Claude Opus 4.7 costs 20–30% more per session
Anthropic's Claude Opus 4.7 quietly inflates token counts by 1.3-1.45x for similar content, effectively increasing per-session costs by 20-30% while maintaining the same sticker price. This subtle shift significantly impacts user budgets and rate limits, prompting HN to question the value of a mere 5 percentage point improvement in strict instruction following. The community widely debates whether this signals diminishing returns on the cost-performance frontier or reflects misaligned incentives from LLM providers.
The Lowdown
A detailed analysis reveals that Anthropic's new Claude Opus 4.7 model employs a tokenizer that substantially increases token consumption compared to its predecessor, Claude 4.6, directly translating into higher operational costs for users.
- Exceeding Estimates: Measurements demonstrate that Claude Opus 4.7 uses 1.2x to 1.47x more tokens for the same content, especially for technical documentation and code, frequently surpassing Anthropic's own stated range of 1.0-1.35x.
- Tokenizer Mechanics: The change suggests a shift to smaller sub-word merges, resulting in more tokens per character for English and code, while CJK content remains largely unaffected.
- Marginal Performance Gain: Despite the significant cost increase, benchmark tests using IFEval showed only a minor +5 percentage point improvement in strict instruction following for 4.7, primarily in specific formatting tasks.
- Real-World Cost Impact: An 80-turn Claude Code session with 4.7 was estimated to cost 20-30% more (approx. $7.86-$8.76) than with 4.6 (approx. $6.65), impacting both monetary expenditure and API rate limits.
- Cache Invalidation: Switching to 4.7 invalidates existing prompt caches, leading to more expensive "cold starts" and increased overall cache storage requirements due to the higher token count.
The author concludes that the modest performance uplift of Claude Opus 4.7 comes at a disproportionate 20-30% increase in per-session cost, forcing users to weigh this trade-off for their specific use cases.
The Gossip
Costly Capabilities
Commenters widely agree with the article's findings, reporting similar 20-30% cost increases for Claude 4.7 with little perceived benefit. Many users express frustration at hitting weekly token limits faster and debate whether the marginal performance gain justifies the substantial financial impact. The discussion delves into the "logarithmic performance/cost frontier" and whether LLMs are reaching a point of rapidly diminishing returns, suggesting Anthropic's higher operating costs might be driving these changes.
Verbose Vexations and Model Regression
A significant theme revolves around the perceived degradation of model quality and increased verbosity in Claude 4.7. Users report the model generating "garbage verbose code" and anecdotally experiencing more hallucinations. Speculation arises that changes to system prompts or an increase in "thinking tokens" might contribute to this verbosity. The debate also touches on the reliability of benchmarks versus anecdotal user experience, with some questioning whether the alleged performance improvements are real or merely "gamed" metrics.
Incentive Imbalance and Model Deprecation
Commenters voice concerns about the long-term strategy of LLM providers, suggesting an inherent incentive to create models that consume more tokens, potentially at the expense of user value. This leads to apprehension about older, more cost-effective models like 4.6 or Sonnet being deprecated, forcing users onto pricier, less efficient versions. The discussion highlights a tension between delivering value to customers and maximizing shareholder returns, questioning the alignment of incentives in the rapidly evolving AI landscape.