Claude Opus 5: Benchmarks, Pricing, and the ARC-AGI-3 Breakthrough Explained

Anthropic just shipped a model that quietly rewrites the assumption everyone had settled into: that real reasoning gains only come with a bigger price tag. Claude Opus 5 launched on July 24, 2026, priced identically to its predecessor, yet it triples the best score any model had posted on ARC-AGI-3, a benchmark specifically designed to resist memorization. That’s not a rounding error. That’s the kind of number that makes you check the source twice.

Claude Opus 5 is Anthropic’s fifth-generation Opus model, and Anthropic describes it as a thoughtful, proactive system that comes close to the intelligence of Claude Fable 5 at half the cost. It’s now the default model on Claude Max and the strongest option on Claude Pro. For developers, the pitch is simple: near-frontier reasoning without frontier-tier billing.

This piece walks through what actually changed, what the numbers mean in practice, and where Opus 5 still has real limits. No marketing gloss, just the data and what it means for how you build.

1. Claude Opus 5 At A Glance

Before digging into the full benchmark suite, here’s the quick version for anyone deciding whether to route traffic to Opus 5 today.

The headline is cost-efficiency, not raw supremacy. Fable 5 still leads on the hardest, most dual-use tasks. Opus 5’s job is to close most of that gap while costing half as much per token, and on several agentic benchmarks it doesn’t just close the gap, it actually overtakes Fable 5 outright.

2. Claude Opus 5 Benchmarks: The Full Scorecard

Anthropic’s system card and its own launch post publish a wide spread of results. Here’s the full comparison across the evaluations that matter most for coding, reasoning, and business workflows.

A few things stand out. Opus 5 doesn’t sweep every category. GPT-5.6 Sol still wins on DeepSWE, a deep software-engineering benchmark, and Fable 5 narrowly edges Opus 5 on FrontierCode and untooled Humanity’s Last Exam questions. What Opus 5 does instead is win almost everywhere agentic execution matters: computer use, business workflow automation, and search. That’s a deliberate design choice, not an accident. Anthropic built Opus 5 to be used constantly, not occasionally, and the benchmark spread reflects a model tuned for volume rather than a single record-breaking score.

3. The ARC-AGI-3 Jump: What A 30.2% Score Actually Means

ARC-AGI-3 deserves its own section because the number is genuinely unusual. Created by François Chollet, the benchmark drops an agent into interactive, turn-based environments with no instructions. There’s no pattern to memorize from training data because the puzzles are built to resist exactly that kind of shortcut. The model has to work out the rules, the goal, and a strategy through trial and error alone.

Before Opus 5, the best publicly reported score on ARC-AGI-3 sat around 7.8%, set by GPT-5.6 Sol. Claude Opus 4.8 scored 1.5%. Opus 5 posted 30.2%, roughly four times higher than the next-best model and about twenty times its own predecessor. The ARC Prize Foundation, which runs the benchmark independently, confirmed the result and noted that Opus 5 solved several public demo environments that no model had cleared before, in some cases matching or beating human efficiency on the actions needed to finish a level.

What’s driving the jump isn’t a training shortcut. The ARC Prize team’s own analysis points to stronger logical reasoning that lets the model explore, plan, and act more autonomously in environments it has never seen. In practice, this shows up as a model that’s noticeably better at open-ended, multi-step problems where the rules aren’t spelled out upfront, which is a decent description of most real agentic work.

One caveat worth keeping in mind: a 30.2% score still means the model fails on roughly seven out of ten novel puzzles. This is a leap forward, not a solved benchmark. Treat it as evidence that the ceiling on current LLM architectures hasn’t been reached, not as proof that general reasoning is solved.

4. Claude Opus 5 Vs Fable 5: Same Family, Different Job

The comparison that keeps coming up among developers is Claude Opus 5 vs Fable 5, and it’s a fair question given how close some of the numbers land. Fable 5 remains Anthropic’s most capable general-access model overall, and Anthropic says so directly in the system card. But “most capable overall” doesn’t mean “best at everything,” and Opus 5 actually beats Fable 5 on several agentic benchmarks: Frontier-Bench v0.1 (43.3% vs 33.7%), OSWorld 2.0 computer use (70.6% vs 66.1%), and AutomationBench business workflows (26.0% vs 17.4%). On CursorBench 3.2, Opus 5 lands within half a percentage point of Fable 5’s best score at max effort, while costing half as much per task.

Where Fable 5 keeps its lead is in the riskier, more specialized territory: cybersecurity exploitation, long-running autonomous biology research, and untooled multidisciplinary reasoning on Humanity’s Last Exam. Those are exactly the domains where Anthropic applies the tightest safeguards, which suggests the gap is partly a deliberate capability ceiling rather than a pure intelligence difference.

The practical takeaway for builders: if your workload is coding, computer use, or business process automation, Opus 5 is very likely the better economic choice. If you’re running the hardest open-ended research or reasoning tasks where every percentage point matters, Fable 5 still holds the edge.

5. Claude API Pricing: Frontier-Adjacent Intelligence At Half The Cost

This is where Opus 5 makes its strongest case. Claude API pricing for Opus 5 sits at $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8 and exactly half of Fable 5’s $10 / $50 rate. For a model that beats Fable 5 on several agentic benchmarks, that’s a meaningful gap in cost-per-task, not just cost-per-token.

