Why are AI model prices dropping? OpenAI and Anthropic's 2026 cuts of up to 80%, checked against their price pages

Why are AI model prices dropping? Between July and October 2026, frontier model API prices fell sharply: OpenAI cut one model by 80% on July 30, 2026, then halved prices again with each new generation, most recently with GPT-6.1 Terra on October 26, while Anthropic cut Claude Opus by 20% and its cache-read prices by up to 75%. The companies say serving their models got cheaper. Data scientist Max Woolf, in an October 25, 2026 post that drew more than 400 points on Hacker News, argues that competition is the bigger reason. This analysis checks every cut against the vendors' own pages and weighs the explanations. For what the cheapest options cost today, see our comparison of frontier APIs and open models.

This analysis is based on OpenAI's and Anthropic's pricing pages, model pages, changelogs, release notes and launch announcements, Max Woolf's post on minimaxir.com, and Bloomberg News reporting, checked October 27, 2026. Prices are list prices per million tokens in US dollars.

The short version

• The 80% is one cut. OpenAI cut GPT-5.6 Luna from $1 input and $6 output to $0.20 and $1.20 on July 30, 2026. No other single price change in 2026 was as large.
• OpenAI kept halving with each generation. GPT-6 Sol ($2/$10) cost half of GPT-5.6 Sol's discounted $4/$20, and GPT-6.1 Terra ($1/$5) cost half or less of GPT-5.6 Terra's $2/$12.
• Anthropic cut less per token. Opus 5.5 ($4/$20) is 20% cheaper than Opus 5, with cache reads 60% cheaper. Fable 5.1 cut cache reads 75% at the same token price.
• The stated reason is efficiency. OpenAI cited faster GPU kernels and better speculative decoding; Anthropic says Opus 5.5 "requires less compute to serve."
• Woolf's reason is competition. He says easy switching between models and cheap open-weight rivals push prices toward the cost of serving them.

How much have AI model prices dropped in 2026?

Ten price changes from July to October 2026 can be checked against OpenAI's and Anthropic's own pages. Four were cuts to an existing model; six came with a new model that replaced an older one at a lower price.

OpenAI's cuts to existing models came first. On July 30, 2026, GPT-5.6 Luna dropped 80%, from $1 input, $0.10 cached input and $6 output to $0.20, $0.02 and $1.20, and GPT-5.6 Terra dropped 20%, from $2.50/$0.25/$15 to $2/$0.20/$12, according to OpenAI's API changelog and developer-forum announcement. On August 21, OpenAI cut GPT-5.6 Sol from $5 input and $30 output to $4 and $20 for three months; its model page says the promotional price runs "at least through November 21, 2026."

Then came the new generations. GPT-6 Sol launched on September 22 at $2/$10, and GPT-6 Luna at $0.10/$0.50, roughly half the GPT-5.6 prices they replaced. GPT-6.1 Sol kept GPT-6 Sol's $2/$10 on September 29 but halved cached input to $0.10. GPT-6.1 Terra arrived on October 26 at $1/$5, against $2/$12 for GPT-5.6 Terra, as our GPT-6.1 Terra story details.

Output price per million tokens, before and after each 2026 change

US dollars per million output tokens

GPT-5.6 Luna cut$6.00$1.20

GPT-5.6 Terra cut$15.00$12.00

GPT-5.6 Sol promo$30.00$20.00

GPT-5.6 Sol to GPT-6 Sol$20.00$10.00

GPT-5.6 Luna to GPT-6 Luna$1.20$0.50

Opus 5 to 5.5$25.00$20.00

GPT-5.6 Terra to GPT-6.1 Terra$12.00$5.00

• Before
• After

Show the data

Change	Before	After
GPT-5.6 Luna cut	$6.00	$1.20
GPT-5.6 Terra cut	$15.00	$12.00
GPT-5.6 Sol promo	$30.00	$20.00
GPT-5.6 Sol to GPT-6 Sol	$20.00	$10.00
GPT-5.6 Luna to GPT-6 Luna	$1.20	$0.50
Opus 5 to 5.5	$25.00	$20.00
GPT-5.6 Terra to GPT-6.1 Terra	$12.00	$5.00

Anthropic's moves were narrower. Claude Opus 5.5, released September 22, costs $4 input and $20 output, which Anthropic's announcement calls "20% less than Opus 5" ($5/$25), and its cache reads cost $0.20, "60% less than Opus 5." Claude Fable 5.1, released September 1, kept Fable 5's $10/$50 but cut cache reads from $1 to $0.25. On October 7, Anthropic halved Sonnet 5.5's cache reads from $0.20 to $0.10. And on August 10, it scrapped a planned increase that would have raised Claude Sonnet 5 from its introductory $2/$10 to $3/$15 on September 1, according to its release notes.

