GPT-5.5 Codex Burns More Tokens

The chart is a useful sanity check, not a verdict: OpenAI says GPT-5.5 is pricier per token than GPT-5.4, but tuned in Codex to get better results with fewer tokens for most users. If this specific workload shows the opposite, the likely explanation is task mix, tool calls, or a longer reasoning trace, not some hidden pricing trick.

• –GPT-5.5 pricing is higher than GPT-5.4 across input, cached input, and output, so it does not win on price-per-token
• –Cached tokens are discounted for both models, but the discount is proportional, so caching alone does not make GPT-5.5 relatively cheaper
• –The meaningful metric is cost per completed task or cost per successful result, not just total tokens burned
• –A chart like this can still be real if GPT-5.5 spends more tokens but avoids retries, improves output quality, or finishes harder tasks more reliably
• –Any Cursor comparison is noisy unless the benchmark controls for prompt length, tool usage, context reuse, and model routing