Anthropic released Claude Opus 4.7 this month with headline pricing unchanged from Opus 4.6, but independent measurements published on April 19, 2026 show the new model's tokenizer can silently inflate real bills by up to 47% on the same workloads.
The story, surfaced by The Decoder citing work from Claude Code Camp, is the first clear data point showing how a tokenizer swap, not a price sheet, has become the pricing lever that matters for frontier models. Developers who migrated their agents over the weekend were the ones to catch it.
What changed under the hood
Opus 4.7 ships with a new tokenizer that splits the same text into more pieces than Opus 4.6's did. Because Anthropic charges per token, a rate card that looks identical produces a noticeably larger invoice for the same prompts and outputs.
Anthropic's own migration guide acknowledges the shift, telling developers to expect input to map to roughly 1.0 to 1.35 times more tokens than on 4.6. The company did not raise its per-token rates alongside the change, and there is no accompanying statement from leadership explaining the rationale.
The measurements from the field
Developer Abhishek Ray ran a controlled test and reported a 1.325x average token count on real Claude Code content, rising to 1.445x for a typical CLAUDE.md project instructions file and 1.47x for technical documentation. A separate community evaluation hosted at tokens.billchambers.me recorded a 37.4% average token increase across 483 submissions.
Prose and natural language text land at the low end of that range. Code sits at the high end. Chinese and Japanese inputs, Ray found, show minimal impact, presumably because the new tokenizer's changes are concentrated in the segments used most heavily by English-language code and markdown.
In a worked example, an 80-turn agent session that previously cost about $6.65 on Opus 4.6 would now cost roughly $7.86 to $8.76, a 20 to 30 percent jump on identical work.
Why Anthropic likely did this
Opus 4.7 isn't a pure repricing. Anthropic paired the tokenizer swap with real capability gains: the company reports a five-percentage-point improvement on the IFEval benchmark for strict instruction following over Opus 4.6. A tokenizer retrained on more code-heavy data is a plausible ingredient in that kind of gain.
The trade-off is that engineering teams budgeting for agentic workloads now need to retest assumptions. Sticker-price parity no longer guarantees cost parity, especially for code-heavy agents and long Claude Code sessions where the inflation stacks across every turn.
Implications for the market
For enterprises, the episode underscores how opaque frontier-model pricing has become. Benchmarks, context windows, and rate cards are the public numbers, but tokenizer efficiency increasingly determines who wins a bake-off on dollars-per-task. Expect procurement teams to start demanding tokenizer-adjusted cost estimates alongside quality benchmarks when they evaluate Claude, GPT, and Gemini.
It also hands competitors a marketing opening. A rival lab that can match Opus 4.7 on IFEval while keeping token counts flat now has a concrete, defensible pitch to Anthropic's heaviest Claude Code users.
