its own tests rather than repeating either company's. On version 4.3.2 of its Intelligence Index it scores Opus 5.5 at 58 at maximum effort, the highest of the 172 models it lists, and GPT-6 Sol at 48. At medium effort the two read 51 and 40.
Cost runs the other way. Opus 5.5 costs US$5.98 per task at maximum effort and US$1.34 at medium; Sol costs US$1.06 and US$0.25. Opus leads by ten or eleven index points and costs between five and six times as much to run.
Read across the effort levels rather than down them and one rung inverts. Sol at high effort scores 43 for US$0.37 a task, while Opus 5.5 at low effort scores 42 for US$0.55. At that rung the cheaper model is ahead on both numbers. Above it the ordering returns: Opus at its default medium setting scores 51 for US$1.34, beating Sol's best result of 48 for about a quarter more per task.
Every price in this article is taken from the two companies' own published pages rather than from any third-party index. Artificial Analysis builds its cost per task from input, cache and reasoning tokens as well as answer tokens, which is why the gap between the two models is wider than the rate cards alone would suggest.
One caution about the index itself. In early September this desk reported that Artificial Analysis scored GPT-6 Astra at 61, level with GPT-5.6 Sol at the time. On version 4.3.2 Astra reads 53 at maximum effort. Scores are not comparable across index versions, and the page carries no note explaining the change.
Where maximum effort broke down
Simon Willison tried Opus 5.5 at maximum thinking on 22 September with the SVG drawing prompt he runs against every new model. It returned nothing at all. He reports that the model has a "128,000 maximum output token limit" and hit that ceiling while still reasoning about the picture. Two attempts, each costing about US$2.56 and taking close to twenty minutes, produced no answer.
His judgement was that a model which over-thinks to breaking point on a trivial drawing prompt cannot be trusted not to do the same on work that matters.
Anthropic's own migration notes point at the same edge. They say Opus 5.5 thinks more per turn than Opus 5 at the same effort setting, most of all at the two highest levels, and tell developers to leave room in the output budget for that thinking. Willison's result shows what that instruction looks like at the top setting.
On price Willison is unambiguous, calling it "hard to overstate how competitive this pricing is". He also notes that GPT-5.6 Terra, now priced level with GPT-6 Sol, has lost any remaining reason to be used.
Both companies cut prices on the same morning, and each published a scorecard measuring itself against the other's older model. The independent index says Opus 5.5 is the more capable and Sol the cheaper, which is what the two rate cards already implied. Neither table tells you whether maximum effort is worth paying for on work that is not a benchmark.
View author profile → · Editorial policy
