Claude Opus 5

Claude Opus 5, released July 24, 2026, is Anthropic's most capable Opus-tier model and its strongest daily driver for serious engineering. It brings a 1M-token context window, thinking enabled by default, and five effort settings (low to max) so developers can trade tokens for intelligence. Priced at the same $5/$25 per million tokens as Opus 4.8, it scores at or near the level of the much more expensive Claude Fable 5 on coding, agentic, and knowledge-work benchmarks.

Core Features

• Effort-controlled reasoning: Five effort levels (low, medium, high, xhigh, max) plus thinking on by default; xhigh often beats max at ~15% fewer output tokens.
• 1M-token context: The default and maximum context window with consistent instruction following and tool calling throughout.
• 128K max output tokens: Enough headroom for long agent transcripts and multi-file code generation.
• Fast mode: Research preview that runs about 2.5x faster at 2x base price ($10/$50 per MTok) on the Claude API.
• Fewer refusals: Roughly 85% fewer false-positive safety interventions than earlier Claude generations.
• Wide availability: Claude API, Claude Pro/Max/Team/Enterprise, plus Amazon Bedrock, Google Cloud, and Microsoft Foundry.

Benchmark Highlights

• Artificial Analysis Intelligence Index: 61 (max effort), tied with Claude Fable 5 and ahead of GPT-5.6 Sol (59) and Kimi K3 (57).
• Frontier-Bench v0.1: Surpasses all other models and more than doubles Opus 4.8 at a lower cost per task.
• CursorBench 3.2: Within 0.5% of Fable 5's peak score at half the cost per task.
• ARC-AGI 3: A score three times higher than the next-best model on novel reasoning problems.
• IMO 2026: A perfect 42/42 with zero-shot, self-contained proofs.
• GDPval-AA v2 / AA-Briefcase: New leader in agentic knowledge work (1861 Elo), more than 100 points ahead of Fable 5 and GPT-5.6 Sol.

Use Cases

• Autonomous coding agents: Claude Code and Cursor workflows that need self-verifying, long-running engineering loops.
• Deep knowledge work: Research, document synthesis, and professional deliverables where the model checks its own work.
• Computer use: OSWorld 2.0 and Zapier AutomationBench top results make it a strong choice for end-to-end automation.

Tradeoffs to Know

The model is slower on hard tasks (25-36 minutes on the toughest agentic problems) and its newer tokenizer produces roughly 30-47% more tokens for the same code, which can raise API costs on long sessions unless you use prompt caching. Some developers also report a "max effort trap": pushing effort to max can lead to overthinking and rewrites, while xhigh delivers better results for less. Factual claims under uncertainty still need verification, as independent tests measured a higher hallucination rate on unknown facts.

Getting Started

Call claude-opus-5 via the Claude API, or select it in Claude apps, Cursor, or Claude Code. Start with medium or xhigh effort for interactive work and reserve max for high-stakes autonomous tasks.

Alternatives

• Claude Fable 5: Anthropic's most powerful model when cost is not the constraint.
• GPT-5.6 Sol: OpenAI's more usable daily driver with a different reasoning profile.
• Kimi K3: Open-weight 2.8T model with comparable agentic scores at much lower API prices.

Conclusion

Claude Opus 5 is the strongest value in the Claude lineup: near-frontier intelligence at the workhorse price point. If you build agents or write serious code and want the reliability of the Claude ecosystem without paying Fable 5 prices, it is the model to start with.