Claude Opus 5.2 Benchmarks: What Anthropic's New Flagship Can Actually Do
Claude Opus 5.2 tops the Valves index and Arc AGI 3, builds an OS from scratch, designs hardware, and formalizes Fermat's last theorem in 30 hours.
What is Claude Opus 5.2?
Claude Opus 5.2 is Anthropic’s new top-tier model for the hardest technical work, released alongside a sister model called Claude Mythos 5.2. It leads the Valves index, scores 97.7% on Arc AGI 3 with a provider harness, and set a record on Epoch AI’s capabilities index at 174. In demos, it built a working operating system from an empty folder using more than 1,000 coordinated agents, and a team of its agents formalized Fermat’s last theorem in roughly 30 hours.
TL;DR
- Opus 5.2 and Mythos 5.2 are the same underlying model, with Opus carrying stronger safety restrictions on cyber security and biology so it can be released broadly, while Mythos (loosened safeguards) is limited to vetted organizations.
- It got cheaper and faster at the same time: $8 per million input tokens and $40 per million output tokens, down from $10/$50 for Opus 5.1, with outputs up to 30% faster and an optional fast mode up to 2.5x quicker.
- On the Valves index it ranks number one at 71.31% accuracy, beating Claude Sonnet 5.5, Gemini 4 Carbon, Opus 5.5, and GPT-6 Astra.
- On Arc AGI benchmarks it hit 97.7% on Arc AGI 3 (provider harness), 81.2% (standard harness), and 93.8% on Arc AGI 2 for under a dollar per task.
- Anthropic’s own engineers used it to build a prototype operating system that boots on real hardware, coordinating over 1,000 sub-agents across 8 days of continuous work.
- It designed three working electronic devices from a single paragraph of prompting, with circuit boards ready for manufacturing in under a day and no post-build electrical fixes needed.
- Anthropic says the model helped solve two open problems in theoretical physics related to black holes and holographic duality, with results written up as preprints for peer review.
Nothing gets lost.
How does Opus 5.2 compare to Mythos 5.2?
Opus and Mythos 5.2 share the same base model. The difference is in what gets filtered out before a response reaches the user. Opus ships with strong safeguards around cyber security and biology, which is what lets Anthropic make it available to the general public. Mythos strips back some of those restrictions and is only accessible to vetted organizations through Anthropic’s life sciences and cyber verification programs.
This mirrors how Anthropic rolled out the Mythos line in the first place. Back in April, an early Mythos preview was considered so capable in cyber security contexts that Anthropic didn’t release it publicly at all. By June, a version with safeguards attached shipped as the first publicly usable model in that capability class, called Opus 5. Opus 5.1 followed a few months later, and Opus 5.2 now sits a clear step above both. The pattern tells you Anthropic treats the underlying model and the safety layer as separate decisions, tuning who gets access to which version of the same brain.
What do the benchmark numbers actually show?
The headline number comes from the Valves index, a composite benchmark spanning coding, finance, and science, where Opus 5.2 posted 71.31% accuracy and claimed the top spot across seven benchmark categories, ahead of Sonnet 5.5, Gemini 4 Carbon, Opus 5.5, and GPT-6 Astra.
Arc AGI, a benchmark specifically designed to test reasoning that’s easy for humans and historically hard for AI, is where the model posted its most eye-catching scores: 97.7% on Arc AGI 3 with a provider harness, 81.2% with the standard harness (a new high), and 93.8% on Arc AGI 2 while costing under a dollar per task.
Epoch AI’s capabilities index, which tracks frontier model progress over time, put Opus 5.2 at 174, topping the previous record of 169 set by GPT-6 Astra. Epoch also noted the score sits above what the existing trend line for reasoning models would have predicted, meaning the jump wasn’t just incremental progress along the existing curve.
On Anthropic’s own benchmark table: 69.9% on Frontier SWE (agentic coding), 89.3% on OS World (computer use), and 74.2% on Humanity’s Last Exam with tools. Visual chart recognition jumped from 44.8% on Opus 5.1 to 73.4% on 5.2. On Code Arena’s webdev leaderboard, where humans vote on generated websites, it ranks first, ahead of Opus 5.5.
It isn’t a clean sweep everywhere. On Harvey’s legal benchmark, Opus 5.2 scored lower than both Opus 5.1 and Sonnet, and a separate evaluation flagged weaker performance on accounting judgment tasks. For a launch with this many records, those are notable but isolated gaps.
