Claude Opus 5: Pricing, Benchmarks & What Changed
The Short Answer
Claude Opus 5 launched July 24, 2026 at $5 / $25 per million tokens — the same price as Opus 4.8 and half the cost of Fable 5. It offers near-Fable-5 intelligence, a 1M-token context window, and state-of-the-art agentic coding, and is now the default Opus model across the Claude API, AWS Bedrock, and the Claude apps.
What Changed vs Opus 4.8
Same price, meaningfully more model:
- Deep reasoning — a “step-change” in sustaining multi-step analysis across complex problems.
- Agentic coding & long-horizon tasks — holds focus over extended tool-use sessions, completes multi-file features and large refactors without leaving loose ends, and recovers from its own errors.
- Self-correction — catches its own mistakes and pushes back on unsound plans more reliably.
- Effort toggle — low / medium / high / xhigh / max, so you trade quality for cost per call.
- Alignment — Anthropic’s “most aligned model to date,” with lower deceptive-behavior rates.
- Developer features — mid-conversation tool changes (beta), a default fallbacks mode (beta), and a lower prompt-cache minimum (512 tokens, down from 1,024).
Unchanged: $5/$25 pricing, 1M-token context, and 128K max output — the same envelope as Opus 4.8.
Benchmarks (as of July 2026)
| Benchmark | Claude Opus 5 |
|---|---|
| Frontier-Bench (SOTA) | 43.3% |
| ARC-AGI-3 | 30.2% |
| SWE-bench Pro | 79.2% |
| Artificial Analysis Coding Index | 78.0% (better than 99% of models) |
| CursorBench | within ~0.5% of Fable 5 |
In head-to-head comparisons, Opus 5 outperforms GPT-5.6 Sol on six benchmarks including agentic coding and computer use, while GPT-5.6 Sol retains a lead on terminal coding (Terminal-Bench 2.1). Note: Anthropic did not publish an official SWE-bench Verified or Terminal-Bench figure for Opus 5 at launch — the SWE-bench Pro number above is from third-party runs.
Pricing in Context
| Model | $/MTok in | $/MTok out | 30K-in/5K-out task* |
|---|---|---|---|
| Claude Opus 5 | $5 | $25 | ~$0.28 |
| Claude Opus 4.8 | $5 | $25 | ~$0.28 |
| Claude Fable 5 | $10 | $50 | ~$0.55 |
| GPT-5.6 Sol | $5 | $30 | ~$0.30 |
| Claude Sonnet 5 (intro) | $2 | $10 | ~$0.11 |
Task-cost math = 30×(in/1000) + 5×(out/1000).
Who Should Use It
- Complex coding, multi-repo refactors, long-horizon agents → Opus 5 is now the strongest Opus.
- Document-heavy enterprise analysis → notable accuracy gains over 4.8.
- Cost-sensitive daily work → still consider Sonnet 5 ($2/$10 intro) and escalate to Opus 5 on failure.
- Peak-benchmark coding regardless of cost → Fable 5 or GPT-5.6 Sol still edge specific leaderboards.
Sources
- Anthropic — Claude Opus 5 announcement (July 24, 2026): anthropic.com/news/claude-opus-5
- Claude docs — what’s new in Opus 5: platform.claude.com/docs/en/about-claude/models/whats-new-opus-5
- AWS — introducing Claude Opus 5 on AWS: aws.amazon.com/blogs/machine-learning/introducing-claude-opus-5-on-aws
- OpenRouter — Claude Opus 5 pricing: openrouter.ai/anthropic/claude-opus-5