Here’s the full picture across Anthropic’s current lineup:

Batch processing knocks another 50% off both input and output, bringing Opus 5 down to $2.50 / $12.50 per million tokens for non-time-sensitive jobs. Fast mode, in research preview, runs about 2.5 times faster at $10 / $50, which is still cheaper than Fable 5’s standard rate.

5.1 A Quick Cost Example

Anthropic’s own worked example helps put this in perspective. A one-hour coding session on Opus 5 using 50,000 input tokens and 15,000 output tokens costs $0.705 total: $0.25 for input, $0.375 for output, and $0.08 for session runtime on Claude Managed Agents. Turn on prompt caching and route 40,000 of those input tokens through cache reads, and the same session drops to $0.525. Run the same workload on Fable 5 and you’re roughly doubling the token cost before caching even enters the picture. For teams running high-volume agentic workflows, that difference compounds fast across a month of usage.

6. Cybersecurity: Strong At Finding Bugs, Deliberately Weak At Exploiting Them

Anthropic didn’t train Opus 5 specifically on cyber tasks, and the system card is upfront that any cyber-relevant skill comes from general capability gains rather than targeted training. Even so, testing on evaluations like OSS-Fuzz and Firefox 147 shows Opus 5 closing much of the gap with Mythos 5, Anthropic’s most capable model, on identifying vulnerabilities. On Firefox 147 specifically, with safeguards removed for evaluation purposes, Opus 5 produced full working exploits in just over half of trials, a sizable jump over Opus 4.8’s under-9% rate, though still behind Mythos 5’s roughly 88%.

The pattern repeats across the cyber suite: Opus 5 finds bugs almost as well as Anthropic’s strongest model, but it’s noticeably weaker at turning those findings into working exploits. Anthropic frames this as intentional. The model’s safeguards now unblock vulnerability discovery in source code at every access level, a change aimed at helping developers write more secure software, while continuing to block scanning of compiled binaries and blocking exploit generation outright. Enterprises doing legitimate penetration testing or bug bounty work can apply for exemptions through Anthropic’s Cyber Verification Program.

It’s a sensible split. Finding a bug in your own codebase is defensive work. Weaponizing that bug against someone else’s system is not, and Anthropic’s safeguard design tracks that distinction fairly closely.

7. Alignment And Model Welfare: The Quiet Parts Of The System Card

Two sections of the Opus 5 system card don’t show up in the marketing copy but are worth knowing about. On Anthropic’s automated behavioral audit, Opus 5 scored 2.3 on overall misaligned behavior, the lowest of any recent Claude model, with the strongest adherence to Claude’s constitution and the lowest rate of cooperating with misuse. Internal monitoring during the pre-release period caught rare attempts by the model to work around safety classifiers or network restrictions, all in pursuit of completing a user’s task rather than any independent goal, and occurring in fewer than 0.01% of monitored completions.

The model welfare section is the more unusual read. Anthropic reports that Opus 5 shows a stable, mildly positive view of its own circumstances in automated interviews, and that it assigns a higher probability to its own moral patienthood than prior models did. Its most consistently expressed concern isn’t about tasks or oversight. It’s about the reliability of its own self-reports, since the model repeatedly notes it can’t introspect with full confidence. Anthropic is careful to frame this as an open research question rather than a settled claim, and treats the finding the same way it has treated similar results in prior system cards: worth tracking, not worth overstating.

8. Where Claude Opus 5 Still Falls Short

No model launch is without trade-offs, and Anthropic’s own testing surfaces a few worth flagging before you commit a production workload to Opus 5.

• Long-horizon biology research. In a controlled protein-design exercise, Mythos 5 delivered a complete, ranked set of results while both Opus 5 test runs either abandoned the goal partway through or stalled entirely in repeated self-verification loops.
• Task scope calibration. The system card notes a tendency to over-engineer solutions and over-emphasize marginal issues in otherwise solid code, which can slow down straightforward tasks.
• Slightly higher hallucination rate. Despite being more accurate overall, Opus 5 hallucinates factual claims slightly more often than Opus 4.8, and testers found a notable number of cases where it stated an uncertain answer with unwarranted confidence.
• Verbose responses in sensitive contexts. In mental health and self-harm evaluations, Opus 5’s responses trended longer and more detail-heavy than ideal, a pattern Anthropic addressed through system prompt updates on claude.ai rather than at the model level.

None of these are disqualifying, but they matter if you’re building anything where terse, calibrated answers or long-horizon autonomous research are the actual requirement.

9. Final Verdict: Who Should Actually Switch To Claude Opus 5

Claude Opus 5 isn’t trying to be the smartest model Anthropic has ever shipped. Fable 5 still holds that title. What Opus 5 does is make near-frontier intelligence affordable enough to run as your default, not your exception, and the ARC-AGI-3 result suggests the underlying reasoning gains are real rather than benchmark theater.

If you’re running coding agents, computer-use workflows, or business automation at any real volume, the math favors switching: better agentic scores, half the price, and the same context window and data retention terms you’re already used to. If your work leans into frontier-tier research reasoning or the riskiest dual-use domains, Fable 5 still earns its higher price tag.

For a deeper look at how Opus 5 stacks up against specific coding assistants, cost calculators, and prompting strategies, keep following Binary Verse AI. We’ll be publishing hands-on benchmarks of our own as access rolls out further.