이미지: Ars Technica
Summary
- OpenAI cut GPT-5.6 Luna pricing 80% on both input and output tokens
- Anthropic launched Claude Opus 5 at half the price of its flagship Fable 5 and scrapped a planned Sonnet 5 price hike
- As more companies switch to Chinese models like Moonshot and DeepSeek, SiliconData's token price index has fallen roughly 25% since mid-July
- GPT-5.6 Luna 입력가
- 100만 토큰당 1달러→0.20달러(80% 인하)
- GPT-5.6 Luna 출력가
- 100만 토큰당 6달러→1.20달러(80% 인하)
- Claude Opus 5 가격
- 입력 5달러·출력 25달러(100만 토큰당), Fable 5의 절반
- Fable 5 가격
- 입력 10달러·출력 50달러(100만 토큰당)
- 토큰가격지수 변화
- 실리콘데이터 집계, 7월 중순 이후 약 25% 하락
- 중국 모델 도입 사례
- 도어대시·에어비앤비가 비용 절감 위해 중국산 모델 사용 시작
A price that was $6 per million tokens is now $1.20
OpenAI has cut the output token price of GPT-5.6 Luna from $6 to $1.20 per million tokens. Input tokens dropped from $1 to $0.20. Both are 80% cuts. This matches the pricing OpenAI had already rolled out on Amazon Bedrock on July 30. Anthropic followed the same trend. Its newly released Claude Opus 5 is priced at $5 per million input tokens and $25 per million output tokens — half the price of Anthropic's flagship model, Fable 5. Anthropic also scrapped a Sonnet 5 price increase that had been planned for this week.
How much did prices drop
| Model | Before cut (input/output, per million tokens) | After cut |
|---|---|---|
| GPT-5.6 Luna | $1 / $6 | $0.20 / $1.20 |
| Claude Opus 5 | – | $5 / $25 (half of Fable 5) |
| Fable 5 | $10 / $50 | Unchanged |
According to SiliconData's token pricing index, average pricing across major US lab models has fallen nearly 25% since mid-July. Ars Technica noted this as a sign that closed-model companies, which had been focused primarily on performance competition, are now entering price competition for the first time.
Why cut prices now
Behind this lies weak sales performance for top-tier models. As covered in The Most Expensive AI Model, and Companies Aren't Opening Their Wallets, Fable 5 accounted for only 6% of Anthropic's total token sales volume and 11.4% of revenue in its first month on the market. By contrast, the relatively cheaper GPT-5.6 Sol captured 25% of OpenAI's tokens and 23% of its revenue. This shows that cheaper mid-tier models are outselling expensive flagship models. On top of this, Chinese open models like Moonshot and DeepSeek — which can be freely downloaded and modified — have been rapidly closing the performance gap, adding pressure by reducing the case for paying a premium for expensive US closed models.
Companies are already switching
DoorDash and Airbnb both said they have actually adopted Chinese models to cut their bills. This shift coincides with OpenAI and Anthropic moving some enterprise customers from flat-rate plans to usage-based billing. Usage-based billing charges companies for the actual compute resources consumed, which has a side effect: as agentic workloads increase token consumption, bills for heavy users balloon beyond expectations. OpenAI's introduction of a $125/month premium seat for ChatGPT Business on August 11 came in this same context. Some companies are responding by capping usage outright or testing cheaper alternatives.
So what does this change
This round of cuts signals that US closed-model companies, which had competed purely on performance, are now also wielding price as a weapon. OpenAI and Anthropic are both pursuing IPOs at valuations in the trillions, and investors are demanding evidence that massive AI investment is translating into real returns. In that context, losing customers to cheap Chinese models is a heavy burden. For users, on the other hand, this means more options to get the same performance at a lower price. Whether this round of cuts is a one-off or the start of a sustained trend remains to be seen.
