Claude Opus 5.5 vs GPT-6 Sol and Luna: prices, benchmarks, platforms

On September 22, 2026, Anthropic released Claude Opus 5.5, the first model of its Claude 5.5 family, at $4 per million input tokens and $20 per million output tokens, 20% below Opus 5. About an hour later, by Simon Willison’s account, OpenAI released GPT-6 Sol and GPT-6 Luna at half the price of their GPT-5.6 predecessors. All three models reached Amazon Bedrock, Vercel’s AI Gateway and GitHub Copilot the same day.

Each announcement shows only its own price line, and each benchmarks against models the other vendor had already shipped, not against the one released that day. This piece puts the two side by side: the prices including cache reads, what each vendor claims and against which model, where the models run today, and the two Opus 5.5 API changes that turn working requests into errors.

Both vendors cut the price

The table compares the list price of each new model with the model it replaces, per million tokens. The cache read column matters for agents, because a long session resends its context on every turn and most of that is billed at the cache rate.

The Anthropic figures come from its announcement and pricing page. The OpenAI figures come from its announcement and its pricing page, read in a capture from September 24, 2026. OpenAI’s announcement gives only the input and output prices of GPT-5.6 Sol and Luna, so their cached input prices in the table, $0.40 and $0.02, come from Willison’s table, which lists the same models with the current prices.

Anthropic says Opus 5.5 “costs 40% less to run than Opus 5”. That is not a 40% cut to the token price. A post on the Claude blog by Addy Osmani, published on September 25, 2026, explains that the 40% is an estimate for typical workloads at default settings and assumes the model uses fewer tokens per task.

Priced on the same token counts, the post’s example session costs about 31% less, and the largest drop in that example is the cache read line, down 60%.

OpenAI describes the Sol and Luna cut as 50% “compared with their GPT-5.6 promotional pricing”. Willison notes that GPT-5.6 had a 25% price increase scheduled for November, so the new models are half the promotional price, not half of what GPT-5.6 was about to cost.

The prices of the two top models did not move. Claude Fable 5.1 lists at $10 input, $50 output and $0.25 per cached read on Anthropic’s pricing page, and GPT-6 Astra at $10 input, $50 output and $1.00 per cached read on OpenAI’s.

Neither vendor has benchmarked the other’s new model

Anthropic’s announcement compares Opus 5.5 with OpenAI’s GPT-6 Astra and GPT-5.6 Sol. The table lists its figures and, for each competitor score, who measured it, as the page’s footnotes state.

The page adds that “benchmark margins have become a less reliable guide to real-world differences”.

OpenAI’s announcement compares in the opposite direction, against the Claude models that existed before September 22. Effort is the setting that controls how much a model thinks per request, on a scale from low to max, and OpenAI states it for each score.

The FrontierCode row comes from the chart data embedded in the page; the prose says only that Sol “is able to match Claude Fable 5.1 xhigh at much lower cost”. The same chart puts Fable 5.1’s best point at 50.9% at medium effort and $3.28 per task, above Sol at any effort level.

The AutomationBench row needs OpenAI’s own note: the Fable 5.1 run fell back to Opus 5 on about 40% of tasks, and the cost figure “understates its actual cost, as it omits the cost of the Opus 5 fallbacks”. OpenAI’s note on the whole set says competitor scores “were taken from publicly available reports” and that Fable 5 scores were used where Fable 5.1 had none.

Neither page has a number for the model the other vendor released that day. The AutomationBench rows show how different the two framings are: both cite Zapier, Anthropic lists Opus 5.5 at 40.0% and OpenAI lists Sol at 33.2%, and Anthropic’s footnote says its Opus 5.5 run came from Zapier’s evaluation during early access, with safeguard interventions counted as failures.

All three models run on Bedrock, Vercel and Copilot since launch day

The table lists where each model is available, from Anthropic’s announcement, the AWS announcements for Opus 5.5 and Sol and Luna, Vercel’s changelog entries for Opus 5.5 and Sol and Luna, and GitHub’s changelog entries for Opus 5.5 and Sol and Luna.

