Claude Opus 5 Benchmarks Explained
Claude Opus 5 Benchmarks Explained
Anthropic released Claude Opus 5 on July 24, 2026. The pitch is unusually direct: near Fable 5 intelligence at half the price. $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8, and half of Fable 5's $10/$50 rate card.
Here's the thing, though. On several of the benchmarks Anthropic published, Opus 5 surpasses Fable 5 outright. And on one test designed to resist every trick labs use to inflate their numbers, it posts a result nobody in the field expected.
1. Coding: Frontier-Bench v0.1
Frontier-Bench v0.1 is an agentic terminal coding benchmark. Tasks involve real software engineering work: multi-file changes, debugging, building features from specifications. It's the benchmark that most directly maps to what enterprise development teams actually pay for.
Frontier-Bench v0.1
| Score, % (higher is better) | Opus 5 | Fable 5 | Opus 4.8 |
|---|---|---|---|
| 43.3 | 33.7 | 18.7 |
Source: Anthropic, Introducing Claude Opus 5. Scores via VentureBeat.
Opus 5 scores 43.3%, more than double Opus 4.8's 18.7% and nearly 10 points clear of Fable 5's 33.7%. This is the result that reframes the whole launch. The pre-launch consensus on X was that Opus 5 would land "comparable to Fable 5, but won't surpass it." Pankaj Kumar tweeted exactly that the day before launch. It was wrong. Opus 5 clears Fable 5 by a wide margin on coding while costing half as much per task.
On CursorBench 3.2, the picture is similar but tighter. At max effort, Opus 5 lands within 0.5% of Fable 5's peak score, but at half the cost per task. Anthropic also claims it achieves greater performance at a given cost than every other model on high, xhigh, and max effort settings. Scott Wu at Cognition, the company behind Devin, confirmed on FrontierCode 1.1 that "Claude Opus 5 approaches Fable-level performance at half the cost," with particular strength in debugging and root-cause analysis.
2. Novel problem-solving: ARC-AGI 3
ARC-AGI 3
| Relative Human Action Efficiency, % (higher is better) | Opus 5 | GPT-5.6 Sol | Opus 4.8 |
|---|---|---|---|
| 30.2 | 7.8 | 1.5 |
Source: ARC Prize Foundation, verified scores. Via OfficeChai and Anthropic.
ARC-AGI 3 is the benchmark François Chollet built to measure something closer to fluid intelligence than pattern-matching. Instead of static grid puzzles, it drops an AI agent into interactive, turn-based environments with no instructions, no stated rules, and no stated goal. The agent has to work out everything through trial and error. It's scored by Relative Human Action Efficiency, which compares how many actions the model needs against how many the second-best human took on the same level. Humans clear 100% of the environments.
When ARC Prize Foundation released ARC-AGI 3 in March 2026, the best AI model scored 0.37%. By launch day, the strongest reported score 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's roughly four times GPT-5.6 Sol's result, and twenty times Opus 4.8's. Anthropic's framing of "three times the next-best model" is, if anything, conservative.
A 30% score on ARC-AGI 3 means Opus 5 is genuinely working out unfamiliar rule systems through interaction, not recalling something adjacent to what it has already seen. 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. That's a much harder property to fake.
3. Knowledge work and computer use
On Zapier AutomationBench, which measures whether a model can complete business tasks from start to finish, Opus 5's pass rate is roughly 1.5x the next-best model for the same cost per task. Even at its lowest effort setting, it passes more tasks than any other model at any effort level. Wade Foster at Zapier confirmed that on a raw account-health workbook, Opus 5 "ran a full churn-prevention sequence end to end: flagging at-risk accounts, alerting the right owner, and summarizing for retention ops. Previous models didn't pass. Opus 5 hit 100%."
On OSWorld 2.0, a computer use benchmark, Opus 5 outperforms every other model at any given cost, surpassing Fable 5's best result at just over a third of the cost. On GDPval-AA, a knowledge work evaluation, Opus 5 is the new state-of-the-art.
4. Science: life sciences and visual outputs
Opus 5 is Anthropic's most capable generally available model for scientific research. It outperforms Opus 4.8 on every life sciences evaluation Anthropic tracks, with the biggest gains on organic chemistry (10.2 percentage points higher on their internal benchmark, including inferring molecular structures from spectroscopy data) and protein-related tasks (7.7 percentage points higher, including predicting how variations in a protein's sequence affect its function).
The model also produces substantially stronger visual outputs. Anthropic showed Opus 5 visualizing airflow over aerodynamic objects and building a simplified, interactive illustration of a cell. On X, early-access users were particularly struck by the 3D generation quality. sui (@birdabo) tested Opus 5 against Fable 5 Max and called it "quite the AGI beast in 3D, the texture and details are absolutely insane," while noting that Fable 5 one-shot the same task where Opus 5 needed multiple attempts.
5. It checks its own work
The benchmark numbers tell you what Opus 5 scores. The stories from the announcement 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. The model was intentionally given no way to directly view the drawing. Opus 5 responded by writing its own computer vision pipeline to pull the geometry from the raw pixels, then reconstructed the full machine part. It succeeded repeatedly. No competing model with the same setup could solve it after five attempts.
