Anthropic has launched Claude Opus 5, which the company claims is approaching the capabilities of its flagship model, Fable 5. The new model is now available across all platforms.

https://twitter.com/claudeai/status/2080699495453528290

Claude Opus 5 has been designated as the default model in the Max subscription, with pricing remaining the same as Opus 4.8 — $5 per million input tokens and $25 per million output tokens.

According to Anthropic, Opus 5 has achieved performance levels close to those of Claude Fable 5 while utilizing half the token costs. The company also stated that this model has set a new benchmark in programming skill assessments and cognitive tasks, such as Frontier-Bench and GDPval-AA, although it still lags behind Mythos 5 in cybersecurity tasks.

Source: Anthropic.

In the Frontier-Bench v0.1 benchmark, Opus 5 outperformed Opus 4.8 by more than double while incurring lower costs per task. In CursorBench 3.2, it achieved a score that was less than 0.5% behind the peak result of Fable 5 at half the cost.

In the ARC-AGI 3 test, Opus 5 delivered results three times higher than its closest competitor. On the Zapier AutomationBench platform, which evaluates business problem-solving, the success rate was approximately 1.5 times higher for equal financial expenditure. In the OSWorld 2.0 test, which assesses computer performance, Opus 5 surpassed the best result of Fable 5.

Anthropic highlighted advancements in scientific tasks as well, noting that Opus 5 outperformed Opus 4.8 across all evaluations in life sciences. In their internal benchmark for organic chemistry, the score improved by 10.2 percentage points, while in protein-related tasks, it rose by 7.7 percentage points.

The company stated that Opus 5 is the most compliant model according to its standards. In an automated behavioral audit, it recorded a total inconsistency score of 2.3, the best result among recent Anthropic models.

Source: Anthropic.

For cybersecurity tasks, Anthropic has relaxed restrictions compared to Fable 5, estimating that classifier interventions will occur approximately 85% less frequently. Requests flagged by security systems in Claude.ai, Claude Code, and Claude Cowork will, by default, be rerouted to Opus 4.8, with a similar switch available in the API.

Additionally, Anthropic noted that biological queries that were blocked on Fable 5 will now be directed to Opus 5 instead of Opus 4.8.

Alongside the new model, Anthropic has introduced two beta updates: the ability to change available tools during a dialogue without resetting the prompt cache and automatic switches to backup models in the API.

It is worth noting that Opus 4.8 was launched in May, coinciding with the addition of dynamic workflow features to Claude Code.