AI BriefingThe tokenizer drove up the cost
Key point
Opus 4.7 keeps the same price, but token counts can increase up to 1.35x.
Details
Opus 4.7 kept its $5/$25 per million tokens pricing, but due to the new tokenizer, the same input can be counted as up to 1.35x more tokens.
Comparing a simple Python binary search bug-fix task:
- Opus 4.7: 1,381ms, input 202, output 141, cost $0.0136
- Sonnet 4.6: 14,142ms, input 170, output 795, cost $0.0124
- Both models successfully fixed the answer
In this test, Opus was 10x faster while costing about the same as Sonnet, and Sonnet's advantage from its cheaper per-unit price was largely offset by excessive explanatory output.
When the same task was handed to a routing layer, gemini-2.0-flash was recommended, suggesting that for such simple bug fixes, a cheaper model than Claude's higher-tier models is suitable in the first place.
The key point is that Claude Code's approach of sending every turn in a session to Opus by default is inefficient, and in 4.7, the tokenizer change on top of that has raised the perceived cost even further.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.