Cheap AI models can still wreck your margins
Three matching token prices can produce very different bills once you count retries, caching and correction time.
The dots
- Anthropic’s Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens, unchanged from Sonnet 5. Yet Anthropic reports up to 30% lower costs per task because it uses fewer tokens. Source.
- OpenAI prices GPT-6.1 Sol at the same $2/$10, but cached input costs $0.10 per million tokens, half Sonnet’s $0.20. That difference can add up when an agent repeatedly uses the same documents or instructions. Source.
- Google announced Gemini 4 Argon at an introductory $2/$10, with a maximum output allowance of one million tokens, up from 64,000. Access initially goes to selected cyber defenders; most builders still have to wait. Source.
Our take
Picking a model from its price table is getting harder. Three matching rates can produce different bills once you count how long each model works, what gets cached and how often it needs another attempt.
There’s a revealing detail in Anthropic’s announcement: at higher reasoning settings, Sonnet can approach Opus’s performance at a similar cost. Turning every setting up can eat the saving. Anthropic’s analysis.
We think builders will increasingly reserve expensive models for difficult cases. That only pays if the cheaper model finishes routine work reliably. Someone still gets the bill for attempt number four.
Watch
Total cost per accepted result, including failed attempts, tool charges and human correction time. Compare models on the same tasks, with the same standard for a usable answer.