Claude Opus 5 vs Opus 4.8: Same Price, Double the Agentic Performance

TL;DR

Claude Opus 5 costs exactly what Opus 4.8 cost — $5 per million input tokens, $25 per million output — and roughly doubles agentic performance. SWE-bench Pro jumps from 69.2% to 79.2%, Frontier-Bench performance doubles at lower cost, and the model adds a 1M-token context window, an effort toggle, and extended thinking on by default. Anthropic considers Opus 4.8 superseded. There is no scenario where staying on 4.8 makes sense.

What Changed

Raw capability. The ten-point jump on SWE-bench Pro (69.2% → 79.2%) understates the difference in practice. On Frontier-Bench v0.1, Anthropic reports Opus 5 doubles Opus 4.8's performance while spending less per task, because it wastes fewer tokens on dead ends and verifies its own work more reliably. Context. Opus 5 moves to a 1M-token context window as both default and maximum, with 128K max output (300K via the Message Batches API beta). Whole-repository analysis that required chunking on 4.8 now fits in a single request. Effort control. A new per-request toggle — low, medium, high — lets you trade capability for cost and latency. Low effort makes Opus 5 economical for tasks you'd previously have routed to a smaller model; high effort buys longer reasoning on the hardest problems. Extended thinking by default. Opus 5 reasons before answering out of the box, which shows up most clearly on debugging and multi-step planning. Iterative verification. The launch materials highlight much stronger self-checking: Opus 5 re-runs and validates its own outputs during agentic tasks rather than declaring victory early. In one showcased example it built a custom computer-vision pipeline when direct visualization wasn't available, and performed root-cause debugging of open-source packages.

What Didn't Change

• Pricing: identical at $5/$25

• API surface: standard Messages API; migration is a model-ID swap to claude-opus-5

• Platform support: claude.ai, Claude Code, Claude Cowork, API

Migration Notes

1. Swap the model ID to claude-opus-5 in staging

2. Re-run your eval suite — expect fewer iterations per task, which usually lowers total token spend

3. Review prompts that compensated for 4.8 weaknesses (forced chain-of-thought, manual verification steps) — many are now redundant and just add cost

4. Experiment with effort settings: medium is a sensible default; drop to low for mechanical tasks

Conclusion

Same price, double the agentic performance, five times the context. Opus 4.8 had a good five-month run, but this is the easiest migration call of the year: upgrade, re-baseline your evals, and simplify the prompt scaffolding you no longer need.