What happened?
OpenAI and Anthropic have slashed prices on their mid-tier models to retain cost-sensitive customers. OpenAI cut the price of GPT-5.6 Luna, which it markets as its "fastest and most affordable model," by 80%; the input token price dropped from $1 to $0.20 per million, and the output token price fell from $6 to $1.20 per million.
Anthropic, meanwhile, launched Claude Opus 5 at half the price of its most capable model, Fable 5. Opus 5 is priced at $5 per million input tokens and $25 per million output tokens. The company also canceled a planned price increase for Sonnet 5 that was set to take effect in September.
According to Silicon Data's token price index, prices for models from leading US labs have fallen by about a quarter since mid-July.
Why does it matter?
The price cuts stem from mounting competitive pressure as Chinese developers like DeepSeek and Moonshot close the performance gap. "Open" Chinese models, which can be freely downloaded and modified, offer a price advantage over the closed systems of US companies.
Companies such as DoorDash and Airbnb have said they've started using Chinese-origin models to cut their bills. In addition, Anthropic and OpenAI are shifting some enterprise customers from fixed subscriptions to usage-based billing, which is intensifying cost pressure on companies.
These developments come as OpenAI and Anthropic are planning IPOs at trillion-dollar valuations. Investors are looking for evidence that the industry's massive AI spending will pay off.
How do the prices compare?
Upfront token prices don't offer a direct comparison between models, since more capable models can complete a task with fewer tokens or fewer attempts. Models can also run at different "effort" settings, which affects both performance and cost.
| Model | Input price (per million tokens) | Output price (per million tokens) |
|---|---|---|
| GPT-5.6 Luna (new) | $0.20 | $1.20 |
| GPT-5.6 Luna (old) | $1 | $6 |
| Claude Opus 5 | $5 | $25 |
According to Artificial Analysis benchmarks, Opus 5 at "medium" effort delivered similar performance and per-task cost to Moonshot Kimi K3 at "maximum" effort. GPT-5.6 Luna at "maximum" effort showed performance close to DeepSeek V4 Flash, but cost nearly twice as much per task.
What's next?
- Anthropic and OpenAI are trying to maintain their position at the top end by keeping prices for their flagship models steady or raising them.
- According to Mantas Lukauskas, AI technical lead at Hostinger, this will be the "first real test" of whether US labs can maintain pricing for their top-tier products.
- A source close to Anthropic said Opus 5 being priced lower than Fable 5 is a natural part of the company's "model family" structure and is not related to competitors.