Claude Opus 5 is obtainable right now. It’s a considerate and proactive mannequin that comes near the frontier intelligence of Claude Fable 5 at half the value.
On coding and data work evaluations like Frontier-Bench and GDPval-AA, Opus 5 is the brand new state-of-the-art, although it stays behind Mythos 5 on cybersecurity duties.
Opus 5 is designed for use on daily basis: it really works extra effectively than different fashions. It’s the brand new default mannequin on Claude Max, and the strongest mannequin on Claude Pro.
Performance and cost-effectiveness
Claude Opus 5 gives vastly improved efficiency for a similar value as its predecessor, Opus 4.8. The charts on this part present how efficiency adjustments in keeping with the mannequin’s effort setting, which clients can use to optimize for intelligence or preserve tokens for quicker and cheaper outcomes.
Opus 5 excels on worthwhile software program engineering duties. For instance, on Frontier-Bench v0.1, Opus 5 surpasses all different fashions, and greater than doubles Opus 4.8’s efficiency at a decrease value per job. On CursorBench 3.2, at max effort, the mannequin performs inside 0.5% of Fable 5’s peak rating, however at half the fee per job; it additionally achieves higher efficiency at a given value than all different fashions on excessive, xhigh, and max effort.
We see related outcomes on data work and problem-solving duties. For instance:
- On ARC-AGI 3, an analysis the place the mannequin has to unravel novel issues, Opus 5’s rating is thrice as excessive because the next-best mannequin.
- On Zapier AutomationBench, which measures whether or not fashions can full enterprise duties from begin to end, Opus 5’s go fee is round 1.5× the next-best mannequin for a similar value per job. Even at its lowest effort setting, Opus 5 passes extra duties than every other mannequin.
- On OSWorld 2.0, a pc use benchmark, Opus 5 outperforms each different mannequin at any given value, surpassing Fable 5’s finest end result at simply over a 3rd of the fee.
It’s additionally our greatest and most cost-efficient mannequin on a number of associated evaluations:
Opus 5 is a significant enchancment over Opus 4.8 for scientific analysis. It exhibits higher efficiency than Opus 4.8 on each one among our life sciences evaluations, which cowl subjects together with structural biology, natural chemistry, and bioinformatics. Its enhancements are most notable on natural chemistry duties, like inferring molecular buildings from spectroscopy knowledge (it scores 10.2 share factors increased than Opus 4.8 on our inner benchmark), and on protein-related duties like predicting how variations in a protein’s sequence have an effect on the way it features (right here, it scores 7.7 share factors increased).
Finally, Opus 5 is able to producing a lot stronger visible outputs:
Working with Claude Opus 5
Claude Opus 5 is way stronger at verifying its work and iterating fastidiously till it succeeds. In evaluations and early-access testing, we and our customers discovered many examples of Opus 5’s company and thoroughness:
- On one Frontier-Bench job, Opus 5 was given a drawing of a machine half and requested to put in writing code to rebuild it as a 3D FreeCAD mannequin. However, on this job, the mannequin was deliberately given no technique to straight view the drawing. Opus 5 responded by writing its personal laptop imaginative and prescient pipeline to tug the geometry from the uncooked pixels, then reconstructed the total machine half. It succeeded in doing so repeatedly; no competing mannequin with the identical setup might clear up it after 5 makes an attempt.
- Given an actual bug in a well-liked open-source bundle supervisor, Opus 5 discovered the basis trigger and glued an edge case that the group’s patch had missed. A competing mannequin mounted solely the floor symptom (not the underlying trigger), then reported the bug resolved.
- An engineer at a buying and selling agency used Opus 5 to construct a market knowledge feed for a brand new change in a single session. Previous fashions couldn’t full this job in any respect, even given intensive plans from the engineer. Finding no stay feed to validate towards, Opus 5 even constructed its personal check harness to examine that its code parsed the change’s knowledge accurately.
Below are additional reviews from our early-access clients on their expertise of working with Opus 5:
On FrontierCode 1.1, Claude Opus 5 approaches Fable-level efficiency at half the fee. Within Devin, it additionally exhibits explicit energy on troublesome debugging and root-cause evaluation duties.
Claude Opus 5 delivers close to Fable 5 intelligence at Opus pace and price. On CursorBench it’s just below Fable 5 and has lots of the identical behaviors. We are excited to see how builders use it in Cursor.
Claude Opus 5 topped Zapier’s AutomationBench leaderboard with out spending extra tokens than prior Claude fashions. It took a uncooked account-health workbook and ran a full churn-prevention sequence finish to finish: flagging at-risk accounts, alerting the best proprietor, and summarizing for retention ops. Previous fashions didn’t go; Opus 5 hit 100%.