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 patch had missed. A competing model fixed only the surface symptom and reported the bug resolved.
An engineer at 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 at all, even with extensive plans. Finding no live feed to validate against, Opus 5 built its own test harness to check that its code parsed the exchange's data correctly.
This is the pattern across the customer testimonials: the behavioral shift is that 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."
6. Cybersecurity: close on finding, far on exploiting
Opus 5 closes much of the gap with Mythos 5 on finding software vulnerabilities, but the two models pull apart sharply on exploit development. On OSS-Fuzz, Anthropic's vulnerability-and-exploit benchmark, Opus 5 tracks close to Mythos 5 on discovery and falls well behind on exploitation. Mythos 5 can turn a found bug into a working attack. Opus 5 cannot.
This is by design. Anthropic intentionally avoided training Opus 5 on cyber tasks. The model improved on cybersecurity as a side effect of becoming more generally capable, but the safeguards are calibrated to allow defensive work (finding vulnerabilities in source code) while blocking offensive capabilities (binary-based scanning, penetration testing, exploit generation). Anthropic says Opus 5's classifiers will trigger roughly 85% less often than Fable 5's did. When they do trigger, flagged requests fall back to Opus 4.8 automatically.
7. The most aligned Opus yet
On Anthropic's automated behavioral audit, Opus 5 scores 2.30 on overall misaligned behavior. That's the lowest of any recent Claude model, ahead of Opus 4.8, Sonnet 5, and Fable 5. Anthropic frames this as the least deceptive behavior and the strongest adherence to Claude's Constitution of anything they have shipped.
The alignment story matters here because of the context. Fable 5's launch was rocky. Amazon researchers found ways to bypass its safeguards, which triggered a U.S. government export control directive and a two-week takedown in June 2026. Users complained the guardrails were too strict, making the model less useful for coding, cybersecurity, and biological research. Opus 5's alignment improvement, combined with the 85% reduction in classifier interventions, is Anthropic's answer to both critiques: safer and less in the way.
The effort dial
Opus 5 ships with an effort setting: low, medium, high, and max. Developers can toggle how much reasoning the model expends on a given task. Lower effort means fewer tokens, faster responses, lower cost. Higher effort means more thinking, better results, more spend.
This matters because the benchmark charts in Anthropic's announcement are all cost-versus-performance curves, not single-point scores. Opus 5's advantage isn't just that it scores higher at max effort. It's that at every effort level, it delivers more performance per dollar than every other model on the same benchmark. On the trading benchmark Anthropic referenced internally, Opus 5 reaches the 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.
The open question
The day before launch, Yamikishi (@Yamik1shi) posted a take on X that cuts to the heart of Opus 5's commercial positioning: "If Opus 5 drops today, it has one job. Not win another benchmark. Prove Anthropic can still charge a frontier premium." The context: Kimi K3 from Moonshot sits just above Opus 4.8 in Text Arena at $3/$15 per million tokens versus $5/$25 for Opus, and Moonshot plans to release full weights on July 27. GPT-5.6 Sol has been marketed heavily on cost per task. The field has gotten price-conscious.
Opus 5 doesn't undercut anyone on raw price. It's still above Sol and well above K3. Anthropic's argument is that performance per dollar, not price per token, is the metric that matters once you account for how many turns and how much reasoning each model actually needs to finish a task. The Frontier-Bench and OSWorld results back that argument. The ARC-AGI 3 result makes it hard to dismiss as marketing.
Takeaways
- Opus 5 beats Fable 5 on Frontier-Bench v0.1 (43.3% vs 33.7%) and ARC-AGI 3 (30.2% vs the next-best 7.8%) at half the price. The "near Fable 5" framing undersells it.
- ARC-AGI 3 is the standout: a 20x improvement over Opus 4.8 on a benchmark designed to resist memorization. This is the result most likely to shift expectations for the field.
- The effort dial (low/medium/high/max) turns benchmark wins into cost savings. Opus 5 leads on performance per dollar at every effort level, max included.
- Self-verification is the behavioral shift. Opus 5 checks its own work, catches its own errors, and builds its own test harnesses. The customer testimonials are unanimous on this point.
- Cybersecurity is deliberately split: close to Mythos 5 on finding vulnerabilities, far behind on exploiting them. Classifiers trigger 85% less than Fable 5.
- Most aligned model to date (2.30 misalignment score) with the strongest adherence to Claude's Constitution. The answer to both the Fable 5 safety crisis and the "guardrails too strict" complaints.
- Priced at $5/$25 per million tokens, same as Opus 4.8. Available today on Claude.ai, the Claude API, Claude Code, and Claude Cowork. Default model on Claude Max, strongest on Claude Pro.
If you're running an AI assistant on Opus 5, the model's self-verification and effort calibration map directly to the work your assistant does every day: checking its own outputs before showing them to you, spending more reasoning on harder tasks, and backing off when something is simple.