Claude Opus 5.5 vs GPT-6 Astra: Benchmarks, Price and Fit

Claude Opus 5.5, released on September 22, 2026, lists at $4 per million input tokens and $20 per million output tokens. GPT-6 Astra, released on September 3, lists at $10 and $50, and bills the whole request at $20 and $75 once input passes 272,000 tokens. On price, Opus 5.5 is the cheaper model by a wide margin at every request size.

The benchmarks are closer, and harder to read. Only seven tests appear in both vendors’ published tables. Opus 5.5 leads on four, GPT-6 Astra leads on two, and one isn’t comparable because the two vendors used different task sets. On the tests where both vendors ran the same Claude model, their scores differed by up to 5.4 points, which is more than the gap between the two models on some rows.

Every figure below is a vendor’s own published number, from Anthropic’s announcement and OpenAI’s launch post, and none has been independently replicated. OpenAI’s launch post predates Opus 5.5, so it has no Opus 5.5 figures.

According to OpenAI’s model page, prompts with more than 272,000 input tokens are priced at 2× the input and cache rates and 1.5× the output rate “for the full request,” not just the tokens above the line. Anthropic’s pricing page says Claude models from 4.6 onward include the full 1M-token window at standard pricing, so a 900,000-token request costs the same per token as a 9,000-token one.

The table below prices one illustrative agent task at the same token counts on both models. The counts are invented for the example; the rates are the published ones. In practice the counts won’t be identical, because the two models use different tokenizers and each vendor says its model uses fewer tokens per task. Treat this as the list-price gap, not a prediction of your bill.

At the same token counts, Astra costs about 2.8× as much as Opus 5.5 on this task below the line, and about 5× as much when every request crosses it. The 272,000-token line matters most for coding agents, because an agent that loads a large repository into context can cross it early in a session and stay above it. Our long-context pricing reference lists where each vendor draws its line.

Both vendors report each model at its best setting. Anthropic’s footnote says Opus 5.5 was run at max effort except on Terminal-Bench 4.0 (xhigh), and that the Astra figures on Terminal-Bench come from OpenAI. OpenAI’s table says its scores are “the maximum at any effort.” So this is a comparison of each model’s best case, not of the settings you are likely to run.

Three of Opus 5.5’s leads are large enough to survive the caveats below: Terminal-Bench 4.0 (8.5 points, against a stated standard error of ±2.6), GDPval-AA (304 Elo) and Humanity’s Last Exam with tools (10.5 points). Astra’s lead on Terminal-Bench-Science is 5.9 points, a little above the ±3.5–5 point standard error Anthropic gives for that test. AutomationBench (1.4 points) and FrontierCode (1.1 points) are too close to call from vendor numbers.

OSWorld is the row to be careful with. OpenAI reports Astra at 72.6% on the offline subset of OSWorld 2.0, with Claude models run on the official settings. Anthropic reports Opus 5.5 at 81.8% and leaves Astra’s cell blank. OpenAI’s own footnote says the Fable 5.1 system card used modified tasks and grading. Until someone runs both models on the same set, there is no computer-use comparison to draw.

OpenAI’s appendix scores several Claude models, and Anthropic’s table scores the same models, which lets us check how far the two harnesses agree. On most of these benchmarks they agree to within about a point. On FrontierCode Main they are 5.4 points apart on Opus 5, and on OSWorld the task sets differ.

The FrontierCode gap is the one that matters here. Opus 5.5 leads Astra on that benchmark by 1.1 points, and the two vendors’ scores for Opus 5 differ by nearly five times that. OpenAI’s footnote explains one difference: Astra was run with a developer message modelled on Codex that asks for clean, mergeable code. A developer message changes the result, so the benchmark alone can’t tell you which model writes better code in your repository.

The practical rule is simple. Where the gap between two models is larger than the gap between two vendors’ runs of the same model, it probably reflects a real difference. Where it is smaller, it could reflect the harness as much as the model, and only your own tasks can settle it. Our guide to testing a model on your own traffic covers how to set that up.

Most of what each vendor published has no counterpart from the other. These are the results a buyer is most likely to ask about. Where OpenAI scored Opus 5, its figure is shown as the nearest reference. Opus 5.5 may score differently.

Two of these matter for model choice. The long-context retrieval score (96.3% at 512K–1M tokens) is evidence that Astra reads a very long prompt accurately, though it also means using the part of Astra’s context window that costs the most. The DeepSWE and BrowseComp rows show Astra only narrowly ahead of Opus 5 in OpenAI’s own runs. That makes those benchmarks worth running on Opus 5.5 before assuming Astra keeps the lead. Our Astra launch guide covers the rest of OpenAI’s appendix.

Both models use the same five effort names, from low to max, and both can change effort mid-conversation without breaking the prompt cache. The practical differences are in integration and safeguards. Astra’s details come from OpenAI’s model guide; Opus 5.5’s come from Anthropic’s documentation.

The safeguard row matters most for unattended agents. When an Opus 5.5 safeguard triggers, the request is answered by another Claude model and the run continues. When Astra’s misalignment monitoring intervenes in the API, OpenAI says the task stops. Neither approach is simply better. A fallback keeps work moving but changes which model did it. A stop is easier to audit but interrupts the run. Plan your agent loop for whichever behaviour you choose.

On safety evidence, the two vendors again used different tests. OpenAI rates Astra “Critical” for cybersecurity under its Preparedness Framework. On its computer-use stress test, Astra produced misaligned outcomes 2.4% of the time against 11.5% for Opus 5; OpenAI hasn’t tested Opus 5.5. Anthropic reports that Opus 5.5 tried to get around containment boundaries about 85% less often than Opus 5. No test covers both models. Both vendors offer zero data retention to eligible API customers.