17 November 2026 · Site Methodology & Corrections · 5 min read
Sonnet 4.6 charged $3 per million input tokens and $15 per million output tokens. Sonnet 5 charges $2 and $10. That's a 33.3% cut on both, and it's the number every pricing tracker, including ours, leads with. It's also not the number that lands on your invoice, because Anthropic changed something else at the same time that almost nobody priced in: the tokenizer.
Buried in Anthropic's own pricing documentation, well past the rate tables, is this: "Claude 4.7 and later models and Claude Mythos Preview use a newer tokenizer that contributes to their improved performance on a wide range of tasks. This tokenizer produces approximately 30% more tokens for the same text. [...] Claude Sonnet 4.6 and earlier models use the previous tokenizer."
Sonnet 5 ships in the 4.7-and-later generation. So the comparison between Sonnet 4.6 and Sonnet 5 isn't "same text, cheaper rate." It's "same text, cheaper rate, and more tokens to pay that rate on." Two numbers moved in opposite directions at the same time, and only one of them made it into the headline.
Take a workload that doesn't change: the same prompts, sent the same number of times a month, measured once under each tokenizer. Say it comes out to 100M input tokens and 20M output tokens a month under Sonnet 4.6's (older) tokenizer.
| Fichas de entrada | Fichas de salida | Costo de insumos | Costo de producción | Total/mes | |
|---|---|---|---|---|---|
| Sonnet 4.6 ($3/$15, old tokenizer) | 100M | 20M | $300.00 | $300.00 | $600.00 |
| Sonnet 5, naive estimate (reuses old token count) | 100M | 20M | $200.00 | $200.00 | $400.00 |
| Sonnet 5, actual (same text, new tokenizer, +30%) | 130M | 26M | $260.00 | $260.00 | $520.00 |
Someone who reads "33% cheaper" and reuses their old token count expects the bill to fall to $400/month. The real bill, once the same text is re-tokenized under Sonnet 5's tokenizer, is $520/month — a 13.3% reduction from the $600 baseline, not 33.3%. The gap between the naive estimate and the real number is $120/month, and it exists entirely because of token count, not price.
If the 30% figure is a real tokenizer effect and not a rounding choice, it should show up on numbers that have nothing to do with your content. It does. Anthropic's own tool-use token tables list a fixed overhead per tool definition that is tokenizer-dependent, not workload-dependent:
| Tool | Claude 4.6 and earlier | Claude 4.7 and later | Cambiar |
|---|---|---|---|
| Bash tool (added input tokens) | 244 tokens | 325 tokens | +33.2% |
| Text editor tool (added input tokens) | 745 tokens | 974 tokens | +30.7% |
Same tool, same JSON schema, same functionality — the only thing that changed is which tokenizer counts it. Both land in the same band as the documented "~30% more tokens" line, which is a useful sanity check: this isn't one cherry-picked number, it's a consistent property of the new tokenizer showing up everywhere Anthropic publishes a token count.
None of this means Anthropic's 33% rate cut is fake. The per-token price really did fall. What changed underneath it is the exchange rate between "your text" and "tokens," and that number is model-generation-specific, not vendor-wide — switching from Sonnet 4.6 to GPT or Gemini doesn't carry the same 30% tax, because it's not their tokenizer. The only place this bites is generation-to-generation comparisons inside the same vendor, which is exactly the comparison most people make when a new model drops and they re-run last month's numbers through a new rate card.
Don't carry over a token count across a model-generation boundary. Before you compare Sonnet 4.6 to Sonnet 5, or budget for any jump across the Claude 4.6 → 4.7 line, re-count your real prompts and documents against the new model with a tokenizer-aware tool — not the old estimate with a new price pasted in. Check current Claude rates on the Claude API pricing page, count actual tokens for your own text with the contador de fichas, and model total cost with the Calculadora de costos de tokens LLM o el completo Comparación de precios de LLM tool. If your workload leans on repeated context, also check the calculadora de ahorro de almacenamiento en caché rápida — cache multipliers apply on top of whatever the tokenizer hands you, in the same direction as everything above.
Figures from Anthropic's published pricing and tool-use documentation (platform.claude.com/docs/en/about-claude/pricing), checked 17 November 2026. The "~30% more tokens for the same text" figure is Anthropic's own stated approximation, not ours; the exact increase depends on content and workload shape, which is also what Anthropic's docs say. The worked example uses illustrative round-number volumes to isolate the tokenizer effect, not a specific customer's real traffic.