AI Gets Dramatically Cheaper: Inference Prices Fall by Up to 80% as Open-Weight Rivals Force a Price War

The price collapse, in one timeline

Something remarkable has happened to the price of artificial intelligence in four months. Between June and September 2026, OpenAI and Anthropic — the two companies whose models power a large share of the world's AI coding assistants, chatbots, and business tools — cut the prices of their flagship large language models by as much as 80%. The scale and speed of the decline surprised even close observers, and it has changed the economics of building with AI in ways that are still working their way through the industry.

The timeline, tracked in detail by former BuzzFeed senior data scientist Max Woolf on his blog, tells the story step by step. All prices below are given in the industry's standard form: input tokens, output tokens, and cache reads, per million tokens. On 26 June 2026, OpenAI announced GPT-5.6 Sol and Luna. Sol, the direct successor to GPT-5.5, launched at 5 dollars per million input tokens, 30 dollars per million output tokens, and 50 cents per million cache reads. Luna, a smaller model for cheaper use cases, launched at 1 dollar, 6 dollars, and 10 cents. Four days later, on 30 June, Anthropic announced Claude Sonnet 5, at launch pricing of 2 dollars, 10 dollars, and 20 cents — pricing initially supposed to be temporary, and expected to rise.

What happened next reversed that expectation. On 24 July, Anthropic announced Claude Opus 5, its long-awaited major-version update, at 5 dollars, 25 dollars, and 50 cents — similar to GPT-5.6 Sol. Then, on 30 July, OpenAI cut the price of GPT-5.6 Luna by 80%, to 20 cents, 1.20 dollars, and 2 cents. Anthropic made Sonnet 5's temporary launch pricing permanent on 10 August. On 22 September, Anthropic announced Claude Opus 5.5 at 4 dollars, 20 dollars, and 20 cents — a 20% cut on input and output prices, and a larger cut on cache reads, working out at a 40–60% total cost reduction against Opus 5. OpenAI responded days later with GPT-6.1 Sol, at half the input and output cost of its predecessor and a cache-read price matching Anthropic's at 5% of input cost, making the family cumulatively about 75% cheaper. As Woolf put it: "Starting from the release of the GPT-5.6 model families in June to now, the cost of using these popular LLMs has cratered, which is not something I would have expected a year ago."

The Wall Street Journal confirmed the same turning point from the industry side, noting that many founders and analysts point to OpenAI's June release of the GPT-5.6 models — designed with cost efficiency in mind — as the catalyst, and that shortly after launch OpenAI lowered Luna by 80% and Terra by 20%. Retool founder and chief executive David Hsu told the Journal that the GPT-5.6 release was "the main catalyst" for why his company now mostly uses OpenAI models, adding: "We want to find the cheapest models because our revenue goes up the cheaper the models are." He estimated he now spends about 20% less using OpenAI's models than Anthropic's.

Why prices fell: competition, efficiency, and perishable capacity

Three forces explain the collapse, and they point to different futures. The first is competition from open-weight models, above all from China. DeepSeek's Flash V4 was reported to earn 38 dollars per graphics processing unit (GPU) per hour at 60 cents per million output tokens, and its successor, DeepSeek Flash V4.1, was listed at 15 cents per million input tokens, 60 cents per million output tokens, and a fraction of a cent for cache reads. When a capable model is available at that price, through dozens of inference providers on routing platforms such as OpenRouter, the frontier labs cannot simply hold their prices and expect developers to stay. OpenRouter's head of insights, Peter Walker, told the Journal that GPT-5.6 "opened up this new lane for OpenAI in a way that the Anthropic family of models doesn't really have" — a cheaper tier competing directly with open-weight alternatives rather than asking customers to pay a frontier premium for every task.

The third force is the economics of GPU capacity itself. As Woolf's analysis notes, GPU capacity is perishable, like airline seats: an idle GPU hour cannot be stored and sold tomorrow. A lab with vast computing capacity therefore has a strong incentive to fill it, even at a lower price, rather than leave it empty while waiting for a customer willing to pay the old rate. That logic pushes prices down fast, because every provider faces the same perishability at the same time. It also raises an uncomfortable question: if prices fall faster than usage grows, revenue can fall even as adoption rises, and enormous capital spending on data centres and chips becomes harder to justify.

