- Anthropic’s Opus 4.7 carries the same per-token price as Opus 4.6, but a revised tokenizer encodes identical text into significantly more tokens, raising effective per-request costs.
- Developer Abhishek Ray’s measurements found token counts 1.325x higher on average for Claude Code content and up to 1.47x higher for technical documentation.
- A community dataset of 483 submissions on tokens.billchambers.me recorded a 37.4% average increase in token counts and costs per request.
- Opus 4.7 demonstrated a five-percentage-point improvement over Opus 4.6 on the IFEval instruction-following benchmark across 20 prompts.
What Happened
Anthropic’s Claude Opus 4.7 was released with the same nominal per-token price as its predecessor, Opus 4.6, but early measurements by developer Abhishek Ray—published on Claude Code Camp and reported by The Decoder—show that a new tokenizer causes the model to consume between 1.325x and 1.47x more tokens per request depending on content type.
Anthropic’s own migration guide acknowledges a 1.0 to 1.35x token increase range. Ray’s measurements align with that band for some content categories but exceed the upper bound for technical documentation and configuration files, putting the effective cost increase outside the range Anthropic cited for typical use.
Why It Matters
Token consumption is the primary cost variable in production deployments of large language models. A tokenizer that encodes more tokens per unit of text raises effective per-request costs even when the stated per-token rate stays flat—a discrepancy that does not appear on a model’s pricing page and requires direct measurement to detect.
The pattern is not new to the industry. OpenAI’s rollout of the GPT-4o tokenizer and Google’s varying tokenization behavior across Gemini model generations both produced similar hidden cost shifts in prior model transitions. The Opus 4.7 case is notable for the speed at which structured, reproducible measurements emerged from the developer community.
Technical Details
Ray’s analysis covered three content categories. Real Claude Code session content produced a 1.325x token count increase over Opus 4.6; a CLAUDE.md configuration file registered 1.445x; and technical documentation reached 1.47x. Code content showed the largest inflation; standard prose saw a smaller increase; and Chinese and Japanese texts were largely unaffected by the tokenizer change.
A separate community benchmark aggregating 483 submissions on tokens.billchambers.me found a 37.4% average increase in both token counts and costs per request across the full dataset. Ray modeled a representative 80-turn Claude Code session and estimated total cost rising from $6.65 under Opus 4.6 to between $7.86 and $8.76 under Opus 4.7—an 18% to 32% increase in practice.
On the capability side, Opus 4.7 showed a measured gain in instruction adherence: in a 20-prompt evaluation using the IFEval benchmark, the model followed strict formatting and structural constraints five percentage points more reliably than Opus 4.6.
Who’s Affected
Developers using Claude Code—Anthropic’s AI-assisted coding environment—face the steepest exposure, as code and technical documentation produce the largest token inflation in Ray’s measurements. Teams that maintain large CLAUDE.md context files or that pass extensive technical reference material into prompts will see costs closer to the 1.47x upper bound.
Enterprise customers who modeled Opus 4.7 budgets by extrapolating from Opus 4.6 token volumes will need to revise projections before migrating. Organizations processing East Asian language content or primarily prose-based workloads will see a smaller impact than code-heavy deployments.
What’s Next
Anthropic’s migration guide addresses the tokenizer change directly, indicating the company anticipated developer scrutiny. The company has not announced plans to revise its public pricing documentation to surface effective per-request cost comparisons alongside per-token rates.
The community dataset on tokens.billchambers.me continues to collect submissions beyond its current 483-entry baseline, and Ray’s per-content-type methodology provides a replicable framework that other developers can apply to their specific workloads before committing to a migration.
