It is now common to hear that AI models keep getting cheaper. We counted. Between 18 August and 17 September 2026 we recorded 127 price changes, of which 96 were cuts — three times the 31 increases.
So far it matches the story everyone tells. Then you look at who changed those prices.
The cuts come from one camp
| Vendor | Price changes |
|---|---|
| DeepSeek | 87 |
| Moonshot AI | 16 |
| Zhipu AI | 11 |
| Alibaba | 7 |
| Meta | 3 |
| 1 | |
| Mistral AI | 1 |
| NVIDIA | 1 |
The top four (DeepSeek, Moonshot AI, Zhipu AI, Alibaba) account for 121 of 127 changes — 95%. DeepSeek alone made 87, or 68%. Some names are missing entirely: OpenAI and Anthropic changed no prices at all in this window.
So "AI is getting cheaper" is only half true. The open-weight camp moves prices often and by large amounts. The commercial API vendors leave prices alone and answer with new models instead. For a buyer the difference matters: with the former your bill can differ from last month, with the latter it stays put unless you switch models.
The biggest moves were increases
Rank the changes by size and the picture shifts again. gemma-4-26b went from $0.04 to $0.09, up 125%. deepseek-v4-pro went from $0.73 to $1.60, up 119%. Then glm-5.3-flash at 114%, glm-5.2 at 106%, deepseek-v4.1-flash at 100%. Every one of the ten largest moves was an increase.
Cuts are far more numerous but much smaller. While cheap models get slightly cheaper, the same camp occasionally doubles a price overnight. If you picked a model for its price, there is no guarantee that price holds.
What one point of benchmark costs
Price alone does not decide anything. Here are the models that carry a benchmark score, with their input price per million tokens.
| Model | Score | Input | Per point |
|---|---|---|---|
| claude-fable-5.1 | 53.4 | $10.00 | $0.187 |
| gpt-6-astra | 52.8 | $10.00 | $0.189 |
| claude-opus-5 | 50.7 | $5.00 | $0.099 |
| gpt-5.6-sol | 47.1 | $2.00 | $0.042 |
| qwen3.8-max-0902 | 45.4 | $2.00 | $0.044 |
| glm-5.3 | 44.9 | $1.40 | $0.031 |
| grok-4.6 | 44.4 | $2.00 | $0.045 |
The leader, claude-fable-5.1, scores 53.4 at $10 — about $0.187 per point. glm-5.3 scores 44.9 at $1.40, or $0.031 per point, six times cheaper. The gap is 8.5 points, 16%. Whether that 16% is worth paying six times more depends on the work: for code generation the last few points can decide the output, for classification or summarisation they often do not.
How to read these numbers
Two caveats. The score is Artificial Analysis' composite of ten evaluations, not a percentage. And many models carry no score yet — including $30 models such as gpt-5.4-pro and gpt-5.5-pro. The table above compares the models that have been scored, not the whole field.
Prices keep moving. You can check the current values on the pages below.
See the price tableSee price change historySee model rankings