Anthropic released Claude Sonnet 5.5 on September 28, claiming 30% faster output and up to 30% lower cost per task than Sonnet 5 despite unchanged token prices. That makes it a credible upgrade candidate, not an automatic replacement. Operators should compare accepted-output cost on their own workflows before changing production routing.

Key Takeaways

• Sonnet 5.5 keeps Sonnet 5’s $2 input and $10 output token prices.
• Anthropic attributes lower task cost to fewer tokens, not a cheaper rate card.
• One migration setting changes for applications that disable thinking.
• Operator posture: run a small test before changing model routes.

What did Anthropic actually release?

Sonnet 5.5 is available across Anthropic, AWS, Google Cloud, and Microsoft Azure under the model name claude-sonnet-5-5. Anthropic positions it between lower-cost, high-volume models and Opus 5.5, reserving Opus for complex work requiring sustained judgment. TechCrunch describes Sonnet as Anthropic’s mid-tier model for coding and office-document work.

The rate card did not fall: input remains $2 per million tokens and output $10. Anthropic says the model usually needs fewer tokens for the same work, producing its “up to 30%” task-cost claim. Its published benchmarks and customer quotes are launch evidence, not proof of lower cost in a buyer’s configured workflow. The accompanying system card documents the evaluation and safeguard scope, but cannot substitute for deployment testing.

Why is this a routing decision, not a price cut?

For operators, the useful comparison is cost per accepted output: total tokens, retries, tool calls, latency, reviewer time, and failure recovery. A faster model can lower queue time while still losing money if it needs more correction. A model with the same token price can become cheaper if it reaches the quality bar in fewer steps.

That makes Sonnet 5.5 a routing candidate for well-scoped coding, document, slide, and spreadsheet tasks—not a blanket replacement for judgment-heavy work. The distinction fits the broader need to track agent costs beyond seat licenses and to treat large model funding and capacity claims as procurement evidence, not outcomes.

There are two operational wrinkles. Anthropic says high-risk cyber requests may visibly fall back to Sonnet 5, so specialized security teams should record fallback rates and output differences. Applications that run with thinking disabled must also adopt the new between_tools setting; Anthropic’s migration guide should be treated as a release prerequisite, not optional reading.

How should operators test Sonnet 5.5 now?

Run a small test on 20–50 low-risk, repeatable tasks already handled by Sonnet 5 or another incumbent. Keep prompts, tools, permissions, effort settings, and acceptance rubrics constant. Measure accepted-output rate, total tokens, elapsed time, retries, reviewer minutes, and fallback events.

Do not reroute every workflow from one aggregate benchmark. Promote Sonnet 5.5 only where it lowers accepted-output cost or materially improves speed without increasing errors or supervision. Watch next for independent workload tests, production customer evidence, and clearer reporting on cost variance by effort level and task type.

FAQs

Is Sonnet 5.5’s API price lower than Sonnet 5’s?

No. Anthropic lists the same $2-per-million input and $10-per-million output token rates. Its savings claim comes from using fewer tokens per task in Anthropic’s tests.

Should businesses switch from Opus 5.5?

Not broadly. Anthropic still positions Opus for complex, open-ended work requiring sustained judgment. Test Sonnet on bounded tasks and route work by accepted-output quality, cost, latency, and required supervision.

What could break during migration?

Applications with thinking disabled need the new between_tools setting. Security workflows may also encounter model fallback on higher-risk requests. Validate configuration, logs, fallback behavior, and output quality in staging before changing production traffic.