The real prices of frontier models. Tokens * Price, right?
The true cost of using Large Language Models isn't just about the sticker price per token; it's about how efficiently a model's tokenizer processes your content. This deep dive reveals that Anthropic's new tokenizers can inflate token counts (and thus costs) by 30-73% compared to OpenAI for the same code, a hidden fee that has developers on Hacker News verifying their bills and rethinking their LLM choices.
The Lowdown
The article dissects the often-overlooked factor of tokenization in LLM pricing, arguing that the advertised "dollars per million tokens" can be misleading. A token is not a fixed unit, and different models' tokenizers process the same text into varying numbers of tokens, directly impacting the final bill.
Key findings include:
- Silent Price Hike: Anthropic's new tokenizer, used in models like Opus 4.8 and Sonnet 5, creates approximately 30% more tokens for the same code than its predecessor, effectively increasing costs at the same list price.
- Cross-Vendor Disparity: For code, Claude's new tokenizer generates 1.5x to 1.73x more tokens than OpenAI's
o200ktokenizer, with TypeScript being the most affected. GPT's tokenizer is particularly efficient for web-related languages. - Effective Pricing: When actual token counts are factored in, models like Claude Opus 4.8 and Fable 5 become significantly more expensive than their sticker prices suggest, while Gemini 3 Flash remains cost-effective despite slight tokenizer inefficiency.
- Content Matters: The degree of token inflation varies by content type; code and English prose are significantly impacted, while Chinese text shows different patterns.
The authors recommend comparing models based on their performance on your specific content, treating tokenizer changes as price adjustments, and measuring costs in "dollars per task" rather than just "dollars per token." This approach reveals the actual expenditure developers face when building with LLMs.
The Gossip
Anthropic's Token Troubles
Many commenters confirm the article's findings, sharing their own experiences of Anthropic's tokenizers being significantly less efficient than OpenAI's. There's a consensus that Claude models consume substantially more tokens for equivalent content, especially codebases, leading to higher costs. Some users label this as "shrinkflation" by Anthropic, while others simply note the practical implications for their budgets.
Beyond the Token Count
A significant portion of the discussion expands on what truly constitutes the 'real price' of LLM usage beyond simple input token counts. Users highlight other critical factors like model 'verbosity' (how much output it generates), KV cache writes/reads (especially for large contexts and agentic workflows), and the overall 'dollars per task' performance. The sentiment is that while tokenization matters, model behavior and the efficiency of agentic loops often have a much greater impact on total spend.
Fable's Fantastic, but Fiscally Formidable
Commenters discuss the performance and pricing of specific Anthropic models, particularly Fable 5. While acknowledging Fable's superior capabilities for complex or creative tasks, its effective cost is often deemed prohibitive for routine use. Users frequently opt for cheaper alternatives like Opus 4.8 or Sonnet 5, despite their token inefficiencies, finding Fable only justifiable for very specific, challenging problems.
Critique of Content Craftsmanship
Some users express skepticism about the article's authorship, speculating that it might be LLM-generated due to its writing style, which some describe as 'unpalatable' or having 'Claude-isms.' This leads to concerns about the effort put into fact-checking and the overall trustworthiness of the content, highlighting a broader sensitivity on Hacker News to AI-generated text.