GPT-6 Astra vs Claude Opus 5.5 on Science Tasks: 4x Cheaper per Task?

On Artificial Analysis's Intelligence Index, GPT-6 Astra at Max effort costs $3.26 per task and Claude Opus 5.5 at Max costs $5.98 (captured 14 October 2026, independent). That is about 1.8x, not 4x. The 4x figure does exist, but it describes output tokens: 27k for Astra against 119k for Opus 5.5 per task.

On science scores the two models split by task type. Astra leads the agentic Terminal-Bench-Science 0.1, while Opus 5.5 leads Artificial Analysis's SciCode and Humanity's Last Exam numbers. The only verified cost figures are index-wide, not science-only.

Is GPT-6 Astra really 4x cheaper per task than Claude Opus 5.5?

No. The 4x is a token ratio, and the cost ratio at the same effort setting is smaller. Artificial Analysis (AA) reports 119k output tokens per task for Opus 5.5 and 27k for Astra, a ratio of about 4.4 (my arithmetic).

Cost per task is a different number. AA's page lists $5.98 for Opus 5.5 and $3.26 for Astra, a ratio of about 1.8. The gap narrows because the per-token prices run the other way: Opus 5.5 is $4 input and $20 output per million tokens, while Astra is $10 and $50, both as of 15 October 2026. A 4.4x token gap divided by a 2.5x price gap gives roughly 1.8x on output spend alone.

Opus 5.5 is faster per token but slower per task, because it generates far more tokens. All figures are AA's, and AA computes cost from input, cache hit, cache write, reasoning and answer token prices weighted across index tasks.

Does the cost gap change with reasoning effort?

Yes, and it flips below the top settings. According to Lindy, which reproduces AA's per-effort results, Opus 5.5 is cheaper per task at low and medium effort and Astra is cheaper from high effort up. Only the Max row is confirmed on AA's own page; the other rows rest on Lindy's table.

Source: Lindy's reproduction of Artificial Analysis data, index v4.3.2, independent, captured 14 October 2026. The Max row matches AA's page.

Read the table by score, not by label. Opus 5.5 at High scores 54 for $1.82, which is above Astra at Max (53 for $3.26). On that matched-score comparison, Opus 5.5 is the cheaper model by about 1.8x, so the headline claim inverts. Opus 5.5's own default is medium effort, according to Anthropic's docs. Astra's reasoning.effort accepts low, medium, high, xhigh and max.

For a related look at how cost-per-task numbers move with setup, see Gemini 4 Argon Cost per Task and Benchmark Cost Per Task and Fallback Routing.

Which model scores higher on science benchmarks?

It depends on the benchmark. Astra leads Terminal-Bench-Science 0.1 in lab-reported and independent runs. Opus 5.5 leads on AA's SciCode and Humanity's Last Exam, and the two tie on CritPt.

Different benchmark versions and settings sit in this table, so read the "version and setting" column before comparing rows.

Sources: OpenAI's GPT-6 Astra announcement for Astra's lab numbers, and aicatchup's reproduction of Anthropic's launch table and Cosmic JS for Opus 5.5's. AA figures come from its comparison page and its Terminal-Bench-Science leaderboard.

Lindy notes that Anthropic did not publish GPQA Diamond or FrontierMath for Opus 5.5, so those two cannot be compared head to head. Astra's 96.0% and 97.6% are the company's own claims. For context on why FrontierMath-style scores stop separating models, see AI Benchmark Saturation 2026.

What is Terminal-Bench-Science 0.1, and who runs it?

It measures whether an agent can complete research workflows in a terminal. AA describes it as a 70-task benchmark of workflows authored and reviewed by domain experts across the life, physical, earth, mathematical and engineering sciences, each checked by its own set of tests.

AA scores it as the share of tasks passed (pass@1), averaged over three repeats. The Terminal-Bench team says the tasks are drawn from real scientific workflows contributed by practicing scientists. It is a terminal task set, not a question-answering test like GPQA Diamond, so it rewards tool use and persistence as well as domain knowledge.

The benchmark is an academic collaboration. The Terminal-Bench team says it is led by researchers at Stanford University and built with domain experts at research institutions. Snorkel AI's page lists Stanford and the Laude Institute as hosts, and Snorkel supports it through a grants program.

The same Snorkel page lists the leads as Steven Dillmann, Sanmi Koyejo and Ludwig Schmidt, with Schmidt's affiliation given as Stanford University and Anthropic. That is a disclosed link to one of the two labs in this comparison. It is not evidence of bias, but readers weighing Anthropic's lab-reported score should know it. AA, which produced the independent numbers, is a separate operator from the benchmark's hosts.