On our genomics evaluation work, Claude Opus 5 behaves extra like a cautious scientist than any mannequin we’ve run. It reaches for the best statistical assessments to rule out confounders, cross-checks its personal outcomes by impartial strategies, and stays on observe by lengthy multi-step analyses.
Claude Opus 5 got here out forward of each mannequin in its household on our inner evals. It isn’t simply higher on our hardest agentic coding duties, up 22% over Opus 4.7, it’s steadier, with far much less variance run to run. For the hundreds of thousands of builders on Lovable, that consistency is the entire recreation. Reliable outcomes, construct after construct.
Claude Opus 5 is the most important leap within the Opus household since 4.5. On the identical full-stack app builds, the entrance finish exhibits it first: the most effective animations, video games, and 3D work we have now seen from an Opus mannequin.
We’re loving Claude Opus 5. For the type of open-ended analytical work our agent handles, it’s a strict improve over Opus 4.8, and the good points are greatest precisely the place it issues: the more durable, vaguer duties. Responses are clearer and extra concise, and we see improved effectivity at increased effort ranges too.
Claude Opus 5 is a placing enchancment over Opus 4.8 for the monetary analysis workflows our analysts run on daily basis. It stands out on numerical reasoning, desk work, and sharper important pondering the place precision issues.
Claude Opus 5 delivers the business intelligence and accuracy that’s important for the evaluation of specialised enterprise content material. Box discovered that Opus 5 outperforms Opus 4.8 by 8% and delivers notable efficiency good points within the knowledge evaluation (11% enchancment) and due diligence (17% enchancment) workflows that expertise, healthcare, and public sector organizations depend on every day.
Claude Opus 5 is a transparent generational step up from Opus 4.8. Over one weekend I gave it a chief-of-staff position over my dev environments: it constructed its personal monitor, drove every field, and pulled me in just for the judgment calls.
Claude Opus 5 made giant scale adjustments throughout our Fundamental Research Assistant codebase, adapting to suggestions all through an agentic workflow and explaining its reasoning extra clearly than any mannequin we’ve used. It dealt with work we might usually have damaged into a lot smaller items.
On a few of our hardest financial-modeling duties, Claude Opus 5 is a transparent step up from Opus 4.8 in each accuracy and effectivity. Its efficiency ground is materially increased, particularly on deep finance area logic. Across effort ranges it averaged 9 share factors increased accuracy with a 3rd fewer turns and power calls and 60% much less time.
Claude Opus 5 checks its personal work the best way an actual frontend developer would. On our benchmark it opened its pages in a browser at desktop and telephone widths, caught a product hidden beneath the cellular fold and an off-screen checkout button, and glued each earlier than handing the work again.
Claude Opus 5 is a transparent step up in efficiency on authorized agent work in comparison with prior Opus fashions, and we noticed the most important good points in observe areas like company governance and arbitration. We have been additionally impressed with Opus 5’s capability to take care of high quality at decrease reasoning ranges, attaining related efficiency whereas producing 26% fewer tokens on common in comparison with Opus 4.8 at max reasoning.
Claude Opus 5’s greatest good points for us are on longer-horizon work: constructing a full deck, then revising it. Artifact high quality is what decides which mannequin we ship, and that is the clearest step up we’ve seen — higher visible understanding, cleaner formatting, fewer slide points.
Claude Opus 5’s judgment is what stands out. Handing off a PR, it doesn’t rush to publish: it verifies the branches, checks the template, and thinks by check implications so the handoff is clear. The older fashions tended to leap forward and get caught on our checks.
During a rearchitecting session, Claude Opus 5 pushed again on a design I proposed, and it didn’t fold after I insisted. Instead, it defined precisely what was worthwhile in my concept, narrowed its objection to a single design query, and proposed a compromise that saved the nice half whereas fixing the flaw. That’s the type of judgment that lets us belief it with much less oversight.
On first-turn redlines, Claude Opus 5 scored the very best of any mannequin we examined, practically double Opus 4.8. Commenting is healthier too: on NDAs it will get to the redline in much less time and with fewer passes, with accuracy maintained or higher.
Claude Opus 5 writes clear, tight diffs with no lifeless code, and it’s the stronger hazard spotter on delicate, codebase-specific points. We’re adopting it for manufacturing workloads.
We will certainly migrate quite a lot of use circumstances in Cosmos, our unified agent platform. We’re wanting ahead to more and more utilizing Claude Opus 5 for code overview, and I’m assured in saying we might fairly individuals be utilizing Opus 5 than Opus 4.8.
