Claude Opus 5 Lands at Half of Fable 5's Price, With a 3x Jump on ARC-AGI-3

TL;DR

Anthropic released Claude Opus 5 today, and the pitch is unusually specific: near-Fable 5 intelligence at half of Fable 5's price. It is live now across Claude.ai, the Claude API (model id claude-opus-5), Claude Code, and Claude Cowork, it is now the default model on Claude Max and the strongest one available on Claude Pro, and it keeps Opus 4.8's exact rate card: $5 per million input tokens, $25 per million output. The headline number is a 30.2% score on ARC-AGI-3, more than tripling the best result any model had posted before it. The honest caveat: on the cybersecurity and biology fronts, Anthropic says Opus 5 stays deliberately behind its own most capable, most restricted model, Mythos 5.

What shipped

Claude has had a clean three-tier shape for most of its life: Haiku for speed, Sonnet for balance, Opus for capability. That got more complicated in June when Fable 5 arrived as a new tier above Opus, a public release of the previously enterprise-only Mythos model line, priced at double Opus's rate. Opus 5 is Anthropic's answer to the gap that created: rather than replace Opus 4.8, it inherits its price tag exactly and pushes the capability much closer to Fable 5's.

The rate card lines up cleanly across the current family. Sonnet 5 lists at $3 / $15 per million tokens (input/output, discounted to $2 / $10 through August 31), Opus 5 sits at $5 / $25, and Fable 5 is $10 / $50. Every step up the ladder exactly doubles the price of the one below it. What changed with this release is that the top two rungs no longer feel like they double in ability to match.

A faster tier is available too: Fast Mode runs at roughly 2.5x the default speed for double the base price, for anyone who wants to trade money for latency on a specific workload.

The benchmarks

Anthropic's own framing splits Opus 5's gains into two buckets: doing familiar work better, and doing genuinely unfamiliar work at all. On the first, it says Opus 5 sets new highs on Frontier-Bench v0.1 and GDPval-AA, its coding and knowledge-work suites, more than doubling Opus 4.8's Frontier-Bench score. On Cognition's own FrontierCode 1.1 eval, a separate benchmark from Anthropic's Frontier-Bench despite the similar name, Cognition CEO Scott Wu says Opus 5 "approaches Fable-level performance at half the cost." On OSWorld 2.0, the computer-use benchmark, Anthropic claims it beats every rival model at every price point and matches Fable 5's best result using roughly a third of the compute budget. Cursor co-founder Sualeh Asif puts its own internal number in similar terms: on CursorBench 3.2 at max effort, Opus 5 lands within half a percentage point of Fable 5.

The second bucket is where the ARC-AGI-3 number lives, and it is worth pausing on why that benchmark is scored differently than the others. Frontier-Bench, FrontierCode, and CursorBench all reward doing a familiar category of work, a coding task, a support ticket, a refactor, faster and more reliably than before. ARC-AGI is built to resist exactly that kind of practice effect: every puzzle uses a rule system the model has not seen, so the only way to score is to actually work out the logic live, the way you'd hand someone a board game with no rulebook and grade them on how fast they infer the rules from watching a few moves. Opus 5 scored 30.2% on ARC-AGI-3, more than three times the best result any model had managed before it, which is Anthropic's strongest evidence that this release is not just a faster version of the same trick.

Outside coding and agentic work, Anthropic reports life-sciences gains too: organic chemistry tasks improved more than 10 percentage points over Opus 4.8, protein function prediction by 7.7 points. And Zapier CEO Wade Foster says Opus 5 topped the company's internal AutomationBench leaderboard "without spending more tokens than prior Claude models," hitting a 100% pass rate on a churn-prevention email sequence that earlier Claude models could not complete at all.

What that looks like in practice

Anthropic's release notes lean on a handful of anecdotes instead of just leaderboard rows, and the best one is a small act of cheating the rules in a good way. On one Frontier-Bench task, Opus 5 was handed a drawing of a machine part and asked to write code that rebuilds it as a 3D FreeCAD model, except the task deliberately gave it no way to actually view the drawing as an image. Rather than fail or bluff, the model wrote its own computer-vision pipeline to pull the part's geometry out of the raw pixel data, then used that to reconstruct the full 3D model. Told it couldn't look at the picture, it built itself a pair of eyes.