Why do lab-reported and independent science scores disagree?

The settings are not the same, and the gaps are small enough for settings to matter. OpenAI states that its evaluation scores are the maximum at any effort. AA reports named effort levels: on its leaderboard, Astra (Max) scores 63.3%, Opus 5.5 (Xhigh, Default Fallback) 61.9% and Opus 5.5 (Max, Default Fallback) 59.0%.

Two details deserve attention:

• Opus 5.5 is higher at Xhigh than at Max on AA's run (61.9% vs 59.0%). That is a 2.9 point swing within one model, comparable to the Astra-to-Opus gap.
• Fallback routing. AA labels the Opus 5.5 entries "Default Fallback". Anthropic's launch table notes that when production safeguards step in, an older model completes the task, which Anthropic says likely pulls Opus 5.5's scores down. The AA page does not say whether fallback calls are included in its cost and token figures. For the mechanics, see Benchmark Cost Per Task and Fallback Routing and Claude Opus 5.5 on Terminal-Bench 4.0.

The two labs' tables also disagree slightly on an older model. Anthropic's table lists Opus 5 at 29.0% on this benchmark, while OpenAI's table and the Terminal-Bench team's release page give 30.0%.

What are the weaknesses of this comparison?

Sample size is the main one. With 70 tasks, one task is about 1.4 points (my arithmetic). A naive binomial standard error at a 60% pass rate and 70 independent tasks is about 5.9 points. Three repeats reduce run-to-run noise but do not add tasks, so that figure is an upper-end sketch, not a published interval. aicatchup states a standard error of 3.5 to 5 points per model without naming the method (single source). Either way, AA's 4.3 point gap between Astra Max and Opus 5.5 Max is not clearly outside noise.

Other limits:

• No confidence intervals appear on the pages cited here for the index or the benchmark scores.
• Version drift. OpenAI's table lists Astra at 61.2 on AA Intelligence Index v4.1.1. AA's current page uses v4.3.2 and shows 53. Scores across index versions do not carry over.
• Index cost is blended. AA's cost per task averages coding, knowledge work and science tasks, so it is not a science cost. SciCode and CritPt are flagged "under review" by AA.
• Asymmetric effort reporting. Astra's lab scores are max at any effort, while Opus 5.5's effort on Terminal-Bench-Science is not stated in the sources cited. aicatchup says Opus 5.5 was run at max effort unless noted.
• Contamination. The sources cited do not describe contamination controls for this benchmark, so none can be assumed.

Should you trust these numbers, and for what?

Trust them as direction, not as a ranking. A rough scorecard for the science claims, scored 1 to 5 from the evidence above:

For procurement, the safer reading is: Astra has a small, uncertain edge on agentic science workflows, and Opus 5.5 is ahead on AA's static science components. Neither result supports a "4x cheaper" claim. If you run science workloads, measure cost per resolved task on your own tasks at the effort setting you will deploy, because the verified data shows the cost ordering changes with effort.

What would a better science comparison look like?

It would fix effort, report cost per resolved task on science tasks specifically, and publish intervals. It would also state fallback handling and tool access for every run.

Concretely: report cost per task per benchmark, not per index. Compare the models at matched scores as well as matched effort labels. Publish per-task results so readers can compute paired differences, which are tighter than comparing two marginal pass rates on 70 tasks. Disclose which headline scores are maximum-over-effort selections, which inflate lab numbers relative to a fixed-setting independent run.

What's still unclear?

• Science-only cost. No science-only cost per task is published in the sources cited here. The $5.98 and $3.26 figures are index-wide.
• Per-effort rows. Only the Max row of the per-effort table is confirmed on AA's own page. The rest comes from Lindy.
• Opus 5.5 lab figures. Opus 5.5's lab-reported numbers (58.7% on Terminal-Bench-Science 0.1, 67.7% on Humanity's Last Exam with tools) are cited from outlets that reproduce Anthropic's launch table, not from Anthropic's own page.
• Fallback effects. Whether AA's Opus 5.5 cost and token counts include fallback calls to an older model is not stated.
• Confidence intervals. None are published on the pages reviewed, so whether Astra's Terminal-Bench-Science lead is statistically meaningful is open.
• Weakest point of this post. The "1.8x cheaper at Max" result and the "Opus cheaper at matched score" result both come from one index, and the second leans on a single secondary table.