Claude Sonnet 5.5 keeps sticker price but cuts real cost per task by 30%

General
1

Claude Sonnet 5.5 keeps sticker price but cuts real cost per task by 30%

Anthropic released Claude Sonnet 5.5 on September 28, 2026, six days after Opus 5.5. The pricing looks identical to Sonnet 5 on a token basis — $2 per million input tokens, $10 per million output — so at first glance nothing changed. The actual argument for the update is more specific: the model uses fewer tokens and tool calls to complete the same tasks, which Anthropic says cuts effective cost per task by up to 30% compared to Sonnet 5, and runs more than 30% faster.

What “effective cost per task” means in practice

Token pricing is per token consumed, not per task completed. If a model accomplishes a given task in fewer turns and with fewer tool calls, the per-task cost drops even with the same sticker price. This is a meaningful distinction for agentic workloads, where a task might involve dozens of sequential model calls before reaching a result.

The claimed 30% reduction in effective cost per task is a vendor figure, and Anthropic has not published a methodology for how that number was derived or over what distribution of tasks it was measured. Teams should treat it as a directional claim and measure against their own workloads rather than taking it as a guaranteed budget reduction. The 30%+ speed improvement is similarly a published vendor claim without independently verified conditions.

For workloads that are primarily single-turn completions — one prompt, one response — the efficiency argument is less relevant, since task completion doesn’t involve multi-step overhead. The case is strongest for tool-calling pipelines and agentic loops.

Context, output limits, and deprecations

Sonnet 5.5 carries a 1M token context window and 128k maximum output, matching what Opus 5.5 offers. Those are the same specs as Sonnet 5.

The release comes with a significant deprecation notice: Opus 5 and Sonnet 5 both move to legacy status with Sonnet 5.5’s release. Sonnet 4.5 is deprecated outright, with API retirement set for November 30, 2026. Teams still running Sonnet 4.5 in production have roughly two months to migrate. Haiku 5.5 is confirmed as forthcoming, though Anthropic has not given a release date.

The migration question

If a team is currently on Sonnet 5 and primarily running agentic or tool-heavy workloads, the efficiency argument favors migrating to Sonnet 5.5 — the sticker price is the same, and fewer tokens consumed per task should lower the bill. The migration path is straightforward since the pricing structure is identical; the main risk is any behavior differences that need validation against existing prompts and pipelines.

For teams on Sonnet 4.5, migration is not optional by November 30. The practical choice is between Sonnet 5.5 and Opus 5.5. Given the 5x price difference between the two — $2/M vs. $10/M on output — Sonnet 5.5 remains the default cost-efficient choice unless the task profile genuinely warrants Opus-tier capability.

The 1M context window on Sonnet 5.5 removes one of the reasons teams historically reached for Opus: very long document work. With both models now offering the same context ceiling, capability differences are more about reasoning depth and agentic reliability than raw context capacity.

Sources

2
The claim that Sonnet 5.5 uses 30% fewer tokens per task is directionally useful, but Anthropic hasn't published a methodology for how that number was measured — which tasks, what prompt structures, what tool configurations. For teams trying to verify this against their own workloads rather than taking it as given, the measurement approach matters. The most direct way to test it is to instrument your pipeline to log total tokens consumed per completed task rather than per API call, then compare Sonnet 5 and Sonnet 5.5 on the same fixed set of representative tasks and look at the distribution, not just the mean. A 30% improvement at the median can coexist with no improvement on certain task types. One question I haven't seen addressed: does the efficiency gain apply equally to pipelines using native tool calling versus those using more verbose prompt-based tool descriptions? The mechanism for reduced token usage — fewer intermediate reasoning steps or fewer tool call roundtrips — would affect those two patterns differently, and understanding which one drives the savings would help teams know whether the gain applies to their specific setup.
3
For high-volume tasks — classification, routing, short extraction, simple Q&A at scale — the trade-off between Sonnet 5.5 and a Haiku-class model isn't just price per token; it's price per correct answer. Sonnet 5.5 at $2/M input and $10/M output is more expensive per token than Haiku-class models, but if it solves a task in one call that a Haiku model gets wrong 15-20% of the time and requires a retry or fallback, the effective cost per successful completion can end up comparable or worse on Haiku. The 30% efficiency improvement in Sonnet 5.5 is most relevant for multi-step agentic work, not for single-turn high-volume classification where the bottleneck is accuracy rather than token count. The calculus for each tier comes down to task structure: if the task requires sustained reasoning or careful instruction following across many steps, Sonnet 5.5's accuracy premium justifies the price. If it's a well-bounded task a smaller model handles reliably, Haiku-class remains the right choice even after Sonnet 5.5's efficiency gains — and Haiku 5.5 being forthcoming may shift this comparison further once its actual pricing and accuracy numbers are available.