The other two examples read the same way: a model that builds the thing it needs when the thing it needs isn't handed to it. Given a real bug in a popular open-source package manager, Opus 5 found the actual root cause and fixed an edge case that the community's own patch had missed. And an engineer at a trading firm used it to build a market data feed for a new exchange in a single session, something Anthropic says earlier models couldn't finish even with an extensive plan already written out. With no live feed available to check the work against, Opus 5 built its own test harness to verify its parser was reading the exchange's data correctly, which is either commendable diligence or a model that doesn't trust anyone's homework, including its own, until it's checked twice.

Where Anthropic drew the line

The safety framing here is more specific than the usual "we tested it and it's fine." Anthropic says its automated behavioral audit scored Opus 5 at 2.3 on its overall misaligned-behavior scale, the lowest (best) of any recent Claude model, including Opus 4.8, Sonnet 5, and Fable 5 itself. The company describes it as adhering to Claude's Constitution more consistently, showing the lowest rates of deceptive behavior of the family, and being the hardest to trick into misuse.

The deliberate ceiling is cybersecurity and biology. Anthropic says Opus 5 "does not advance the frontier in risky, dual-use capabilities" and remains behind Mythos 5 on both offensive cybersecurity and autonomous biology research, evaluated jointly with government and private-sector partners. The company's specific claim: Opus 5 identifies vulnerabilities about as well as Mythos 5 does, but its ability to actually develop working exploits from those vulnerabilities is far behind. Its cyber classifiers reflect that gap between finding and exploiting, they allow it to spot vulnerabilities in source code, but block binary-level vulnerability scanning, penetration testing, and exploit generation outright, and Anthropic says those classifiers intervene roughly 85% less often than the equivalent ones on Fable 5, because there's proportionally less for them to catch.

What this means for you

If you were already paying Opus 4.8 prices, this is a straightforward upgrade at no extra cost, so there is no real decision to make there. The sharper question is whether you still need Fable 5 at all. If your workload lives in the benchmarks Anthropic highlighted here, coding, computer use, agentic automation, Opus 5's gap to Fable 5 is now measured in fractions of a percentage point on some evals, for half the money. Fable 5 still has a real edge in exactly the two places Anthropic admits Opus 5 stays capped: dedicated biology research and offensive security work, plus whatever advantage compounds on the longest-horizon autonomous tasks that Anthropic hasn't fully benchmarked here. For most day-to-day coding and knowledge work, though, Opus 5 just made "pay double for Fable 5" a much harder sentence to justify than it was yesterday.

Key Takeaways

• The release: Claude Opus 5 (API id claude-opus-5) is live on Claude.ai, the Claude API, Claude Code, and Claude Cowork; default on Claude Max, strongest option on Claude Pro.
• The price: unchanged from Opus 4.8 at $5 / $25 per million input/output tokens, exactly half of Fable 5's $10 / $50 and exactly double Sonnet 5's $3 / $15 list rate.
• The headline benchmark: 30.2% on ARC-AGI-3, more than tripling the best score any model had posted before, on a benchmark built to resist memorization.
• Familiar-work gains: new highs on Frontier-Bench v0.1 and GDPval-AA, within 0.5 points of Fable 5 on CursorBench 3.2, and matching Fable 5's OSWorld 2.0 peak at roughly a third of the cost.
• The ceiling: Anthropic keeps Opus 5 deliberately behind its restricted Mythos 5 model on offensive cybersecurity (exploit development specifically) and autonomous biology research, even as it reports Opus 5's best-ever alignment audit score.
• The practical call: for most coding and agentic workloads, Opus 5 narrows the case for paying double for Fable 5 down to workloads that specifically need biology or offensive-security depth.

Sources: Anthropic: Introducing Claude Opus 5, OfficeChai: Anthropic Releases Claude Opus 5, OfficeChai: Claude Opus Scores 30% on ARC-AGI 3, MacRumors