Why OpenAI and Anthropic slashed LLM API pricing

How far OpenAI and Anthropic cut prices since June

API providers bill by the token, a chunk of text a few characters long, and quote prices per million tokens. Input (the text you send) and output (the text the model writes) have separate rates, and output costs more.

Sol's input price went from $5 to $2 and Opus's from $5 to $4 within three months, and Luna lost 80% of its price in a single step. Woolf sees a tit-for-tat in the dates, with each lab answering the other. He also points out that nobody drops a price by 80% without some outside pressure, and he names three forces behind the cuts.

OpenAI cut GPT-5.6 Luna by 80% in July

About a month after launch, OpenAI dropped Luna to a fifth of its original input and output price on July 30. Luna, the small model in the GPT-5.6 family, had debuted on June 26 at $1 per million input tokens and $6 per million output tokens. After the cut it cost $0.20 and $1.20.

Anthropic moved in the same weeks. Claude Sonnet 5 launched on June 30 with a $2 / $10 price that Anthropic first called temporary, then made permanent on August 10. On August 21, OpenAI took about 20% off GPT-5.6 Sol for three months, bringing it to $4 / $20.

GPT-6 Sol and Claude Opus 5.5 launched on the same day

On September 22, GPT-6 Sol arrived at half the input and output price of the already discounted GPT-5.6 Sol. It costs $2 per million input tokens and $10 per million output tokens. OpenAI shipped GPT-6 Luna the same day at $0.10 / $0.50, also half the previous Luna.

Anthropic launched Claude Opus 5.5 on that date too, at $4 / $20, which is 20% below Opus 5's $5 / $25. A week later GPT-6.1 Sol halved the cache read price from $0.20 to $0.10. By Woolf's count, the Sol family now costs 75% less than it did before the cuts began.

OpenAI API pricing and Claude API pricing, before and after

Output price per million tokens, before and after the cuts

GPT Sol

GPT-5.6 Sol's output price = 100.

$30GPT-5.6 Sol

$10GPT-6.1 Sol

0102030$

GPT Luna

June GPT-5.6 Luna's output price = 100.

$6GPT-5.6 Luna (June)

$0.50GPT-6 Luna

0246$

Claude Opus

Opus 5's output price = 100.

$25Opus 5

$20Opus 5.5

01020$

Bar height uses each family's pre-cut output price as 100. Source: Max Woolf's blog.

Cache read prices fell along with input and output prices, so real bills for agent workloads dropped further than the headline rates. Sol's output rate is a third of what it was in June, and Luna's is under a tenth. Opus moved less on the headline, from $25 to $20, but Woolf puts its total cost 40% to 60% lower once the cheaper cache reads are counted.

Cost per token fell to 43.3% of April's level

Every model in the table now charges less on all three rates than it did in June or July: Sol's output fell from $30 to $10, and Opus's cache reads from $0.50 to $0.20. Vercel runs a gateway that routes requests to many AI models, and its data puts the average cost per token at 43.3% of the April 2026 level. The table lists the old and new prices for the three families, per million tokens.

Cheaper tokens have not shrunk total spending. Vercel's aggregate spend kept climbing, and Woolf's explanation is that the models are improving faster than they are getting cheaper.

Cache reads drive what coding agents actually cost

Because 96% to 99% of a coding agent's token spend goes to cache reads, a 75% cut there lowers the cost of using a model by 70% to 75%. An agent resends the whole conversation on every turn, and most of that text comes back from the cache rather than as fresh input. Cache hit rates run at roughly 90% to 100%.

Fable 5.1, launched on September 1, shows why this matters. It kept Fable 5's $10 / $50 input and output prices but cut cache reads from $1 to $0.25. The list price looks unchanged, yet an agent running on it costs far less.

Reason 1: switching models takes one setting

Woolf argues that users can move to another lab's model with almost no effort, which leaves OpenAI and Anthropic little room to hold prices up. He tests new models in agents as they come out and finds them hard to tell apart beyond subjective impressions. When two models feel about the same, the cheaper one wins the traffic.

OpenCode and OpenRouter make models interchangeable

A toggle in OpenCode or Zed Agent, or a different model string on OpenRouter, is all it takes to try another company's model. OpenRouter is a single API that forwards each request to whichever provider hosts the chosen model.

Woolf dates the shift to November 2025, when agentic coding took off. Since then, he writes, models have behaved more like a commodity: you can point the same codebase at a new model in seconds. That leaves the labs with little moat, meaning little that keeps a customer from walking.

Google and Meta are not the only alternatives

Smaller labs such as DeepSeek and Moonshot AI also ship models good enough to replace OpenAI's and Anthropic's. Woolf counts the many rivals beyond the big tech companies as a source of steady pricing pressure. DeepSeek V4 Flash and Kimi K3.1 Pro are open-weight models, so companies other than their makers can host them and sell API access.

Reason 2: open-weight models set a price floor

Model quality on the Artificial Analysis index

Bar height uses 60 as 100. Blue bars are the newer open-weight models. Source: Max Woolf's blog.

