The three frontier models are within two points of each other on the independent index, and fivefold apart in price. Every other comparison you read is a lab quoting its own benchmark.
| Model | Lab | Input / MTok | Output / MTok | AA index v5.0 | Benchmark the lab reports |
|---|---|---|---|---|---|
| Gemini 4 Carbon | $2 | $10 | 59 | SEC-bench Pro 88.2% | |
| Claude Opus 5.5 | Anthropic | $4 | $20 | 58 | Terminal-Bench 4.0 66.4% |
| GPT-6 Astra | OpenAI | $10 | $50 | 57 | Terminal-Bench-Science 0.1 64.6% |
The numbers that do agree
On Terminal-Bench-Science 0.1, OpenAI reports Astra at 64.6% and Claude Fable 5.1 at 52.6%. Anthropic's own Opus 5.5 table reports Astra at 64.6% too, with Opus 5.5 at 58.7%. Two labs, one test, the same number for the same model. That is what agreement looks like.
The numbers that do not
Anthropic reports Astra at 57.9% on Terminal-Bench 4.0. OpenAI's Astra announcement does not mention Terminal-Bench 4.0 at all. Google's Carbon announcement leads with SEC-bench Pro, a cybersecurity test, and Collinear's patching benchmark at 71.0%. Three labs, three different first table.
None of them is lying. Each is choosing.
Price is the part nobody argues about
Carbon launches at $2 and $10 per million tokens, rising to $4 and $20 after the introductory period. Opus 5.5 is $4 and $20. Astra is $10 and $50. If the scores are within two points, the price is the decision.
Anthropic says as much on its own page: "at these levels of capability we've found that benchmark margins have become a less reliable guide to real-world differences."
Two things to check before you switch
Carbon is limited at launch to a set of cyber defenders through Google's Fairwind Program, so a general customer may not get it yet.
And Astra's own headline scores — 98% on FrontierMath Tier 4, 99.9% on ARC-AGI-3 — are tests the other two labs do not publish results for. A number with no second reading is a claim, not a comparison.