A lower per token price does not guarantee a lower bill after a model upgrade. Anthropic's documentation says models from Claude 4.7 onward produce approximately 30 percent more tokens for the same text than earlier models. A 20 percent price cut with 30 percent more tokens raises cost by about 4 percent.
What Anthropic documents
The pricing page and the token counting page, both read on 19 October 2026, say that Claude 4.7 and later models use a newer tokenizer that produces approximately 30 percent more tokens for the same text, with the exact increase depending on the content and workload shape. Earlier models, such as Sonnet 4.6, use the previous tokenizer. The token counting page tells you to recount prompts against the model you plan to use rather than reuse counts from an older model, and says billing on newer models reflects the new tokenizer's counts.
Two worked examples
These use list prices from the same page and assume the same text, the same output length, and exactly 30 percent more tokens. Real results vary with content, so treat them as illustrations.
Sonnet. Sonnet 4.6 lists $3 input and $15 output per million tokens. Sonnet 5.5 lists $2 and $10. The price ratio is 0.67. Times 1.3 tokens is about 0.87, so the same text costs roughly 13 percent less on the newer model.
Opus. Opus 4.6 lists $5 and $25. Opus 5.5 lists $4 and $20. The price ratio is 0.8. Times 1.3 is about 1.04, so the same text costs roughly 4 percent more.
The lesson is not that newer is worse. It is that you cannot read a bill change off the price list alone.
A migration test in an afternoon
- Take 100 real prompts from production logs.
- Count them with the old model ID and the new one using the provider's token counter. Compare the input totals.
- Run both models on the same prompts and compare output tokens, since newer models may write longer or shorter answers, and may think more.
- Compute cost per prompt for each at list price, including caching if you use it.
- Check quality on your acceptance set. See choose a model tier.
- Compare rate limits. Anthropic's limits are set per model class, so your capacity may change too.
Other things that move on a migration
- Context thresholds, such as Haiku 5.5's higher price for prompts over 100,000 tokens. See long context pricing.
- Cache read multipliers, which Anthropic lists as lower on a few newer models.
- Retirement dates for older models. Several earlier Anthropic models are marked retired except on some cloud platforms.
Make it a habit
Treat each new model as a cost change to test, not a free upgrade. Record the price date, as in scheduled price changes in AI APIs.
Quick check
Does the tokenizer change affect output tokens too?
The page describes more tokens for the same text. Measure your own outputs on both models, because answer length also depends on the model.
Can I estimate without running the model?
You can count input tokens without generating output. For output, you need real runs.
Sources
Last reviewed: 19 October 2026.