58Claude Opus 5.5

55Kimi K3.1 Pro

45DeepSeek V4.1 Flash 1017

24DeepSeek V4 Flash

02040

Because anyone can host open-weight models, dozens of providers sell the same model and push its price toward cost. The more hosts a model has, the harder it is for any of them, or for a closed-model lab competing with it, to charge much above the cost of running it.

Woolf links the Luna cut to DeepSeek V4 Flash

DeepSeek V4 Flash launched in April at $0.14 input and $0.28 output, far below Luna before its cut. It drew wide attention, and at the time OpenAI had no small model aimed at simple tasks (related: DeepSeek V4 Flash 0731: Benchmarks and MIT License).

DeepSeek also priced cache reads at 2% of input, against 5% to 10% at OpenAI and Anthropic, which lowers the real cost further. Woolf suspects the July Luna cut was aimed squarely at small open-weight models like this one. He reads OpenAI's own price-performance chart the same way: without the 80% cut, Luna would have lost to DeepSeek V4 Pro.

29 providers serve DeepSeek V4.1 Flash on OpenRouter

The number of OpenRouter providers for DeepSeek's Flash model grew from 13 in May to 29 on October 24, and Cloudflare and DeepInfra come within 10% of DeepSeek's own effective cost. Any host can list a model on OpenRouter, and developers can ask for the cheapest one in the request.

The newer V4.1 Flash 1017 scores 45 on the Artificial Analysis index, up from 24 for V4 Flash. Bought from DeepSeek, it costs $0.15 / $0.60 / $0.003, roughly the old input and cache prices with output about twice as high.

Kimi K3.1 Pro nears Opus 5.5 at half the price

Kimi K3.1 Pro scores 55 on the Artificial Analysis index, three points behind Opus 5.5. It costs $2 / $10 / $0.20, half of Opus 5.5's input and output price with the same cache read rate (related: Kimi K3.1 Pro benchmarks, pricing and open weights).

The K3.1 Pro weights are now on Hugging Face. Its predecessor, Kimi K3, already runs on more than 20 providers, and Woolf expects most of them to add K3.1 Pro at about the same price. Moonshot itself priced K3.1 Pro below K3's $3 / $15 / $0.30.

Reason 3: newer models cost less to serve

Woolf also accepts a less cynical explanation: the infrastructure behind LLMs improved, so serving them got cheaper. The first two reasons are pressure from outside. This one is a change inside the labs, and it lets them lower prices without necessarily giving up margin.

Anthropic says Opus 5.5 needs less compute than Opus 5

Anthropic ties the lower Opus 5.5 price directly to the smaller amount of compute the model needs. In Anthropic's own tests at default settings, Opus 5.5 costs 40% less than Opus 5 on typical workloads (related: Claude Opus 5.5 pricing, benchmarks, API changes).

Woolf writes that he believes Anthropic on this point. Of his three reasons, it is the only one that does not depend on what competitors do.

Why cut prices right before an IPO?

Cutting the main revenue source by up to 80% runs against what a company would do to look good in an IPO filing. OpenAI is rumored to go public in 2027 and Anthropic within a couple of months, and investors will read their numbers closely.

Leaked financials show both companies losing billions on research and scaling. On the other hand, some reports put the profit margin on inference itself at around 80%, with caveats. Woolf wants to see how much of that margin is left after the cuts.

Will LLM prices go back up?

Woolf rejects the idea that cheap tokens are an Uber-style bait that will end in a price spike. AI skeptics keep predicting that the free ride ends once users are hooked. Woolf counters that neither behavioral economics nor the last few months of falling prices support that story.

Unlike Uber and Lyft, this market is not a duopoly

With many providers able to serve the same model at the same quality, Woolf expects the price of a token to settle near marginal cost. Ride-hailing in the 2010s had two companies that could set prices. LLM inference has many, the barrier to entry is modest, and switching costs almost nothing, so prices can keep falling.

The 29 providers on OpenRouter have no reason to keep serving a model at a loss; they would drop it and use the hardware for a profitable model instead. In response, both labs are building moats of their own: enterprise contracts, and personal agents such as OpenAI Dots and Claude Duet.

Woolf expects subscription limits to shrink first

When OpenAI launched GPT-6.1 Sol, it halved the usage allowance on the $200-a-month Codex tier. OpenAI said the change reflects its increasingly efficient models. Woolf expects subscriptions such as Codex and Claude Code to lose value before list API prices go up, if the labs ever run short of cash.

He changed his own plan for the same reason. In the GPT-5.6 Sol days he used up the $100-a-month Codex plan (5x Pro). Once GPT-6.1 Sol made the same work about 75% cheaper, he cut his quota to a fifth, since four times the efficiency left him with roughly the same amount of work.

What the cuts mean for API and subscription users

Looking at API prices and subscription allowances together shows where the savings reach users and where they are already being clawed back. On the API side, the cuts reached cache reads, and Woolf's math puts coding-agent costs 70% to 75% lower. Over the same weeks, the $200 Codex tier lost half its allowance.

API prices face a check that subscriptions do not: if a lab charges too much per token, developers can move the same workload to another model or another host. A subscription allowance is set by one company for its own product, with no OpenRouter list of 29 rival sellers. Tracking the two separately makes it easier to see which side of the bill the next change will hit.