Claude Sonnet 4.6 Guide: The Last Sonnet on the Old Tokenizer

Sonnet 4.6 is now two Sonnet generations old, and on a price sheet it looks beaten: Sonnet 5.5 lists at $2 / $10 per million tokens against Sonnet 4.6's $3 / $15. But the rate card is not the bill. Sonnet 4.6 is the last Sonnet on Anthropic's older tokenizer, and it still accepts several request parameters the newer models reject. For some workloads those two facts outweigh a lower list price.

This guide covers what Sonnet 4.6 is, why the per-token comparison misleads, and a plain rule for when to keep it.

Near-Opus at a Sonnet price

Anthropic released Sonnet 4.6 on 17 February 2026 at the same price as Sonnet 4.5, and made it the default model for Free and Pro users on claude.ai at the time. Its pitch was Opus-class work at a Sonnet price, across coding, computer use, long-context reasoning, agent planning and knowledge work.

The supporting evidence in the announcement was mostly preference data rather than leaderboard scores:

• In Claude Code testing, users preferred it to Sonnet 4.5 roughly 70% of the time, and to Opus 4.5 59% of the time.
• Testers rated it less prone to overengineering and laziness than Opus 4.5, with fewer false claims of success.
• Anthropic reported 80.2% on SWE-bench Verified with a prompt modification, and 60.4% on ARC-AGI-2 at high effort.

Anthropic still recommended Opus 4.6 for the deepest reasoning, such as large refactors and multi-agent coordination. That division of labour is the useful part: Sonnet for fast turnaround and volume, Opus when a failed run costs more than the tokens.

The context window is 1M tokens, billed at standard rates with no long-context premium, and maximum output is 128K. It launched in beta and is now the default in Anthropic's documentation.

The last Claude on the old tokenizer

Anthropic's pricing page states that Claude 4.7 and later models use a newer tokenizer producing approximately 30% more tokens for the same text, and that Sonnet 4.6 and earlier use the previous one. Sonnet 5 adopted the newer tokenizer, and Sonnet 5.5 kept it.

That changes the comparison. If your prompts tokenize about 30% larger on Sonnet 5.5, its $2 input rate behaves more like $2.60 against Sonnet 4.6's $3, and the $10 output rate more like $13 against $15. Sonnet 5.5 is still cheaper per unit of text, but by a much smaller margin than the rate card suggests, and the real gap depends on your content. Run count_tokens on both with your own prompts before deciding. Tokenizer differences across models explains why.

Controls the newer models removed

Sonnet 4.6 sits before several API changes. According to Anthropic's documentation, it still supports:

• Sampling parameters. temperature, top_p and top_k work. On Claude 4.7 and later models a non-default value returns a 400.
• Turning thinking off. {type: "disabled"} is accepted at any effort. Sonnet 5.5 rejects it and offers a narrower thinking-between-tools mode instead.
• Forced tool use. tool_choice of any or a named tool works. Sonnet 5.5 rejects both.
• Summarised thinking by default. Newer models return empty thinking text unless you opt in.

Two things it does not support: assistant-turn prefill, which was removed for the whole 4.6 family, and the xhigh effort level, which arrived with Opus 4.7. Thinking budgets via budget_tokens still work but are deprecated; use adaptive thinking for new code.

If you run a classification or extraction pipeline that depends on low temperature or forced tool calls, Sonnet 4.6 runs it unchanged. Moving to Sonnet 5.5 means rewriting it around structured outputs. Structured outputs and JSON mode covers that rewrite.

Effort: set it explicitly

Sonnet 4.6 introduced the effort parameter to the Sonnet line, defaulting to high. Anthropic's migration notes warn that moving from Sonnet 4.5 without setting effort can raise latency and token use noticeably. Their suggested starting points:

• Chat, classification, content generation — low, optionally with thinking disabled.
• Most applications and agentic coding — medium.
• Long-horizon autonomous agents and computer use — high with adaptive thinking.

Sonnet 4.6 or Sonnet 5.5?

For new projects with no legacy constraints, start with Sonnet 5.5. It is the current Sonnet, Anthropic lists it as the recommended replacement for older Sonnets, and it is cheaper per unit of text even after the tokenizer adjustment. The Sonnet 5.5 guide covers its own breaking changes.

Keep Sonnet 4.6 when:

• Your pipeline relies on sampling parameters, forced tool calls or fully disabled thinking, and the rewrite is not scheduled yet.
• You have measured Sonnet 4.6 doing better on your task. Preference shifts between generations are real.
• Your cost model, rate limits and max_tokens ceilings are calibrated to its token counts and you need stability for now.

Anthropic lists its retirement as not sooner than 17 February 2027, so plan the move rather than rush it. If you are still on the previous generation, the Sonnet 4.5 guide covers a more urgent deadline.

Variants and a decision rule

The catalogue lists the standard endpoint and a :batch variant: the same model on Anthropic's asynchronous Batch API at half price, $1.50 / $7.50 per million tokens. For offline jobs that halves the bill for the cost of latency.

The rule: if your integration uses any control listed above, Sonnet 4.6 is the safe home until you rewrite it. If it does not, count tokens for your real prompts on both models, multiply by each price, and pick the cheaper one that passes your evals.