Four frontier models have shipped in three weeks: Fable and Mythos 5.1 on 1 September, GPT-6 Astra on the 3rd, Opus 5.5 on the 22nd. Add to that open models that occasionally outperform commercial ones. Here is where the landscape stands — systematically separating what the makers measure from what third parties measure.
Before the tables, a caveat that conditions everything else: leaderboards have become hard to compare, for reasons we have documented one by one.
The five reasons to distrust a leaderboard
The effort level. The same model can vary by tens of points depending on the reasoning budget allowed, and in cost by a factor of five to ten. We detailed this here.
The harness. Leaderboards now compare model + software wrapper pairings, not models alone. Explanation here.
Test decay. Contamination, shortcuts, saturation: an old benchmark mainly measures that the model has seen it. Details here.
The source. A table published by a maker selects its own tests. Anthropic itself acknowledges that benchmark gaps have become an unreliable guide to real-world differences at this level.
The guardrails. An unusual detail noted yesterday: Anthropic's safety systems intervened during its own evaluations, redirecting cyber and biology tasks to weaker models, which lowered its scores on those tests.
The commercial high end
| Model | Release | Input / Output | Key strength |
|---|---|---|---|
| Claude Opus 5.5 | 22 Sept | $4 / $20 | Agentic coding, price-to-capability ratio |
| Claude Fable 5.1 | 1 Sept | $10 / $50 | Independent index, academic tests |
| Claude Mythos 5.1 | 1 Sept | $10 / $50 | Restricted access (cyber, life sciences) |
| GPT-6 Astra | 3 Sept | — | Research maths, business workflows, science in the terminal |
| Claude Opus 5 | 24 Jul | $5 / $25 | Replaced by Opus 5.5 |
Who leads on what
There is no overall winner, and that is the main takeaway.
On the general-purpose independent index, Fable 5.1 was in the lead at our last reading, ahead of Opus 5, with Astra on a par with its predecessor. Opus 5.5 does not yet appear there.
On agentic coding and terminal work, Anthropic places Opus 5.5 ahead of Astra — 66.4% against 57.9% on Terminal-Bench 4.0.
On business workloads and science in the terminal, Astra keeps the advantage in those same tables: 64.6% against 58.7% on Terminal-Bench-Science.
On research mathematics, Astra claimed 97.6% on FrontierMath level 4 against 87.8% for Fable 5.1.
On web browsing, an open agent with 744 billion parameters leads the commercial models.
A ranking limited to commercial models gives a false picture. GLM-5.3-Flash — 320 billion parameters of which 18 are active, a one-million context — is served at scale under an MIT licence. Step 5 Preview offers 600 billion parameters at 1 dollar per million input tokens. K2 Horizon offers a range from 0.9 to 375 billion under Apache 2.0. On generalist benchmarks, the gap between the best closed and the best open model comes down to a few points.
The ranking that really matters: cost per task
It is the only metric that factors in price, tokenizer, cache, effort and quotas. And it produces a different order.
Astra consumes markedly fewer tokens than its competitors for a comparable result in coding — a third of those of its predecessor — but its price has been multiplied by 2.5, which cancels out part of the advantage on general reasoning.
Fable 5.1 was, at our last reading, the most expensive model per task in the independent ranking, which puts its first place in perspective.
Opus 5.5 claims a 40% reduction in real cost compared with Opus 5. If it holds the level of Fable 5.1 as announced, it becomes the best capability-price ratio in the high end. That remains to be verified independently.
What to choose, depending on use
Agents and coding at volume: Opus 5.5 is the most obvious candidate today, for the combination of price and cache reads at 0.20 dollar.
Hard reasoning and mathematics: Astra on its strengths, at high effort, accepting the cost.
Volume, confidentiality, internal hosting: open models, depending on the size that fits your hardware. See our guide to sizes.
Simple tasks in bulk: the smallest model that passes your reference cases. That is where most organisations overpay.
What is coming
Anthropic has announced Sonnet 5.5 and Haiku 5.5 in the coming weeks. If Sonnet 5.5 improves on Sonnet 5 as much as Opus 5.5 does on Opus 5, the best choice on value for money could change again.
This ranking will therefore be out of date. That is exactly why the conclusion does not change: do not choose a model, choose an architecture that lets you change it, and keep around twenty reference cases up to date. They will serve you at every release, and there will be more before the end of the year.