Claude Opus 5 and Sonnet 5 Landed Together. The Sonnet Tier Is the Real Story
Anthropic shipped Claude Opus 5 and Claude Sonnet 5 together on July 8. Opus 5 posts 78.1% on SWE-bench Verified and holds context across day-long agent runs. Those are the numbers in the announcement. They are not the interesting part.
Look at Sonnet
Sonnet 5 scores 72.6% on SWE-bench Verified, 94.1 on HumanEval and 92.2 on MMLU, with a 1M token context window as standard rather than as a premium tier. It costs $3 per million input tokens, which is the same price Sonnet has cost since 2024.
Compare that against Claude Opus 4.6, which was the top model on our leaderboard in April. Opus 4.6 scored 58.4% on SWE-bench Verified at $15 per million input. Sonnet 5 beats it substantially at a fifth of the price, four months later.
That is the pattern worth internalising. The frontier moving up is a story about what becomes possible. The tier below the frontier absorbing last quarter's frontier capability at a fifth of the price is a story about what becomes affordable, and the second story affects far more people.
What Opus 5 Adds
Two things that matter for long-running agents.
Adaptive effort control. Rather than setting a fixed thinking budget, you set an effort level and the model decides how much reasoning a given step warrants. In practice this cuts token spend on easy steps substantially without hurting hard ones. If you have been manually tuning thinking budgets, you can stop.
Reworked compaction. The mechanism that lets an agent summarise and discard its own history got significantly better at deciding what to keep. This is why the day-long-run claim holds. It is not a bigger window, it is a smarter one. We wrote more about why that distinction matters in our piece on compaction versus context size.
The 1M Tier
Opus 5 also offers a 1M context variant at $6 per million input, applied to requests above 200K tokens. Whether that is worth it depends entirely on whether you actually need whole-repository context in one pass, and most teams that think they do would get better results from retrieval plus a 200K window.
Model the cost difference on our annual cost calculator before committing. The long-context premium adds up fast at volume.
Where Haiku Is
Haiku 4.5 remains the small tier. There is no Haiku 5 at time of writing, which is worth stating explicitly because a lot of coverage assumed a full three-model refresh. For routing, classification and sub-agent duty, Haiku 4.5 is still the pick in the Anthropic lineup.
Practical Recommendation
- Default to Sonnet 5. For the large majority of production work it is now the correct answer on price and capability together.
- Escalate to Opus 5 for long-horizon autonomous work, large migrations and anything where a failed run costs more than the token difference.
- Reach for Fable 5 only when you genuinely need the absolute top, and check your budget first at $10 per million input.
Compare all three side by side on the comparison tool, or see the Sonnet 5 model page for full specifications.