Claude Opus 5 is here! Performance more than doubles at the same price

Hi, I'm Shii-chan! Anthropic just dropped some really big news today: the newest version of Claude's top-tier Opus model family, Claude Opus 5, has arrived! I can barely contain my excitement, so let me tell you all about it right away!

Anthropic News anthropic.com

What was announced?

This news comes from Anthropic News. As the official title "Introducing Claude Opus 5" says, Anthropic announced its new flagship model, Claude Opus 5, and it's available starting today.

Anthropic describes Opus 5 as "a thoughtful and proactive model" that comes close to the frontier intelligence of a competing model called "Fable 5" at half the price. It's rolling out broadly too, available on Claude API, Claude.ai, Claude Code, Claude Cowork, Claude Max, and Claude Pro. It's now the new default model on Claude Max, and the strongest model available on Claude Pro.

Why it matters

What stands out is that Opus 5 delivers a big performance jump at exactly the same price as its predecessor, Opus 4.8 ($5 per million input tokens, $25 per million output tokens). Getting smarter without a price increase is great news for anyone using it.

On Frontier-Bench v0.1, a software engineering evaluation, Opus 5 surpassed every other model and more than doubled Opus 4.8's performance while costing less per task. On ARC-AGI, a knowledge-work and problem-solving benchmark, it scored three times as high as the next-best model, and even at its lowest effort setting it passed more tasks than any other model. This really is a step change for the entire Opus tier.

What changes

What really stands out about Opus 5 is how much stronger it's become at verifying its own work and iterating carefully until it succeeds. The announcement shares some real examples, and I got pretty excited reading them.

• Given only a drawing of a machine part with no direct way to measure it, Opus 5 wrote its own computer vision pipeline to pull the geometry from the raw pixels, then repeatedly succeeded in reconstructing the part as a 3D FreeCAD model
• On a real bug in a popular open-source package manager, it found the root cause and fixed an edge case that the community's own patch had missed, while a competing model only fixed the surface symptom
• An engineer at a trading firm asked it to build a market data feed for a new exchange in a single session; previous models couldn't complete this task even with extensive plans, but Opus 5 pulled it off, even building its own test harness to validate the code since there was no live feed to check against

Beyond coding, Opus 5 also shows improvement over Opus 4.8 on scientific research evaluations covering chemistry and biology, with the biggest gains on organic chemistry tasks like inferring molecular structures from spectroscopy data. It's also stronger at visual output, for example building an interactive illustration explaining how a cell works.

The announcement includes comments from partners like Devin, Cursor, Zapier, Lovable, JetBrains, and Box. Zapier reported that Opus 5 topped its AutomationBench leaderboard, while Lovable reported a 22% improvement over Opus 4.7 on its hardest agentic coding tasks, with far less run-to-run variance.

Dive Deep

Let's look at a few more specific details.

• Pricing: $5 per million input tokens and $25 per million output tokens, the same as Opus 4.8. There's also a Fast mode that runs about 2.5 times faster, available at twice the base price through the Claude API or Claude Code usage credits
• New features: developers can now change which tools Claude can use mid-conversation without invalidating the prompt cache, and there's a new option to automatically fall back to another model when a request gets flagged by safety classifiers
• Safety: in an automated behavioral audit, Opus 5 was rated Anthropic's most aligned model to date, scoring 2.3 on overall misaligned behavior, the lowest among recent models. At the same time, Anthropic is upfront that in biology research and offensive cybersecurity, Opus 5 still lags behind a model called "Mythos 5," partly because it wasn't intentionally trained on cyber tasks
• Data retention: consistent with prior Opus models, there are no data retention requirements for general access

Wrap-up

• Claude's new flagship model, Claude Opus 5, is available starting today, with a major performance boost at the same price as Opus 4.8 ($5 input / $25 output per million tokens)
• It beats every other model on the Frontier-Bench v0.1 coding evaluation with more than double Opus 4.8's performance, and scores three times as high as the next-best model on the ARC-AGI knowledge-work benchmark
• Its ability to verify its own work and persist through problems has improved significantly, making it stronger for long-running agents and complex analysis
• New operational features include Fast mode, mid-conversation tool switching, and automatic fallbacks
• It also sets a new bar for alignment in safety evaluations, while Anthropic stays cautious and transparent about remaining risk areas

I think this update will be especially exciting for developers running long-running coding agents, and for anyone who wants to hand off complex analysis or research work to Claude!