Anthropic released Claude Opus 5, the new frontier model positioned between Opus 4.8 and Fable 5 in intelligence, but at half Fable 5’s price. It’s available starting today on Claude Max as the default model and on Claude Pro as the strongest option.

Why it matters to you. If you use Claude for work, the math changes. Opus 5 costs the same as predecessor Opus 4.8 but more than doubles performance on Frontier-Bench, and on CursorBench reaches within 0.5 percentage points of Fable 5 while spending half as much. For anyone running complex tasks on Fable 5 because it was the only one that could handle them, there’s now an alternative that costs half as much with comparable results.

The model also includes an effort setting that lets you choose between maximum intelligence or conserving tokens for faster, more economical results. Even at the lowest level, Opus 5 passes more tasks than any other model, according to Anthropic’s data from July 24, 2026.

The second point is autonomy. Anthropic reports that Opus 5 better verifies its own work and iterates until it succeeds. The clearest example: they gave it a mechanical part drawing without letting it see it directly, and the model built its own computer vision pipeline to extract geometry from pixels and reconstruct the 3D model. No competing model with the same setup managed it in five attempts.

Limitations remain: on cybersecurity Opus 5 lags behind Mythos 5, and the benchmarks come from the lab that built the model. But 799 points and 451 comments on Hacker News in just a few hours show that the technical community is really testing it.

In detail

What came before. Until yesterday, Claude’s lineup had Opus 4.8 as the mid-tier model and Fable 5 at the top. Fable 5 had been extended multiple times in Max and Team Premium plans with reduced capacity (as we reported on July 18), a sign that demand exceeded supply and Anthropic had only one truly frontier model. Opus 5 changes this geometry: it brings performance close to Fable 5 at half the price, and becomes the default on Claude Max.

The numbers in context. Benchmarks cited are measured July 24, 2026 according to the official announcement. On Frontier-Bench v0.1, Opus 5 more than doubles Opus 4.8’s performance at lower cost per task. On CursorBench 3.2 at maximum effort, it reaches within 0.5 percentage points of Fable 5’s peak, but at half cost per task. On ARC-AGI 3, a problem-solving test on novel problems, the score is triple the second-best model. On OSWorld 2.0, the computer use benchmark, it exceeds Fable 5 while spending roughly a third of the cost. On Zapier AutomationBench, which measures ability to complete end-to-end business tasks, pass rate is 1.5 times the next model at the same cost.

The story the numbers tell is efficiency per dollar spent, more than raw power. It’s a direct message to API consumers paying by usage who need to optimize cost-to-result on every call.

The effort setting. Opus 5 introduces a parameter letting you choose spending level for each step: maximum effort for difficult tasks, token savings for simpler ones. For those building agents, this means calibrating cost on each step of a workflow instead of using the same model for everything.

Autonomy in tests. Examples from early-access testers are the most interesting signal. An engineer from a trading firm used Opus 5 to build a market data feed for a new exchange in a single session: previous models couldn’t complete the task even with detailed plans. Finding no live feed to validate the code, the model built its own test harness to verify data parsing was correct. On a real bug in an open source package manager, Opus 5 found the root cause and fixed an edge case that the community patch had missed; a competing model had only fixed the surface symptom and declared the bug resolved.

Limitations to keep in mind. On cybersecurity Opus 5 still lags Mythos 5. Cited testers are Anthropic commercial partners (CEO of Cursor, Zapier, Lovable, Devin), so they have incentive to speak well of the product. Real-world numbers will arrive in coming weeks when professionals start using it on actual tasks.

To orient yourself on model choice, the guide Mollick updated yesterday remains a good starting point. If you were waiting for Claude’s price-to-performance ratio to improve before building serious agents, this release changes the starting conditions.