Claude Opus 4.7 costs 20–30% more per session

What changed — and who pays

Anthropic’s Opus 4.7 shipped with a new tokenizer, and the short version: you’re getting more tokens for the same money. The company’s migration guide warned of “roughly 1.0 to 1.35x as many tokens” versus 4.6. But it has been reported that independent measurements on real technical content found a 1.47x increase on some samples and about 1.45x on a CLAUDE.md file; a seven-sample real-world set weighed out to roughly 1.325x. Same sticker price, same quota. The context window empties faster. Cached prefixes cost more per turn. Your rate limit trips earlier. Ouch.

How the numbers were gathered

To isolate the tokenizer, it has been reported that the tester used Anthropic’s free token counter (POST /v1/messages/count_tokens) and ran identical content through both tokenizers — no inference, just counting. Results varied by content type: CJK, emoji, and symbols barely moved (around 1.01x), while English and code saw the biggest jumps (1.20–1.47x). Chars-per-token dropped — English from ~4.33 to ~3.60, TypeScript from ~3.66 to ~2.69 — consistent with more, smaller subword pieces. Token counts don’t prove which vocabulary entries changed, though; that’s Anthropic’s black box.

The trade Anthropic is apparently making

Why would they do this? The migration notes pitch “more literal instruction following.” It has been reported that an IFEval spot-check (20 sampled prompts from the 541-prompt suite) found a small but consistent improvement on strict instruction-following tasks, while looser evaluation stayed flat. Partner teams — Notion, Warp, Factory — allegedly reported fewer tool-call errors on long runs. Smaller tokens can force attention to words and punctuation, which plausibly tightens formatting and tool precision. But token-count shifts can’t separate tokenizer effects from changes in model weights or post-training tweaks.

Bottom line: who wins, who pays

For heavy code- and docs-heavy users, this is a real wallet hit — roughly 20–30% more token burn in many workflows. For use cases that prize exact formatting, tool reliability, or “do exactly this” prompts, the improvement might be worth the extra cost. So what should you do? Test your own workload. Run the counter, try a few calibration prompts, and ask: is the added precision worth a pricier session? Tradeoffs. Always tradeoffs.

Sources: claudecodecamp.com, Hacker News