Anthropic released Claude Opus 5 on July 24, 2026. It is available in Claude Code, Claude Cowork, the Claude apps, and the API as claude-opus-5, and it is now the default model on Claude Max and the strongest model offered on Claude Pro.

The headline numbers are good. What makes this release worth writing about is not that a frontier lab shipped a better model — that happens on a schedule now — but where the improvement landed. Opus 5 costs the same as Opus 4.8: $5 per million input tokens and $25 per million output tokens. The gains showed up entirely on the other side of the ledger.

That changes the shape of the decision developers have been making for the past two years.

The Old Trade-off Was Intelligence Versus Budget

Anyone who has run an agent loop in production knows the pattern. You start with the strongest available model because it fails less often. Then the bill arrives, and you begin the demotion process: cheap model for the easy steps, expensive model for the hard ones, a router in between, and a growing pile of heuristics deciding which is which.

That routing layer is not free. It is code you maintain, a source of subtle behavioral inconsistency, and a place where the cheap model’s failures leak into the expensive model’s context. Most teams building on LLMs have some version of it, and most would delete it if the economics allowed.

Opus 5 is aimed squarely at that situation. Anthropic’s framing is that it “is designed to be used every day: it works more efficiently than other models.” Efficiency here is not a benchmark category — it is the claim that the model reaches the answer in fewer tokens, which makes the strong model affordable for the boring steps too.

The Benchmark Results, With the Cost Attached

Anthropic’s published comparisons are consistently phrased as score plus cost, which is a more useful format than a bare leaderboard row:

• Frontier-Bench v0.1 — Opus 5 surpasses all other models and more than doubles Opus 4.8’s performance, at a lower cost.
• CursorBench 3.2 — lands within 0.5% of Fable 5’s peak score, at half the cost per task.
• ARC-AGI 3 — Opus 5’s score is three times as high as the next-best model.
• OSWorld 2.0 — beats Fable 5’s best result at just over a third of the cost.
• Zapier AutomationBench — roughly 1.5× the pass rate of the next-best model for the same spend.
• Life sciences — 10.2 percentage points above Opus 4.8 on organic chemistry, 7.7 points higher on protein-related tasks.

On the community leaderboards, Opus 5 is currently sitting at the top of SWE-bench Verified at around 96% — with the usual caveat that different trackers run different harnesses and report numbers a point or two apart.

The ARC-AGI 3 result is the one worth staring at. A 3× margin on a benchmark specifically designed to resist memorization is not the kind of gap you get from more training data on the same recipe. Whether it generalizes to work that looks like your work is, as always, an open question — but it is a different sort of result than another two points on a coding eval.

Sualeh Asif, co-founder of Cursor, summarized it in the terms that matter to anyone shipping a product on top of these models: “Claude Opus 5 delivers near Fable 5 intelligence at Opus speed and cost.”

Effort and Fast Mode: Two Dials Instead of Two Models

The most practically useful change is that the intelligence-versus-cost trade-off has moved inside the model.

The effort setting lets you tell Opus 5 how hard to think. Turn it up to maximize intelligence on a genuinely difficult problem; turn it down to conserve tokens when you want a faster, cheaper answer. This is the router I described earlier, except it is a parameter instead of an architecture, and the model on both sides of the dial is the same model — same knowledge, same tool-calling behavior, same failure modes.

Fast mode is the orthogonal dial: roughly 2.5× the default speed at twice the base price. Note what that actually buys. You are not paying for a better answer, you are paying for the same answer sooner. For an interactive coding session where a human is sitting there waiting, latency is often worth more than tokens. For an overnight batch job, it is worth nothing. Having these as separate knobs means you can make that call per workload instead of per project.

In Claude Code, fast mode toggles with /fast and does not swap in a smaller model underneath — a distinction worth confirming, since “fast mode” in other products frequently means exactly that.

Two API Changes Worth Reading Carefully

Alongside the model, Anthropic shipped two beta features that read like they came directly out of production incident reports.

Mid-conversation tool changes without invalidating the prompt cache. If you have built a long-running agent, you know why this matters. Tool definitions live at the front of the prompt, so changing them — adding a tool as the agent enters a new phase, removing one that is no longer relevant, updating a schema — has historically blown away the cache for the entire conversation. The workarounds were all bad: define every tool up front and accept the context bloat and selection errors, or eat a full re-read on every phase transition. Being able to hand an agent a different toolset partway through a task, without paying for the whole conversation again, makes staged agents meaningfully cheaper to build.

Automatic fallbacks for safety-classifier-flagged requests. When a request trips a safety classifier, it can now be routed automatically to an alternative model — Opus 4.8 by default — rather than failing. If you have ever had a production endpoint return an error because a legitimate user request looked superficially like something else, you understand the appeal. It converts a hard failure into a soft degradation.

Neither of these is glamorous. Both are the sort of thing you only design after watching a lot of real applications break.

The Alignment Number

Anthropic reports that Opus 5 scores 2.3 on overall misaligned behavior — the lowest of any of their recent models. They also note where it does not lead: it remains behind Mythos 5 on cybersecurity exploitation tasks (though comparable at vulnerability identification) and on biology research tasks.

That second sentence is the interesting one. A capability gap disclosed voluntarily in a launch post is a gap the lab decided not to close, and the two named domains are exactly the two where capability and risk are hardest to separate. Reading the omission as a deliberate positioning choice rather than a training shortfall seems reasonable.

What This Means If You Are Building Something

A few concrete implications:

Revisit your model routing. If you built a cheap-model/expensive-model split to control costs, the effort setting may now do that job with less machinery and fewer behavioral seams. Measure before you rip anything out, but measure.

Separate the latency decision from the intelligence decision. These used to be coupled through model choice. They are now two independent parameters. Interactive workloads and batch workloads should almost certainly be configured differently.

Restructure staged agents around the tool-change beta. If you have been stuffing every possible tool into a single definition block to protect the cache, that constraint is loosening. Smaller, phase-appropriate toolsets tend to produce better tool selection.

Do not skip your own evals. Frontier-Bench, CursorBench, ARC-AGI, OSWorld, and AutomationBench are all reasonable proxies, and none of them are your codebase. The cost-per-task framing in Anthropic’s numbers is a good template to copy for your own measurements — score alone has never been the decision.

The Pattern Underneath

For most of the current generation of models, “better” has meant “bigger, slower, more expensive, and worth it.” Opus 5 is a release where better meant the same price, more efficiently spent — beating competitors at a third to a half of their cost per task, and doubling its predecessor without charging more for the privilege.

That is a healthier direction than a leaderboard climb. A model that is 5% smarter and 3× the price changes what is possible in a demo. A model that is substantially smarter at the same price changes what is affordable in production, and production is where the interesting problems have been waiting.

---

Sources: Anthropic — Introducing Claude Opus 5, SWE-bench Verified Leaderboard, Claude Opus 5 Benchmarks Explained — Vellum