The Reversal of Token Costs

Looking only at $/MTok can easily lead to misreading costs. Even with the same input, model-specific token decomposition methods differ, causing the actual billed token count to vary significantly.

• When comparing identical content using official token counting APIs, token count differences reached up to 2.65x+.
• Results varied by content type: text, JSON, YAML, tool definitions all had different tokenization efficiency.
• In the benchmark tests, OpenAI showed the most efficient tokenizer.

When price and token efficiency are considered together, the ranking changes.

• gpt-5.4: baseline
• gemini-3.1-pro-preview: cheaper for text/structured data, but 46% more expensive than OpenAI for tool definitions
• claude-sonnet-4-6 / claude-opus-4-6: rises to 1.4x~2.5x compared to OpenAI depending on input type
• claude-opus-4-7: 5.3x the cost of gpt-5.4 in tool-heavy workloads

The key point is that comparing list prices alone misses the actual cost. When prompt caching, long-context, output tokens, and thinking tokens are included, the difference can grow even larger, so it needs to be measured directly on actual workloads.