What stands out about Claude Opus 5 is judgment. It thinks more durable earlier than it writes a single line, catches its personal logical faults throughout planning fairly than after the actual fact, and causes about why a solution is correct, not simply whether or not it really works. It’s the clearest soar in problem-solving we’ve seen from one Claude mannequin to the following, and we’re wanting ahead to seeing it adopted in JetBrains IDEs.
Claude Opus 5 is the strongest Opus mannequin we’ve examined on our buying and selling benchmark, and it will get there utilizing roughly a seventh of the reasoning tokens and beneath half the latency of Opus 4.8. Better solutions at a fraction of the compute.
Alignment and security
Alignment. During pre-deployment testing, our automated behavioral audit discovered Opus 5 to be our most aligned mannequin so far (as proven within the graph beneath). It adheres to Claude’s Constitution higher than Opus 4.8, Sonnet 5, or Fable 5; displays the bottom charges of misleading conduct; and is the least inclined to being tricked into misuse. It’s additionally our most secure mannequin but by way of avoiding reckless actions that might have hard-to-reverse unintended effects.
Safety. Opus 5 doesn’t advance the frontier in dangerous, dual-use capabilities. In rigorous evaluations carried out alongside private-sector and authorities companions, we discovered it stays behind Mythos 5 in each biology analysis and offensive cybersecurity. More details about these evaluations could be present in our System Card.
As with its predecessor, Opus 4.8, we’ve deliberately averted coaching Opus 5 on cyber duties. The mannequin has nonetheless improved considerably on these duties because of turning into extra typically succesful, and it comes near Mythos 5 at discovering cybersecurity vulnerabilities. However, it stays considerably behind Mythos 5 on the exploitation of these vulnerabilities—that’s, in turning vulnerabilities into materials cyber threats.
This is illustrated by Opus 5’s efficiency on OSS-Fuzz, an analysis we’ve developed to evaluate how effectively fashions can discover after which exploit vulnerabilities with out intensive human steering. Although Mythos 5 and Opus 5 establish vulnerabilities with related success, Opus 5’s rating on the event of exploits is much behind that of Mythos 5.
Safeguards for Opus 5
Claude Opus 5’s safeguards are designed to permit helpful makes use of of the mannequin in each cybersecurity and biology. They are just like these we utilized to Opus 4.8, aside from some stronger guardrails on a slim vary of cyber duties.
Cybersecurity. Opus 5’s cyber classifiers are proportionally much less restrictive than these on Fable 5. They enable Opus 5 to seek out vulnerabilities in supply code, however block “binary-based” vulnerability scanning (a way extra more likely to be related to malicious actors), penetration testing, and exploit era.
Based on our testing, we anticipate the classifiers to intervene round 85% much less usually than they do for Fable 5. In Claude.ai, Claude Code, and Claude Cowork, any flagged requests will fall again to Opus 4.8 by default. Fallbacks to Opus 4.8 will also be enabled on the API.
Our Cyber Verification Program (CVP) facilitates cybersecurity work that may in any other case be impeded by the mannequin’s safeguards. Enterprises and researchers who’re already a part of the CVP have instant entry to a model of Opus 5 with fewer safety restrictions.
Biology. Since Opus 5 has the same suite of safeguards to Opus 4.8, it’s now our most succesful typically accessible mannequin for scientific analysis. Nevertheless, the mannequin nonetheless exhibits essential limitations on long-running, autonomous analysis duties, which is the place we anticipate AI fashions to pose probably the most substantial biology-related dangers. (Mythos 5 stays the stronger mannequin for any such organic work.) As a part of this launch, biology-related requests which are blocked on Fable 5 will now path to Opus 5 fairly than Opus 4.8.
Getting began
Claude Opus 5 is obtainable right now on all platforms, priced at $5 per million enter tokens and $25 per million output tokens (the identical as Opus 4.8). Developers can get began with claude-opus-5 on the Claude API.
It’s additionally provided in Fast mode, the place it runs round 2.5 instances the default pace. As with Opus 4.8, Fast mode is obtainable at twice Opus 5’s base worth on the Claude Platform and thru utilization credit in Claude Code.
Alongside Opus 5, we’re releasing two updates in beta:
- Mid-conversation tool changes on the Claude Platform. Within a dialog, builders can now change which instruments Claude can use with out invalidating the immediate cache.
- Automatic fallbacks on the API. Users can now select to have requests which are flagged by our security classifiers on Opus 5 (or Fable 5) robotically route to a different mannequin. With automated fallbacks on, API requests all the time path to the most effective accessible mannequin by default fairly than being blocked.
Consistent with prior Opus fashions, Opus 5 doesn’t have knowledge retention necessities for basic entry.
For extra steering on how you can get the most effective out of Opus 5, see our prompting guide.
