Anthropic’s pricing page, as of October 14, 2026, lists four primary models with distinct cost structures that defy simple per-token comparisons. While Claude Haiku 5.5 appears cheapest at $0.10 per million input tokens, a critical tokenizer update in Claude 4.7 and later models means the same text now generates approximately 30% more tokens. This technical shift fundamentally alters the value proposition of moving from legacy models like Sonnet 4.6 to newer iterations, where nominal price drops are partially offset by increased token counts.

The Sonnet 5 Price Lock and Model Tiers

Anthropic has made a significant strategic move by making the introductory pricing for Claude Sonnet 5 permanent. Originally launched at $2 for input and $10 for output per million tokens, these rates will remain standard rather than rising to the planned $3 and $15 on September 1. This decision stabilizes costs for developers relying on the mid-tier model, which Anthropic describes as offering the best combination of speed and intelligence. Meanwhile, the premium Claude Fable 5.1 remains at $10 input and $50 output, targeting demanding reasoning and long-horizon agentic work, while Claude Opus 5.5 sits at $4 and $20 for long-running coding tasks.

The 30% Tokenizer Tax

The most deceptive aspect of the current pricing structure is the change in tokenization. Anthropic’s documentation confirms that models from Claude 4.7 onward use a newer tokenizer that produces about 30% more tokens for identical text compared to older models like Sonnet 4.6 or Haiku 4.5. For example, switching from Sonnet 4.6 to Sonnet 5.5 offers a theoretical 33% price reduction per token. However, when adjusted for the 30% increase in token count, the actual cost saving per task is only about 13%. This 'tokenizer tax' means that per-token prices overstate the savings, forcing developers to benchmark costs per task rather than per token to avoid budget surprises.

Haiku’s Long-Prompt Penalty

Claude Haiku 5.5 introduces a sharp pricing cliff for prompts exceeding 100,000 tokens. While prompts under this threshold cost $0.10 per million input tokens, requests over 100,000 tokens jump to $0.50 per millionβ€”a 5x increase. This penalty applies to the entire request, not just the overflow, making Haiku unexpectedly expensive for long-context tasks compared to Sonnet 5.5, which maintains flat pricing up to its 1 million token context window. For instance, a 101,000-token prompt on Haiku costs roughly five times as much for only 2% more text than a 99,000-token prompt.

Key Takeaways

  • Sonnet 5 pricing is now permanent at $2/$10, canceling the planned September increase.
  • Newer Claude models use a tokenizer that generates ~30% more tokens, reducing the effective savings of price drops.
  • Haiku 5.5 costs 5x more per token for prompts over 100,000 tokens, unlike Sonnet and Opus which have flat rates up to 1M.
  • Claude Mythos models are distinct from Fable, available only to verified organizations despite identical listed prices.
  • Batch API discounts apply to all models, halving input and output costs for non-time-sensitive workloads.

The Bottom Line

Developers must stop comparing Claude models on a per-token basis. The 30% tokenizer inflation effectively negates nearly all advertised price cuts for mid-tier migrations, and Haiku's long-context penalty turns its 'cheap' label into a trap for RAG-heavy applications. Benchmark your actual cost-per-task using the new tokenizer before committing to a model migration.