Claude Opus 5 Hits Number One on the AI Leaderboard, and the Real Story Is What It Costs

Claude Opus 5 claimed the top spot on the Artificial Analysis leaderboard the day Anthropic released it, overtaking Claude Fable 5, the company's own frontier model, at roughly half the price 1.

Opus 5 costs $5 per million input tokens and $25 per million output tokens, matching its predecessor Opus 4.8 and running about half of Fable 5's rate 2. It delivers near-frontier benchmarks at that price. And it is already being undercut: China's Kimi K3, an open-weight model at similar performance, charges $15 per million output tokens 2. Ten dollars per million tokens is the kind of spread that rewrites procurement decisions at scale.

The capability story is genuine, just not the differentiator. Anthropic describes Opus 5 as a "thoughtful and proactive" model 1. Anthropic reports that the most striking demonstration came on a Frontier-Bench task: Opus 5 was given a drawing of a machine part and asked to reconstruct it in FreeCAD, with no tool to view the image. The model wrote its own computer vision pipeline from raw pixels, extracted the geometry, and built the full 3D part 1.

The benchmark gains, though, are incremental rather than a leap 2. On Anthropic's own benchmark charts, Opus 5 appears to edge past OpenAI's GPT-5.6-Sol and Anthropic's own Opus 4.8 in nearly every category, but the improvements look modest update to update and only dramatic when you stack a full year of releases 2. The competition between frontier labs is tilting from raw intelligence toward unit economics.

That tilt is already reshaping how the market consumes intelligence. Cursor and Meta are building "model routers," systems that automatically pick the cheapest model capable of handling a given prompt 2. When a cheaper model handles most tasks, the frontier tier becomes a specialty call reserved for the hardest workloads. Anthropic's decision to hold Opus 5 at the Opus 4.8 price point while delivering better benchmarks is a direct response to that pressure. The company is not cutting prices. It is offering more intelligence per dollar, which buys time but not a moat when open-weight rivals keep lowering the floor.

The non-price differentiator is safety. Anthropic intentionally withheld cybersecurity exploitation training from Opus 5, as it did with Opus 4.8 1. The model comes close to Anthropic's restricted Mythos 5 at identifying security vulnerabilities but remains substantially behind on weaponizing them 1. Anthropic frames this gap as deliberate. When token prices converge across labs, a model that can find a vulnerability but will not help exploit it becomes the enterprise-compliant default.

The convergence is already visible. Opus 5 at $25 per million output tokens, Fable 5 at roughly double, Kimi K3 at $15. The spread between the priciest and cheapest near-frontier options is narrowing to a rounding error for anyone running serious volume. A development team that once sent every prompt to a frontier model can now route routine work to a $15 option and reserve the expensive calls for genuinely hard problems. The leaderboard decides who is smartest. The invoice decides who gets used.