Opus 4.7 Burns More Tokens Than 4.6 at Same Price
Anthropic released Opus 4.7 with the same listed price as Opus 4.6. Initial token counts indicate it uses substantially more tokens for each request. Developer Abhishek Ray shared these findings on Claude Code Camp.
Anthropic, a company focused on safe AI systems, built the Claude family of models. Former OpenAI staff founded it in 2021. Opus serves as the top-tier model in this lineup, handling complex tasks like coding and analysis. Pricing follows a token-based system, where costs depend on input and output tokens processed.
Token Increase Matches and Exceeds Guide
Anthropic's migration guide mentions a token rise of 1.0 to 1.35 times. Ray's tests align closely with this for most cases. Real Claude Code content averaged 1.325 times more tokens. A CLAUDE.md file reached 1.445 times. Technical documentation hit 1.47 times.
Code content suffered the largest increase, according to Ray. Prose experienced a smaller rise. Chinese and Japanese texts showed almost no change.
Community Tests Show Even Higher Usage
A community evaluation on tokens.billchambers.me reported stronger results. It covered 483 submissions. Token usage jumped 37.4 percent. Per-request costs followed the same pattern.
These figures come from diverse user tests. They highlight real-world differences beyond official estimates.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
Cost Estimates for Typical Sessions
Ray calculated costs for a sample session with 80 turns. Switching to Opus 4.7 added 20 to 30 percent to expenses. The original bill stood at $6.65. New totals ranged from $7.86 to $8.76.
Such increases matter for heavy users. Developers and businesses track token counts to manage budgets. Flat pricing masked these shifts until hands-on tests appeared.
Gains in Instruction Following
Users receive some benefits. Opus 4.7 follows instructions better in certain tests. On the IFEval benchmark with 20 prompts, it succeeded five percentage points more often than Opus 4.6.
IFEval measures adherence to rules in prompts. This small improvement suits tasks needing precision, like code generation or strict formatting.
Anthropic continues to update Claude models. Opus versions evolve with better capabilities, though efficiency varies. Early adopters now adjust workflows based on these token realities.