On the Claude Platform and in Claude Code, Opus 5.5 also has a fast mode at up to 2.5 times the speed, for $8 input and $40 output per million tokens. Copilot bills all three models under usage-based billing, and GitHub’s note on Opus 5.5 says that in early testing it “resolved tasks comparably to Claude Opus 5 while using significantly fewer steps and tokens”.

Two Opus 5.5 changes that break existing requests

Vercel’s changelog lists two API changes in Opus 5.5 “that can turn previously valid requests into HTTP 400 errors”. Thinking is always adaptive, so a request that disables thinking or sets a fixed thinking budget is rejected. Forced tool use is retired, so a request cannot require a tool call or force a specific tool, and the changelog points anyone who forced a tool call to get JSON to structured outputs instead.

Anthropic’s page confirms the first change: Opus 5.5 “is no longer available with ‘thinking’ mode switched off”.

Three other details affect the bill or the output without breaking requests. The Claude blog post says the default effort on Opus 5.5 is medium, one level below the Opus 5 default of high. At a given level, it adds, Opus 5.5 thinks more per turn than Opus 5, so an effort level carried over from Opus 5 does not mean the same spend.

The same post says that with an API key or a Claude subscription a change of effort keeps the cache. On Amazon Bedrock, Google Cloud’s Agent Platform or a gateway it still clears the cached conversation, and the next request pays the cache write price on all of it. And GitHub’s note says Opus 5.5 watermarks its text outputs without adding tokens or cost, which Anthropic describes as a measure “to comply with the EU AI Act”.

On the OpenAI side, the prompt caching post says GPT-6 models keep the cache when reasoning effort changes between responses, that cached prefixes stay eligible within a 30-minute window, and that explicit cache breakpoints let the developer choose which prefixes to reuse.

Our read: cache reads decide the bill

The headline comparison, $4 against $2 on input, is the least useful line in the table for anyone running a coding agent. Anthropic itself says cache reads “make up the majority of agentic and coding work costs”, and after this cut Opus 5.5 and GPT-6 Sol have the same $0.20 per million cached tokens. What separates them at list price is output, at $20 against $10, and on Opus 5.5 that is where thinking is billed, as the Claude blog post states.

That makes the real question how many tokens and turns each model needs to finish a task, and no source has that number for the two new models against each other. Anthropic claims fewer tokens per task than Opus 5, GitHub reports fewer steps than Opus 5, and OpenAI’s cost-per-task claims are against Opus 5, Fable 5 and Fable 5.1, not Opus 5.5.

Willison’s post adds an independent measurement, and it concerns the effort setting rather than cost per task: at the max thinking level, Opus 5.5 hit its 128,000 output token limit twice while still reasoning about an SVG, at $2.56 per attempt.

What to watch next is the low end and the independent runs. Anthropic says Claude Sonnet 5.5 and Claude Haiku 5.5 “will follow in the coming weeks”, and Sonnet 5.5 arrived six days later, at Sonnet 5’s price; Luna at $0.10 and $0.50 sets the price they get measured against. Until a third party runs Opus 5.5 and GPT-6 Sol on the same tasks with the same harness, the cost-per-task claims on both pages remain each vendor’s own. The follow-ups go in our News section, under the Claude and GPT-6 tags.

Related posts:

Claude Cowork: Anthropic Introduces an AI Agent That Works Directly With Your Files Claude Sonnet 5.5 vs Opus 5.5: cost per task, where each fits GPT-6.1 Sol: what changed from GPT-6 Sol and what a task costs OpenAI Launches ChatGPT Health: New AI Tool Focused on Healthcare with Reinforced Privacy A New DeepSeek Paper Highlights the Growing Push for More Efficient AI Training Gemini 4 Argon: announced at $2 and $10, not in the API yet