A lower price per token does not guarantee a lower bill. The same hypothetical workload (1 million input tokens and 100 thousand output tokens) costs between about $0.42 and $6.00 depending on the model. With 90% of input hitting the cache, it costs between $0.16 and $2.58.
These numbers are our arithmetic from list prices, not a performance test. They assume equal token counts, not equal quality, and leave out cache-write fees, tools, taxes and retries.
The short answer
- Same workload, wide range: from $0.42 to $6.00 with no cache hits, and $0.16 to $2.58 with 90% cache hits.
- Caching and output price matter most. The example shows their effect, not a ranking of quality.
- Prices fell, spending rose: analyst Max Woolf’s 25 October post argues that API prices from OpenAI and Anthropic fell sharply between June and September 2026 while total spending still rose.
- Not subscription fees: these are API list prices, not ChatGPT or Claude subscription fees.
What do the price pages say?
Prices are per 1M tokens, in USD, from the official pages read on 26 October 2026.
| Model (source) | Input (cache miss) | Cache hit | Output |
|---|---|---|---|
| Claude Opus 5.5 (Anthropic, 22 Sep 2026) | $4 | $0.20 | $20 |
| Kimi K3.1 Pro (Kimi API page) | $2 | $0.20 | $10 |
| grok-4.7, prompts under 200k (xAI docs) | $2 | $0.50 | $6 |
| DeepSeek Flash, peak hours (DeepSeek docs) | $0.30 | $0.006 | $1.20 |
| DeepSeek Flash, off-peak | $0.15 | $0.003 | $0.60 |
Anthropic’s page also lists cache writes at $5 per million tokens for Opus 5.5, which the examples below leave out. OpenAI’s GPT-6 prices are not in this table because its pricing pages could not be read.
Same token counts, how different are the bills?
Assume 1M input tokens and 100K output tokens in total, with the same counts for every model.
| Model | No cache hits | 90% of input hits the cache |
|---|---|---|
| Claude Opus 5.5 | $6.00 | $2.58 |
| Kimi K3.1 Pro | $3.00 | $1.38 |
| grok-4.7 | $2.60 | $1.25 |
| DeepSeek Flash (peak) | $0.42 | $0.16 |
| DeepSeek Flash (off-peak) | $0.21 | $0.08 |
For example, Opus 5.5 with 90% cache hits is 0.1M x $4 + 0.9M x $0.20 + 0.1M x $20 = $0.40 + $0.18 + $2.00 = $2.58.
Notice what caching does to Opus: $6.00 drops to $2.58, and $2.00 of that is output, which caching cannot touch. The takeaway is the effect of caching and output price, not a ranking of quality. The cheapest line here may need more retries or produce weaker results for your task.
Why can the bill rise while token prices fall?
- Tokens used per task differ. Artificial Analysis reported that Kimi K3.1 Pro used 55K output tokens per Intelligence Index task against 48K for Kimi K3. A 33% lower token price therefore left the cost per task about the same ($1.98). See the Kimi K3.1 Pro analysis.
- Caching dominates agent workloads. Woolf says cache reads make up 96% to 99% of token spend in agentic coding, which is why price cuts on cache reads matter more than headline input prices. That is his claim from his own analysis, not an industry measurement.
- More usage. Woolf cites Vercel data showing average cost per token at 43.3% of its April 2026 level, while aggregate spending on its gateway kept rising because usage grew. That is one provider’s customers, not the whole market.
What does Woolf argue about the price cuts?
Woolf lists cuts he tracked between June and September 2026, including an 80% cut to one OpenAI model in July. He says some Anthropic and OpenAI price drops were up to 80%. Do not read the 80% figure as a cut that applies to every model or plan.
His explanations (commoditization, pressure from cheap open-weight models and efficiency gains) are his interpretation. He says he lacks firm data on inference margins or subscription profitability.
He also notes that subscription allowances can be cut before list prices: he cites a 50% cut to a Codex usage allowance.
How do you compare costs for your own workload?
We suggest these steps, which are ours and not any provider’s rule.
- Measure your real input, cached input and output tokens on a sample of your tasks.
- Price them with each provider’s page, including cache write fees and any long-prompt tier (xAI charges more above 200k prompt tokens).
- Count retries and the share of tasks that finish, as in why a cheaper model can cost more per finished task.
- Check peak and off-peak pricing where it exists, as in DeepSeek API pricing.
For related pricing pages see Grok API models and pricing and Claude plans and pricing. For reading comparison tables, see five ways AI benchmark tables mislead.
What is still open?
OpenAI’s current GPT-6 prices are missing, so check its pricing page directly. The 90% cache case is an assumption, so replace it with a measurement from your own logs.
Questions and answers
Is a model with a lower token price always cheaper?
No. Cost per task also depends on how many tokens the model uses and on retries, as Artificial Analysis’s Kimi figures show.
Do these prices apply to ChatGPT or Claude subscriptions?
No. They are API list prices per million tokens.
Is the 80% price cut true for every model?
No. Woolf describes cuts of up to 80% on some models between June and September 2026, not on all of them.
Which line in the table is cheapest?
DeepSeek Flash off-peak, at $0.21 with no cache hits and $0.08 with 90% hits. It may still need more retries, so test it on your tasks.
Terms in plain language
- Token: a small piece of text that AI services count and charge for.
- Cache hit: input the service has already seen and can reuse at a lower price.
- Cache write: the fee for storing input so it can be reused later.
- List price: the published price before any discount.
Sources
- Anthropic: Claude Opus 5.5 pricing (read 26 October 2026)
- Kimi API pricing
- xAI Docs: models and pricing
- DeepSeek API Docs: models and pricing
- Max Woolf: analysis of falling LLM prices (25 October 2026)
Facts read on 26 October 2026.
Part of the guide: AI Model Pricing and API Cost Guides
