Anthropic Releases Claude Opus 5 as Safest Model Yet, Priced at Half of Fable 5

Anthropic has released Claude Opus 5, a new flagship model it is calling its safest and most aligned system to date. The model delivers intelligence that comes close to Claude Fable 5 at roughly half the cost, while achieving the lowest rates of deceptive behavior the company has seen in any of its releases.

Opus 5 was announced on July 24, 2026, and is already the default model on Claude Max, Anthropic's highest subscription tier at $100 per month. On Claude Pro ($20 per month), it is the strongest model available.

How Opus 5 compares to Anthropic's other models

According to Anthropic's own benchmarks, Opus 5 outperforms or matches Fable 5 across the vast majority of evaluation suites. The main exceptions are specialized health and legal tasks, where Fable 5 retains an edge.

On Frontier-Bench v0.1, Opus 5 sets a new state-of-the-art among all models tested and more than doubles Opus 4.8's score at a lower cost per task. On CursorBench 3.2 at maximum effort, it performs within 0.5 percentage points of Fable 5's peak score, but at half the cost.

The model also leads on several knowledge-work benchmarks:

• On ARC-AGI 3, an evaluation measuring the ability to solve novel problems, Opus 5's score is three times higher than the next-best model
• On GDPval-AA v2, which measures whether models can complete end-to-end business tasks, Opus 5's pass rate is approximately 1.5 times that of its closest competitor at the same cost
• On OSWorld 2.0, a computer-use benchmark, Opus 5 surpasses Fable 5's best result at just over a third of the cost

Anthropic says Opus 5 is particularly strong on scientific research tasks. It outperforms Opus 4.8 on every life sciences evaluation the company tested, with notable improvements in organic chemistry and protein-related tasks.

Alignment and safety as a product feature

Anthropic is placing unusual emphasis on alignment and safety for Opus 5, framing it as a deliberate product choice rather than a technical afterthought. In the company's pre-deployment behavioral audit, Opus 5 exhibited the lowest rates of deceptive behavior of any Claude model, adhered to Claude's Constitution more closely than Opus 4.8, Sonnet 5, or Fable 5, and was the least likely to be tricked into misuse.

On cybersecurity, Anthropic made a conscious design decision. Opus 5 was intentionally not trained on cyber tasks, a departure from some of its predecessors. It still improves at finding vulnerabilities as a byproduct of general capability gains, but it remains substantially behind Claude Mythos 5 on exploiting those vulnerabilities.

The company says it conducted these evaluations alongside private-sector and government partners, and the findings are documented in Opus 5's System Card.

What this means for the AI industry

Opus 5's release highlights a growing divergence in how AI labs position their products. Anthropic is betting that safety and reliability, not raw benchmark scores, will differentiate its models in a market where frontier performance is becoming increasingly similar across competitors.

The pricing is also notable. At $5 per million input tokens and $25 per million output tokens, Opus 5 is priced identically to Opus 4.8, but delivers significantly more capability. That puts pressure on rivals, including OpenAI's GPT-5.6 Sol, which faces a model that performs within striking distance of Anthropic's most capable tier at a much lower price.

Opus 5 is now available through Anthropic's API and on Claude Pro and Claude Max subscriptions.