Anthropic Releases Claude Opus 5 With ARC-AGI-3 Score Three Times Higher Than Any Prior Model

Quick Facts

• Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens, roughly half the cost of Fable 5 for comparable performance.
• The model scored 30.2% on ARC-AGI-3, more than tripling the previous best result from any model on the benchmark.
• Opus 5 is now the default model on Anthropic’s top-tier Claude Max plan and is available to developers via the API as claude-opus-5.

Anthropic launched Claude Opus 5 on July 24, its fourth model release in under two months. The company positioned it as the default everyday model for most users and developers, sitting below the restricted Mythos 5 and the publicly available Fable 5 in its lineup, but delivering near-comparable results at a lower price.

Opus 5 is priced at $5 per million input tokens and $25 per million output tokens. It ships with a 1 million token context window, 128,000 max output tokens, and thinking enabled by default. A low, medium, and high effort toggle lets developers trade cost for capability on a per-request basis.

Benchmark Performance

The model’s most striking result came on ARC-AGI-3, a novel reasoning benchmark introduced in March 2026. When the benchmark launched, every frontier model scored under 1%, while human testers solved all of it. Opus 5 scored 30.2%. GPT-5.6 Sol scored 7.8%. Opus 4.8, Anthropic’s previous version released May 28, scored 1.5%.

On Frontier-Bench v0.1 agentic terminal coding, Opus 5 reached 43.3%. That beat Fable 5 at 33.7%, GPT-5.6 Sol at 34.4%, and Opus 4.8 at 21.1%. On CursorBench 3.2 at max effort, the model came within 0.5% of Fable 5’s peak score at half the cost per task.

On the knowledge work benchmark GDPval-AA v2, Opus 5 led all models with an Elo score of 1,861, ahead of Fable 5 at 1,747 and GPT-5.6 Sol at 1,736. On OSWorld 2.0, a computer use benchmark, it exceeded Fable 5’s best result for just over a third of the cost.

Opus 5 does not lead every category. On agentic coding via DeepSWE v1.1, GPT-5.6 Sol leads with 72.7%, followed by Fable 5 at 69.7% and Opus 5 at 68.8%. On health and legal benchmarks, Fable 5 and Mythos 5 outperform it, respectively.

Partner Results and Scientific Use

Zapier CEO Wade Foster said Opus 5 topped the company’s AutomationBench leaderboard without spending more tokens than prior Claude models. “It took a raw account-health workbook and ran a full churn-prevention sequence end to end: flagging at-risk accounts, alerting the right owner, and summarizing for retention ops. Previous models didn’t pass; Opus 5 hit 100%,” Foster said.

Cursor co-founder Sualeh Asif described the model as delivering “near Fable 5 intelligence at Opus speed and cost.” The model achieves a 1.5x higher pass rate than its closest competitor on Zapier’s AutomationBench for the same cost, even at its lowest effort setting.

Anthropic also claims Opus 5 is the most capable generally available model for scientific research. Compared to Opus 4.8, it gained 10.2 percentage points on organic chemistry tasks involving deriving molecular structures from spectroscopy data, and 7.7 percentage points on protein-related tasks.

What This Means for Buyers

Anthropic’s pricing structure now gives enterprise buyers a clear choice. Fable 5 remains the top publicly available model for specialized domains like law and medicine. Opus 5 covers agentic coding, knowledge work, computer use, and scientific research at roughly half the price, with a cost-efficiency toggle built in.

Haiku remains the only model in Anthropic’s lineup without a 5-series upgrade. Opus 5 is available now through the API, the Claude Pro plan, and as the default model on Claude Max.

Read more: Anthropic launches Claude Opus 5 with efficiency, safety improvements