Claude Opus 5: The Model That Checks Its Own Work
Anthropic released Claude Opus 5 on July 24, 2026, and the pitch is unusually direct for a company that typically leads with safety framing: near-Fable 5 intelligence at half the price. The benchmark results, however, tell a more interesting story — on several key evaluations, Opus 5 doesn't just approach Fable 5. It surpasses it.
At $5 per million input tokens and $25 per million output tokens — identical to Opus 4.8 and half of Fable 5's $10/$50 — Opus 5 posts a 43.3% score on Frontier-Bench v0.1, an agentic terminal coding benchmark that measures real multi-file software engineering work. Fable 5 scores 33.7%. Opus 4.8 scored 18.7%. The model more than doubles its predecessor and beats the company's own frontier system by nearly 10 points on the benchmark that most directly maps to what enterprise developers actually do.
That alone would make this a significant launch. It is not the most interesting number in Anthropic's announcement.
The ARC-AGI 3 Result: 30.2%
ARC-AGI 3, designed by François Chollet, is not a standard benchmark. It drops an AI agent into interactive, turn-based environments with no instructions, no stated rules, and no stated goal. The agent must figure out everything through trial and error. It is scored by Relative Human Action Efficiency — how many actions the model needs compared to the second-best human. Humans clear 100% of the environments.
When ARC-AGI 3 was released in March 2026, the best AI model scored 0.37%. By launch day, the strongest reported result was GPT-5.6 Sol at max reasoning effort: 7.8%. Opus 4.8 barely registered at 1.5%.
Opus 5 scores 30.2%. That is roughly four times GPT-5.6 Sol's result and twenty times Opus 4.8's. The 30.2% score means Opus 5 is genuinely working out unfamiliar rule systems through interaction — not recalling something adjacent to what it has seen in training.
ARC-AGI 3 tests something closer to fluid intelligence than pattern matching. Benchmarks like SWE-Bench or GDPval reward a model for doing familiar categories of work faster and cheaper. ARC-AGI 3 rewards a model for handling something it has never been trained to handle at all. A jump from 7.8% to 30.2% in a single release is not an incremental improvement. It is a phase change.
The Effort Dial: Pricing That Adapts to the Task
Opus 5 ships with an effort setting: low, medium, high, and max. Developers can toggle how much reasoning the model expends per task. Lower effort means fewer tokens, faster responses, and lower cost. Higher effort means more thinking and better results.
This matters because the value proposition is not about headline benchmark scores at max effort. It is about performance per dollar at every effort level. On every benchmark Anthropic published, Opus 5 leads the cost-versus-performance curve — delivering more capability per dollar than any competing model at every effort setting.
On Frontier-Bench, Opus 5 at medium effort outperforms Fable 5 at any effort setting, including max. On OSWorld 2.0, a computer use benchmark, Opus 5 surpasses Fable 5's best result at just over a third of the cost. On an internal trading benchmark referenced by Anthropic, Opus 5 reaches its best score using roughly a seventh of the reasoning tokens and under half the latency of Opus 4.8.
The effort dial is the feature that turns benchmark wins into billing wins. A model that beats Fable 5 by 10 points at max effort is interesting. A model that beats every competitor at every effort level, at half the cost per task, is the one enterprises actually route traffic to.
It Checks Its Own Work
The benchmark numbers tell you what Opus 5 scores. The customer stories tell you how it works differently.
On one Frontier-Bench task, Opus 5 was given a drawing of a machine part and asked to write code to rebuild it as a 3D FreeCAD model — with no way to directly view the drawing. Opus 5 wrote its own computer vision pipeline to extract the geometry from the raw pixels, then reconstructed the full part. It succeeded repeatedly. No competing model could solve it after five attempts.
An engineer at a trading firm used Opus 5 to build a market data feed for a new exchange in a single session. Finding no live feed to validate against, Opus 5 built its own test harness to verify that its code parsed the exchange's data correctly. Previous models couldn't complete the task at all.
Given a real bug in a popular open-source package manager, Opus 5 found the root cause and fixed an edge case that the community's own patch had missed. A competing model fixed only the surface symptom and reported the bug resolved. The pattern across every customer testimonial is the same: Opus 5 verifies its outputs, catches its own errors, and iterates until it succeeds.
AJ Orbach at Tako put it directly: "Claude Opus 5 checks its own work the way a real frontend developer would. On our benchmark it opened its pages in a browser at desktop and phone widths, caught a product hidden below the mobile fold and an off-screen checkout button, and fixed both before handing the work back."
Safety and Alignment: Learning from Fable 5's Rocky Launch
Fable 5's launch was complicated. Amazon researchers found ways to bypass its safeguards, triggering a U.S. government export control directive and a two-week takedown in June 2026. Users simultaneously complained that the guardrails were too strict, making the model less useful for coding and research.
Opus 5 is Anthropic's answer to both problems. On the automated behavioral audit, it scores 2.30 on overall misaligned behavior — the lowest (best) of any recent Claude model, ahead of Opus 4.8, Sonnet 5, and Fable 5. Anthropic calls it the least deceptive model it has shipped, with the strongest adherence to Claude's constitutional framework.
At the same time, Opus 5's safety classifiers trigger roughly 85% less often than Fable 5's. When they do trigger, flagged requests fall back transparently to Opus 4.8 instead of being hard-blocked. The result is a model that is both safer and less intrusive — a difficult balance that Anthropic appears to have achieved through architectural improvements rather than heavier filtering.
On cybersecurity, Anthropic intentionally split the difference. Opus 5 tracks close to Mythos 5 on finding software vulnerabilities in source code, but falls well behind on exploiting them. The model improved on cybersecurity as a side effect of becoming more generally capable, not through targeted training. Discovery is permitted; offense is not.
Pricing and Availability
Claude Opus 5 is available immediately on all Anthropic platforms at $5/$25 per million tokens — the same as Opus 4.8, half of Fable 5. It is the new default model on Claude Max ($200/month) and the strongest model available on Claude Pro ($20/month). Developers can access it through the Claude API with the model string claude-opus-5, through Claude Code for terminal-based development, and through Claude Cowork for collaborative workflows.
Two beta features ship alongside it: mid-conversation tool changes that do not invalidate the prompt cache, and automatic API fallbacks that route flagged requests to a less capable model instead of blocking them entirely — small quality-of-life improvements that tell you Anthropic is thinking about production deployments, not just demos.
What It Means
Opus 5 resets the value proposition of frontier AI. Anthropic's core argument — that performance per dollar, not price per token, is the metric that matters once you account for how much reasoning each model needs to finish a task — is now backed by data on every published benchmark. The ARC-AGI 3 result, in particular, suggests that the field has been measuring the wrong thing. When a model jumps from 7.8% to 30.2% on a test designed to resist memorization, it is not just getting better at benchmarks. It is getting better at thinking.
For anyone building on top of these models, the takeaway is concrete: Opus 5 has the strongest cost-efficiency profile of any model at the frontier, a self-verification behavior that reduces the need for manual review, and an effort dial that lets you match spend to task complexity. The model that checks its own work also checks your budget.
Comments