OpenAI and Anthropic cut LLM prices by 80%

OpenAI and Anthropic have cut the prices of several of their language models by as much as 80% within a few months, according to a timeline at Max Woolf’s blog. The moves matter because they are large, rapid, and tied to competitive pressure, caching, and the economics of AI agents. They also arrive while both companies are reportedly preparing for IPOs.

What happened

According to the source, OpenAI released the GPT-5.6 family in June 2026. GPT-5.6 Sol was the direct successor to GPT-5.5 and was priced at $5 per input million tokens, $30 per output million, and $0.50 per cache read million. GPT-5.6 Luna was a smaller model for cheaper use cases at $1, $6, and $0.10. Anthropic announced Claude Sonnet 5 in June at $2, $10, and $0.20, which it initially described as temporary. In July, OpenAI announced a cut of about 80% for GPT-5.6 Luna, bringing it to $0.20, $1.20, and $0.02. Anthropic announced Claude Opus 5 at $5, $25, and $0.50. In August, OpenAI cut GPT-5.6 Sol by about 20% to $4, $20, and $0.40 for three months. Anthropic made the Sonnet 5 price cut permanent. Later, OpenAI released GPT-6 Sol at $2, $10, and $0.20, and GPT-6 Luna at $0.10, $0.50, and $0.01. Anthropic announced Opus 5.5 at $4, $20, and $0.20, saying it required less compute than Opus 5. The source describes a sequence of cuts that lowered the cost of using popular LLMs. The source also notes that all pricing numbers were provided as input MTok / output MTok / cache read MTok, and that a Terra model received less attention.

Why did the cuts happen

The source frames the price changes as part of a larger pattern. It says there is a clear tit-for-tat between Anthropic and OpenAI, and that the cuts are reasonable to assume were reactions to one another. But it also asks what else could explain an 80% cut. The author points to behavioral economics and competitor pricing pressure. Since late 2025, agentic LLMs have driven hypergrowth, and models have become more of a commodity. Developers can switch models with a toggle in a harness or change an API endpoint through OpenRouter. OpenAI and Anthropic therefore have less pricing power. The source also notes competition from open-weight models, including DeepSeek and Moonshot AI. Those comparisons may have pushed the larger labs to cut prices.

The timing is also notable because both OpenAI and Anthropic are reportedly close to IPO. The source says leaked financials show both are losing billions due to R&D and scaling, though the reports do not account for the hyperexpansion of agentic LLM use. Investors will scrutinize negative profits and demand plans to return to growth. The source questions why the primary revenue source would be cut by up to 80%, but it notes that inference profit margins may still be healthy. The author writes that the price cuts have affected how they spend on LLM agents, but not always in the way OpenAI or Anthropic might intend.

What does it mean in practice

For agent workflows, the source says cache read costs are the largest part of token spending. In coding agents, cache reads accounted for 96% to 99% of token expenditure, and cache hits were around 90% to 100%. Because of that, a 75% cut to cache read cost can mean a 70% to 75% reduction in overall cost. Anthropic’s Fable 5.1 cut cache read costs by 75% while keeping the same output price. OpenAI’s later moves, including GPT-6.1 Sol cutting cache price in half, also reduced cumulative costs. The source cites Vercel’s AI gateway data showing cost-per-token falling to 43.3% of its April 2026 level. Even so, the source says aggregate spending may increase because models are improving faster than they are getting cheaper.

The practical implication for developers is that cheaper models may now handle more routine tasks. The source says the cuts applied to models used in agentic systems such as Codex and Claude Code. It also says the quality did not appear to be cut in a way that would make the cheaper models useless. For researchers, the lower prices could make it easier to run more experiments or compare models. For creators building AI tools, the shift may affect cost projections and the choice of model for different tasks.

How can readers use this with Grafy

For developers and researchers using AI tools, the story connects to work done in AI coding and multi-model chat. The source notes that agentic harnesses were affected by the price changes. In Grafy, the relevant apps appear to be Grafy Chat and Grafy Code. Grafy Chat lets users chat with multiple AI models side by side, which could help compare a low-cost model against a stronger one. Grafy Code is an AI coding studio that installs packages and runs code, which may support testing agent workflows. The Grafy News Desk’s practical take is to test whether a cheaper model can handle routine coding, documentation, or research tasks, and to keep a more capable model for complex work.

Frequently asked questions

Are the OpenAI and Anthropic price cuts real?

Yes. According to the source, OpenAI cut GPT-5.6 Luna by about 80% in July 2026, and later reduced other model prices. Anthropic made a Sonnet 5 launch discount permanent and cut Opus 5.5 pricing. The author reports these as actual announcements, not speculation.

How much cheaper did the models become?

The largest cited cut was OpenAI’s 80% reduction for GPT-5.6 Luna. Anthropic’s Opus 5.5 was described as having a total cost reduction of 40% to 60% compared with Opus 5. Fable 5.1 cut cache read costs by 75%. GPT-6 Sol and Luna were also priced lower than earlier versions.

Why did the AI labs cut prices?

The source points to competition, model commoditization, and agent adoption. OpenAI and Anthropic may have cut prices to respond to rivals, including open-weight and international model providers. The timing also overlapped with reported IPO preparations, though the source says the reason is not fully settled.

---

Source: Hacker News — front page — OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?