The Hidden 30% Tax in Claude 4.7

Claude 4.7's new tokenizer increases token counts for technical content by up to 1.47x, significantly higher than the average range suggested by Anthropic. This leads to a 20-30% increase in per-session costs and faster rate limit exhaustion for users working with code and English prose. The author concludes that this increased cost is a trade-off for slightly better performance in strict instruction following and tool-call accuracy.

Key Points

• Claude 4.7's tokenizer increases token counts for English and code by 1.2x to 1.47x, often exceeding Anthropic's documented range.
• The effective cost of a long Claude Code session increases by 20-30% due to higher token volume in conversation history and prefixes.
• Technical content like CLAUDE.md and stack traces are more heavily impacted than prose or non-Latin scripts like Chinese or Japanese.
• The higher token count is a trade-off for better precision, with testing showing a measurable improvement in strict instruction following.
• Users on Max plans will hit rate limits faster because the model consumes the context window more quickly for the same amount of text.

Sentiment

The community is skeptical but pragmatic. Most accept the article's factual findings about the tokenizer cost increase, but opinion splits sharply on whether it matters. Cost-conscious developers and AI skeptics see it as part of a concerning pattern of rising costs and declining transparency, while heavy AI users argue the productivity gains still massively outweigh the costs. There is notable enthusiasm for open-source local models as a hedge against proprietary pricing, though most acknowledge these aren't yet at frontier quality. Overall sentiment leans slightly negative toward Anthropic's pricing trajectory while remaining positive about AI coding tools in general.

In Agreement

• The tokenizer change is a real hidden cost increase that Anthropic has not been transparent about, with measured increases exceeding their stated range
• Anthropic is likely raising effective prices because their operating costs are unsustainable as they approach IPO, and current subscription pricing is heavily subsidized
• LLM performance improvements are hitting diminishing returns — like an 8K to 16K display upgrade, the cost premium is not justified by perceptible gains for most users
• The enshittification pattern is emerging: hook users with cheap supply, then gradually increase prices while degrading the product
• Cost unpredictability is a serious business risk — unlike human labor with predictable salaries, AI providers can raise prices overnight with no recourse

Opposed

• Token costs are still negligible as a business expense compared to developer salaries — even at current prices, the productivity gains far outweigh the costs
• Per-intelligence-unit costs are actually dropping dramatically across model generations, so the overall trend is toward cheaper AI, not more expensive
• Real-world results justify the cost: teams report completing months-long projects in weeks using AI coding agents with proper oversight and refactoring processes
• The focus should be on model quality and developer time savings, not marginal token cost differences — right-sizing the model to the task matters more than raw price per token
• Open-source local models are still meaningfully behind frontier models for complex tasks, and the gap matters for production-quality work