Inside the Claude Opus 5 Numbers: Frontier-Bench Doubled, ARC-AGI-3 at Nearly Four Times GPT-5.6

Anthropic shipped Claude Opus 5 on July 24, and the launch-day noise has mostly been about positioning. The benchmark tables deserve a closer read — because the pattern in the numbers says more about where model progress is coming from than any single headline score.

The short version: Opus 5 posts a full generation of gains over Opus 4.8 at the same price, takes outright leads on the hardest new evaluations, and lands within a point of Anthropic's own flagship-class models on coding — at a fraction of the cost per task.

Coding: The Saturation Problem, Then the Real Test

On SWE-bench Verified, the industry's long-standing agentic coding standard, Opus 5 scores 96.0% — a number that essentially retires the benchmark. The action has moved to Scale AI's harder SWE-bench Pro, built by its SEAL evaluation lab specifically because Verified stopped discriminating between frontier models.

There, Opus 5 lands at 79.2%, a ten-point generational jump over Opus 4.8's 69.2%. Notably, the two models ahead of it on the leaderboard are both Anthropic's own: Claude Mythos 5 at 80.3% — the restricted-access, unsafeguarded configuration — and Claude Fable 5, the production-safeguarded version of the same Mythos-class weights, at 80.0%. The top three slots on the industry's toughest coding benchmark now all belong to one lab, separated by 1.1 points.

On CursorBench 3.2, which measures real-world editor-integrated coding, independent trackers place Opus 5 within half a point of Fable 5 — at roughly half the cost.

The New Benchmarks Are the Story

The more interesting results come from evaluations designed after the last generation saturated the old ones. On Frontier-Bench v0.1, Opus 5 scores 43.3% — roughly double Opus 4.8's performance and clearly ahead of both OpenAI's GPT-5.6 Sol (34.4%) and Fable 5 itself (33.7%).

The starkest gap is on ARC-AGI-3, the abstraction-and-reasoning benchmark built to resist memorization: Opus 5 scores 30.2% against GPT-5.6 Sol's 7.8% — nearly a four-fold margin on the evaluation most closely watched by researchers skeptical of benchmark gaming. And on GDPval-AA v2, which measures economically valuable knowledge work as an Elo rating, Opus 5 tops the leaderboard at 1,861, ahead of Fable 5 (1,747) and GPT-5.6 Sol (1,736).

Agentic evaluations follow the same pattern. Trackers report Opus 5 surpassing Fable 5 on OSWorld 2.0 computer-use tasks at roughly a third of the compute budget, and posting pass rates around 1.5 times its competitors on Zapier's AutomationBench.

The Efficiency Lever

The mechanism behind the cost numbers is an "effort" toggle — low, medium and high settings that dial reasoning compute up for hard problems and down for routine ones, with extended thinking on by default and a 1-million-token context window throughout. At $5 per million input tokens and $25 per million output ($10/$50 for a fast mode at roughly 2.5x speed), Opus 5 launches at Opus 4.8's price while, on several evaluations, beating models that cost multiples more per task.

That is the same playbook China's open-weight labs have been running all year — maximum capability per inference dollar — and it is hard not to read Opus 5 partly as an answer to it. As of July 27, BenchLM ranks Opus 5 the top model overall at 85.88/100, but Moonshot's Kimi K3, whose 2.8-trillion-parameter open weights landed the same day, sits at the top of the open class with aggressive sparsity delivering exactly this kind of cost-capability ratio. The competitive frontier is no longer peak score; it is score per dollar.

Caveats

Two things the tables don't show. First, several of the strongest results — Frontier-Bench, GDPval-AA v2 — are new evaluations with short track records, and early leads on young benchmarks have a history of compressing once rivals tune for them. Second, the launch-week numbers mix Anthropic's own reporting with third-party trackers; fully independent replication, particularly on the agentic suites, will take weeks. But the generational delta over Opus 4.8 is consistent across every evaluation with a comparison point — and a ten-point jump on SWE-bench Pro at unchanged pricing is the kind of result that reprices the market whether or not the newer benchmarks hold.