Guides · dev · claude code cost
What a coding session costs: flash models, flagships, and the cache assumption
Token prices flatter models that are cheap to read and expensive to write, and they flatter list prices that are not the host you will call. A session cost is still a model of a session. Ours is written down so you can throw it out.
The mix: 3,000,000 input tokens, 90 percent of them billed at the cache-read rate when we have one, 60,000 output tokens. Cache writes are not included. If your client does not send a stable prefix, set the cache share to zero in the calculator.
OpenRouter list prices
| Model | Session | vs Opus | vs Sol | Cache price |
|---|---|---|---|---|
| qwen3.8-flash | $0.116 | 96% cheaper | 92% cheaper | published |
| glm-5.3-flash | $0.096 | 97% cheaper | 94% cheaper | published |
| deepseek-v4.1-flash | $0.066 | 98% cheaper | 95% cheaper | published |
| minimax-m3 | $0.324 | 89% cheaper | 78% cheaper | published |
| glm-5.3 | $0.522 | 82% cheaper | 64% cheaper | published |
| kimi-k3 | $1.35 | 54% cheaper | 8% cheaper | published |
| Anthropic: Claude Opus 5.5 | $2.94 | — | — | published |
| OpenAI: GPT-6.1 Sol | $1.47 | — | — | published |
The percentages in this table compare session cost. The price table's percentages compare blended token price. They answer different questions.
The GLM-5.3 and Kimi bars here are the trap. They use the card, not the lab.
Lab rates where we have them
| Model | Session | vs Opus | vs Sol | Cache price |
|---|---|---|---|---|
| GLM-5.3 | $1.39 | 53% cheaper | 6% cheaper | published |
| GLM-5.3-Flash | $0.156 | 95% cheaper | 89% cheaper | published |
| DeepSeek-V4.1-Flash | $0.178 | 94% cheaper | 88% cheaper | published |
| DeepSeek-V4-Pro-0813 | $0.752 | 74% cheaper | 49% cheaper | published |
| Kimi K3 | $2.61 | 11% cheaper | 78% dearer | published |
| Anthropic: Claude Opus 5.5 | $2.94 | — | — | published |
| OpenAI: GPT-6.1 Sol | $1.47 | — | — | published |
The percentages in this table compare session cost. The price table's percentages compare blended token price. They answer different questions.
DeepSeek rows are the peak rate. Off-peak is half, on the schedule in the DeepSeek hub. Kimi’s official output rate is why that bar jumps. GLM Flash stays in the cheap group on the official page, which is the interesting part: the discount survives contact with the vendor. The flagship discount does not.
How to use this without lying to yourself
Export one real trace. Count uncached input, cached input, and output. Put those three numbers next to this mix. If output is 400,000 tokens because the agent narrates, every chart on this page is too kind to Kimi and too kind to any model that rambles. If the cache hit rate is 10 percent, the flash models are still cheap and the session is still mostly input price.
The GLM plan is the subscription version of the same question. Credits are not these tokens. Do not convert one into the other without a rate Z.AI did not publish.
Related: price map, calculator.
Questions
Is this my Claude Code bill?
No. The mix is 3 million input tokens, 90 percent cache hits, and 60 thousand output tokens. Change it in the calculator if your traces look different.
Why is the official GLM bar different from the OpenRouter bar?
Because the OpenRouter list price for GLM-5.3 is not Z.AI's price. The second chart uses the docs.
Sources
- OpenRouter models API, accessed 2026-10-16
- Z.AI pricing, accessed 2026-10-16
- DeepSeek API pricing, accessed 2026-10-16
- Kimi context caching, accessed 2026-10-16