Can it really build an operating system from scratch?
According to Anthropic, yes, in a controlled internal test. An engineer gave Opus 5.2 the task of designing a prototype operating system starting from an empty folder. The model coordinated more than 1,000 sub-agents working continuously for 8 days, writing, testing, and reviewing code along the way. The result booted on a real laptop, with its own kernel, drivers for disk, keyboard, and network, persistent file storage, a windowing system, a login screen, an installer, and Claude built directly into it.
Anthropic also says over 90% of changes merged into its own internal codebase are now written by Claude with human guidance, and a team of agents running an earlier Opus 5.2 build migrated a production service from Python to Rust, with the new service starting about 26 times faster. External testers reported similar results: Uber had the model run a 10 to 12 hour security audit using roughly 170 reviewer agents and return a full write-up with fixes, and one company said it built a hypervisor in a single session backed by more than 3,500 machine-checked proofs.
What are early enterprise testers saying?
Anthropic gave pre-launch access to several companies, and the feedback leans heavily positive. Shopify’s Tom Pritchard described it as handling tasks with little intervention, comparing it to “working with another staff-level engineer.” Data Dog said it could handle unfamiliar, ambiguous systems like an incident commander directing faster models to gather information. Oracle tested it against enterprise SQL tasks combining debt summaries and repayment history; Opus 5.2 got the right answer while two other models produced SQL that ran but returned numbers five to eight times too large.
Cognition, the company behind Devin, flagged something specific: the model says it doesn’t know something rather than guessing, including one instance where it told testers its last estimate was wrong and it wouldn’t give a new number until it had actually measured the change. That kind of calibrated uncertainty matters more than raw benchmark scores for anyone trying to use a model for real decision-making.
The publication Every, which runs its own frontier model tests, rated it as top tier for coding work, with one team member building a fully featured video editor through prompting alone, and another noting it’s the first model to proactively give a completion time estimate for long-running tasks rather than leaving the user to guess.
Is it capable of real scientific and hardware work?
Anthropic claims Opus 5.2 contributed to solving two open problems in theoretical physics, both tied to holographic duality in black hole physics. One involved finishing a calculation first outlined in 2003 that had sat incomplete for over two decades because it was too labor-intensive; the model completed it largely on its own. Both results were published as preprints, which means they’re open to scrutiny from the physics community rather than taken on faith.
On the hardware side, Anthropic gave the model a single paragraph of instructions and had it design three electronic devices: a voice assistant hub, a USB speakerphone, and a handheld virtual pet. It wrote specifications, selected real components from distributors, routed four-layer circuit boards, modeled enclosures, and wrote the firmware. Boards were manufacturing-ready in under a day, and all three devices reportedly worked without any post-build electrical fixes.
Without shadow IT.
A team of Opus 5.2 agents also reportedly formalized the proof of Fermat’s Last Theorem, a famously difficult mathematical proof, in about 30 hours, translating it into a form that can be verified line by line by a proof checker.
Frequently Asked Questions
What does Claude Opus 5.2 cost compared to Opus 5.1?
Opus 5.2 costs $8 per million input tokens and $40 per million output tokens, down from $10 and $50 for Opus 5.1. It also generates outputs up to 30% faster, with an optional fast mode offering up to 2.5x the speed.
Is Claude Opus 5.2 the same as Claude Mythos 5.2?
They share the same underlying model. Opus 5.2 has stronger safeguards around cyber security and biology and is broadly available. Mythos 5.2 loosens those restrictions and is restricted to vetted organizations through Anthropic’s verification programs.
Should I use Opus 5.2 for everyday coding tasks?
Anthropic itself recommends starting with Opus 5.5, described as its daily driver, for routine coding and knowledge work, and reserving Opus 5.2 for the most ambitious or technically demanding jobs given its role as an “advisor” or “orchestrator” model.
What is the Arc AGI benchmark testing?
Arc AGI is designed to measure reasoning abilities that come naturally to humans but have historically been difficult for AI systems, as opposed to pattern matching on familiar data. Opus 5.2 scored 97.7% on Arc AGI 3 with a provider harness.
Did Opus 5.2 actually solve real unsolved problems?
Anthropic reports that the model contributed to solving two open theoretical physics problems related to black hole physics, with results submitted as preprints for independent verification. It’s also credited with formalizing a complex mathematical proof and completing agentic engineering tasks like OS prototyping and hardware design.
