Why Claude Sonnet 5 burns so many tokens (and what it costs you)
Claude Sonnet 5 has one of the most token-hungry tokenizers among top models — second-highest token usage on Cursorbench. Here's what that means for your bill.
Why Claude Sonnet 5 burns so many tokens
When a model wastes tokens, the “cheap” per-token price is a lie. Claude Sonnet 5 is the current poster child: its per-token price looks fine, but it’s one of the most token-hungry models among the top tier.
The tokenizer problem
Sonnet 5’s tokenizer is inefficient — it’s reported as the second-highest model for tokens used on Cursorbench. What that means in practice: the same task that another model completes in X tokens takes Sonnet 5 meaningfully more. So even at $2 / $10 per 1M, your real cost per task is higher than the sticker suggests.
The real cost math
| Claude Sonnet 5 | A more efficient rival | |
|---|---|---|
| Sticker (in / out per 1M) | $2 / $10 | $2 / $10 |
| Tokens per task | Higher | Lower |
| Real cost per task | Higher | Lower |
Because it also takes longer per task (it spends more tokens thinking), Sonnet 5 ends up spending too much time and money — practitioners report the output simply isn’t worth it versus Opus 5 or Fable 5.1.
What this means
- Don’t judge models on per-token price alone. Token efficiency is part of the real cost.
- For heavy Claude use, go bigger or different. Opus 5 and Fable 5.1 are more capable, and Fable 5.1 — despite the higher sticker — gets tasks done in one pass.
- The calculator shows sticker cost. For true cost, factor in how many tokens each model actually uses on your task type.
It’s why Claude Sonnet 5 ends up in the “don’t use” category in current practitioner rankings despite the mid-tier price. Compare the Claude family yourself: Sonnet 5, Opus 5, Fable 5.1 — or price out tokens in the AI intelligence calculator .