Claude Sonnet 5's Tokenizer Makes Its Headline Price More Complicated
Anthropic's introductory $2/$10 API rate ends August 31. The same text can also consume up to 1.35 times as many tokens as it did with Sonnet 4.6, so buyers need to measure real workloads.
Claude Sonnet 5 arrives with a price that looks familiar and a tokenizer that makes simple comparisons surprisingly slippery.
Anthropic released the model on June 30 across Claude plans, Claude Code and the Claude API. Introductory API pricing is $2 per million input tokens and $10 per million output tokens through August 31, 2026. On September 1, those rates rise to $3 and $15.
The footnote that matters is the tokenizer. Anthropic says the same input can map to roughly 1.0 to 1.35 times as many tokens as it did with Sonnet 4.6, depending on the content. A lower price per token does not automatically mean a lower price for the same document or codebase.
What developers are getting
Sonnet 5 has a one-million-token context window and supports as many as 128,000 output tokens. Adaptive thinking is enabled by default, while the familiar sampling controls `temperature`, `top_p` and `top_k` are no longer supported.
Anthropic positions the model as a more agentic successor to Sonnet 4.6, with improvements in coding, browser and terminal use, planning and sustained tool work. The company says its higher-effort performance can approach Opus 4.8 on some tasks, while offering a broader range of cost and performance settings.
Those are vendor evaluations, not a promise that every workload will improve. The more immediate operational change is measurable: prompts may tokenize differently, effort settings can increase usage, and the introductory discount expires on a fixed date.
The tokenizer tax
Independent developer Simon Willison tested several large documents and found that the effect varied sharply by content. His examples used about 1.42 times as many tokens for an English document, 1.33 times for Spanish, 1.27 times for Python and almost the same number for Simplified Chinese.
That does not produce one universal “30% price increase.” It shows why teams should replay their own production traffic. An English-heavy coding assistant, a multilingual support system and a Chinese-language document pipeline could experience materially different changes.
Anthropic says the introductory rate is intended to make the transition roughly cost-neutral. That may be true for an average workload during the discount period; it cannot answer what an individual application will cost after August 31.
Safety changes deserve the same nuance
Anthropic reports lower overall rates of undesirable behaviour than Sonnet 4.6, along with stronger resistance to malicious requests and prompt-injection attempts. The company also says Sonnet 5 showed more undesirable behaviour than Opus 4.8 and Claude Mythos Preview on its automated behavioural audit.
In a Firefox exploit evaluation, neither Sonnet 4.6 nor Sonnet 5 produced a fully working exploit. Sonnet 5 did show a slightly higher partial-success rate, and Anthropic enabled cyber safeguards by default.
The useful conclusion is not that the model is simply safer or cheaper. It is that both claims depend on what is measured: content mix for cost, and the specific behaviour or capability test for safety.
Sources
Update note: Substantially revised and source-checked on 2026-07-24. Next checkpoint: standard pricing begins September 1, 2026.
Sources
- Anthropic — Introducing Claude Sonnet 5 — official
- Simon Willison — What's new in Claude Sonnet 5 — analyst
Drafted with AI assistance from verified source material and reviewed for factual accuracy, attribution, clarity, and label integrity.