Claude Sonnet 5: Pricing Correction and Migration Notes

Analysis

The pricing correction comes first

When Sonnet 5 launched, the platform guide listed introductory API pricing of $2 input and $10 output per million tokens through August 31, 2026, with a stated return to $3 / $15. September 1 came and went without the increase. The pricing page now carries $2 / $10 as the standard rate.

This page previously told readers to reconcile intro-window cost against a higher steady state. That advice is retired. If your forecast, chargeback model, or build-versus-route decision used $3 / $15 for September onward, it is overstating Sonnet 5 cost by 50%. Rebuild it from the pricing page.

• Sonnet 5 standard pricing: $2 input, $10 output per million tokens.
• Cache reads on Sonnet 5 follow the usual 10% of base input; the 2.5% rate is specific to Fable 5.1 and Mythos 5.1.
• The Batch API 50% discount still applies to non-interactive Sonnet 5 work.
• Anthropic did not publish a replacement end date, so treat $2 / $10 as current rather than permanent and recheck the pricing page before long-term commitments.

Analysis

Where Sonnet 5 sits on the September board

Sonnet 5 is no longer the top of the recommended path. Anthropic’s selection guidance points to Opus 5 as the starting point for most workloads and describes Sonnet 5 as the speed and intelligence balance. That is a demotion in positioning and an improvement in value: same 1M context window, same 128K max output, adaptive thinking, and a fifth of the Fable 5.1 input price.

The practical read is that Sonnet 5 became the better default for high-volume production traffic while Opus 5 became the better default for interactive and difficult work. Split by task family rather than picking one.

Analysis

Migration mechanics that still apply

Everything the platform migration guide documented at launch remains relevant for anyone moving off Sonnet 4.6 or older: adaptive thinking is on by default, some sampling parameters are unsupported, manual extended-thinking handling changed, and the tokenizer can produce roughly 30% more tokens for identical input. That last one silently breaks tests that assert token counts and budgets that assume a fixed ratio.

Start with representative prompts rather than synthetic demos. Include the cases that depend on exact JSON structure, refusal handling, system-message placement, retrieval chunk sizes, long code context, and hard cost ceilings.

• Inventory every hard-coded model ID and route before switching traffic.
• Re-run any test that asserts token counts, output length, or cache economics.
• Remove or gate unsupported sampling parameters in client libraries.
• Inspect refusal and stop-detail telemetry separately from ordinary errors.
• Recompute cost per successful answer, not cost per token, when comparing against Opus 5.

Analysis

Recommended rollout shape

A careful rollout is narrow: route one task family, run old and new outputs side by side, inspect the failure set, then expand. Pick a first family with clear success criteria and low support blast radius. Do not use a model migration as cover for rewriting prompts, changing retrieval, and changing tool schemas at the same time.

When the new model wins, write down why. The durable artifact is a migration note naming the tasks, model IDs, eval delta, token-cost delta, and client changes required. That note is what makes the next Anthropic release a one-day exercise instead of a one-quarter one.