Developers are facing a sneaky price hike from major AI labs that doesn't show up on the pricing page. By keeping the cost per token static while massively inflating the number of output tokens required for the same task, providers like OpenAI and Anthropic are effectively raising their rates. Recent benchmarks from Playcode reveal that new model versions can cost 3.3x to 3.4x more in practice, despite identical unit pricing.

The GPT Image 2.6 Token Bloat

OpenAI’s release of GPT Image 2.6 highlights the issue with stark clarity. When generating a standard 1024x1024 medium-quality image, GPT Image 2.5 consumed 439 output tokens, costing approximately $0.013. The newer 2.6 model, however, required 1,449 output tokens for the exact same prompt and settings, pushing the cost to $0.044. This 3.3x increase in token usage translates directly to a 3.3x increase in cost per image, even though the per-token price remained unchanged at $30 per million output tokens.

Anthropic’s Fable 5.2 Reality Check

Anthropic is executing a similar strategy with its latest releases. Claude Fable 5.2 launched with lower headline prices than its predecessor, Fable 5.1, at $8/$40 per million tokens compared to $10/$50. However, on a standard MacBook SVG benchmark, Fable 5.2 generated 76,158 output tokens versus 17,933 for Fable 5.1. Despite the cheaper rate, the total cost for the single drawing jumped from $0.90 to $3.05, a 3.4x increase. At higher reasoning levels, Fable 5.2 often exhausted the 128,000 token limit without producing a final answer, costing users $5.12 for zero output.

Hidden Costs in Tokenizer Changes

Beyond raw output inflation, labs are also manipulating how tokens are counted. Anthropic previously adjusted its tokenizer, causing the same text to register as roughly 30% more tokens. This means that even if a model’s efficiency improves, the billing volume increases due to the new counting method. For infrastructure builders, this creates a noisy environment where cost-per-task is no longer a stable metric. You cannot rely on historical cost data when switching model versions, as the underlying token economy shifts without explicit announcement.

Key Takeaways

  • GPT Image 2.6 costs 3.3x more per image than 2.5 due to output token bloat, not price changes.
  • Claude Fable 5.2 costs 3.4x more per task than 5.1 despite lower per-token rates.
  • High-effort modes in new models often fail to complete within token limits, wasting budget.
  • Developers must benchmark output_tokens on specific tasks, not just compare price lists.

The Bottom Line

Stop trusting the price list. If you aren't monitoring output_tokens on your actual workloads, you are paying a hidden premium for every API call.

Developer Action Items

  • Run a curl test on your most common prompts with both old and new model versions.
  • Compare the usage.output_tokens field in the API response, not just the total cost.
  • Set strict token limits for high-reasoning models to prevent runaway costs on failed generations.
  • Consider sticking to stable versions like GPT Image 2.5 for routine tasks where marginal quality gains don't justify a 3x cost increase.