Anthropic's +35% tokenizer tax on code and JSON

Claude Opus 4.7 ships with a quietly retrained tokenizer that splits structural payloads — JSON, Python, TypeScript, regex — into noticeably more tokens than Opus 4.6 did on the same input. Our regression fixtures across 200 mixed prompts show a consistent +35% token count when the body contains structural markers ({, }, def, function, type annotations).

At the sticker price of $5/Mtok input and $25/Mtok output, the underlying rate looks competitive with GPT-5.5 — but on JSON-heavy agentic workloads, the effective cost is closer to $6.75/Mtok input and $33.75/Mtok output. That's a hidden 20–30% premium versus what a buyer would estimate from the price page alone.

The TokenFloor calculator detects this automatically: if the input contains structural markers and the model is in the Claude 4 family, the cost row is shown in EU gold with a tooltip explaining the multiplier. Anthropic has not published a changelog entry for the tokenizer change. For now: budget +35% on code-heavy Claude 4 traffic, or route the same workload to a sibling model.