The cache-read detail that changes everything for coders

The most important technical detail in this price war is also the easiest to miss, because it concerns a price component casual users never see. When an AI coding assistant works through a long task, it re-reads the conversation history — the code, previous answers, context — on every turn. Those re-reads are called cache reads, and in agentic coding workflows they account for 96–99% of total token spend, according to Woolf's analysis. The price of cache reads, not the headline input or output price, is therefore what actually determines the bill for tools such as Codex and Claude Code.

This is why the cache-read cuts matter disproportionately. A 75% reduction in cache-read pricing translates into a 70–75% drop in effective cost for developers using agentic coding tools, even if input and output prices move less. Both labs understood this, and both moved their cache-read prices to the same point: 5% of input cost. For a developer running an agent for hours, that convergence is worth far more than a headline cut to input tokens that make up only a small share of the bill.

IBM's argument: 90% as good, at 10% of the cost, on your own racks

While OpenAI and Anthropic cut API prices, IBM's chief executive, Arvind Krishna, made a bolder claim on the company's third-quarter earnings call: for most enterprise work, companies no longer need frontier models at all. Krishna told investors that open-weight models now make compelling sense for on-premises enterprise use. In 2024, he said, closed frontier models were plainly superior: on a broad-based benchmark, the top model might score 70 while an open model scored 20 or 30. By late 2025, that gap had shrunk to perhaps 70 against 50. Now, he said, open-weight models deliver "anywhere from 25 percent to 10 percent of the token cost of the frontier models" and are "now 90 percent as effective."

Krishna's pitch was deliberately simple: "It's a simple question for any enterprise. Would you like something that's 10 percent of the cost if it's 90 percent as good as the best, or do you want to pay that 10 times higher cost?" He argued that the vast majority of routine business tasks — summarising an email, examining a contract, answering a customer call, composing an email — can be done just as well with the cheaper answer, and that only some tasks genuinely require the frontier model. He said IBM is seeing booming interest in on-premises AI infrastructure, with most businesses able to build useful infrastructure in just two or three racks — while acknowledging that many will struggle to get it running, given the shortage of engineers who understand GPU clustering, networking, and caching.

That admission matters, because it is the main counter-argument to the on-premises pitch. Running open-weight models yourself is cheaper per token, but not free: it requires hardware, power, cooling, maintenance, model updates, security patching, and scarce expertise. For a large enterprise with steady, high-volume, sensitive workloads — a bank reviewing contracts, an insurer processing claims — those fixed costs can be spread across enough usage to make the maths work, and the control benefits (data never leaves the building, no dependence on an external provider's pricing) may matter as much as price. For a small company with spiky, unpredictable usage, paying per token through an API — now much cheaper than in June — may still be simpler and cheaper. The market is therefore likely to split, with routine, high-volume, sensitive work moving on-premises or to private deployments, and frontier or occasional work staying on the big labs' APIs.

What this means, and what to watch

For developers and startups, the immediate implication is that the cost barrier to building with AI has fallen sharply in a single quarter. Experiments too expensive to run at scale in June are affordable now, and products that needed high prices to cover model costs can charge less, or serve more users for the same revenue. That should mean more AI products, more competition, and more pressure on any business whose only advantage was access to a frontier model everyone can now afford.

For the labs, the implication is harder. A price war driven by open-weight competition, efficiency gains, and perishable capacity is difficult to reverse. If the next generation of models is only modestly better than open-weight alternatives, the premium will be hard to sustain. If it is dramatically better at tasks customers care about — complex reasoning, long-horizon agency, scientific work — the premium may survive. The race is no longer only about who has the best model; it is about who can keep a quality gap large enough to justify a price gap, while costs fall everywhere.

Verifying source: Max Woolf's Blog (minimaxir.com), October 2026; The Register, 22 October 2026; The Wall Street Journal, October 2026; ai0.news daily digest, 26 October 2026.