Last verified: October 27, 2026

So the headline needs a qualifier. "OpenAI and Anthropic cut their LLM prices by up to 80%," as Woolf's title puts it, is accurate for OpenAI's Luna cut. Anthropic's largest per-token cut was 75%, on Fable cache reads, and its cut to Opus input and output prices was 20%.

Why are OpenAI and Anthropic cutting prices?

The vendors give one answer; Woolf and the reporting around the cuts suggest others. They aren't mutually exclusive.

What the vendors say: serving got cheaper. OpenAI's July 30 announcement said GPT-5.6 had been used to optimize its own runtime, with "20% lower serving costs through production GPU kernel improvements" and more than 15% better token-generation efficiency from improved speculative decoding, and that it was passing those gains on to customers. Anthropic's Opus 5.5 announcement says the model "requires less compute to serve than Opus 5, and its pricing reflects that." Neither company publishes its serving costs, so these claims can't be checked from the outside.

Woolf's main argument: competition and commoditization. Woolf, a former BuzzFeed senior data scientist, argues that switching models has become trivial, "just a toggle in a harness like OpenCode and Zed Agent," and that on OpenRouter any inference provider can list a model it hosts, so no lab has much pricing power. He points to cheap open-weight models as a floor: he suspects the July Luna cut was aimed at DeepSeek's small V4 Flash models, and notes that Moonshot priced Kimi K3.1 Pro at $2/$10 with scores near Opus 5.5's, as our Kimi K3.1 Pro story covers. He sees "a clear tit-for-tat between Anthropic and OpenAI" in the timing. He also offers a "less-cynical" factor, better infrastructure, and says he believes Anthropic's statement that Opus 5.5 is cheaper to serve.

A price war was reported before it happened. On June 11, 2026, Bloomberg News reported, citing The Wall Street Journal, that OpenAI was considering "significant price cuts" in anticipation of similar moves by Anthropic, possibly a "pricing war in the lead-up to their expected public offerings." People familiar with the matter told the outlets nothing had been decided, and OpenAI declined to comment. The cuts that followed fit that report, though neither company has described them as a response to the other.

The loss-leader theory. Some critics argue AI prices are being held artificially low and will rise. Woolf calls the idea that token prices will suddenly rise "more a fantasy," because many inference providers serve the same open models and keep their prices near cost. He also notes that both companies are close to going public, which he says "raises further questions about the price cuts," and writes that he wants to see Anthropic's prospectus and its inference profit margins. Our story on Anthropic's S-1 covers what the filing shows.

Will AI model prices keep falling?

Woolf expects them to, writing that "the 'true cost of a token' may end up being where price equals marginal cost." Two facts complicate that.

A promotion ends. GPT-5.6 Sol's $4/$20 price is a promotion that OpenAI says runs at least through November 21, 2026. OpenAI hasn't said what happens after that.

The open-weight pressure is real but small in volume. On IBM's October 21, 2026 earnings call, CEO Arvind Krishna put open models' token cost at 25% to 10% of frontier models' and called them "90 percent as effective," as our open-weight vs open-source explainer reports. But Woolf notes that open-weight providers still hold a minority of token share. Citing Vercel's AI Gateway data, he says cost per token there is 43.3% of its April 2026 level while total spending keeps rising, because people use more tokens.

Does a lower token price mean a cheaper bill?

Not necessarily. A model's bill depends on how many tokens it uses per task, and reasoning models spend output tokens on thinking that you pay for, as our reasoning models explainer shows. Anthropic's Opus 5.5 claim that it costs "40% less than Opus 5 on typical workloads" rests on the model using fewer tokens per task, not only on the 20% price cut, and is a vendor claim. Anthropic's pricing page also notes that Claude 4.7 and later models use a tokenizer that produces roughly 30% more tokens for the same text, so per-token comparisons with older Claude models understate the real cost.

Cache pricing matters more than the headline rates for agents. Woolf writes that cache reads are "the vast majority (96% - 99%)" of token spending in agentic coding, which is why the 60% to 75% cache cuts may save heavy users more than the 20% to 50% cuts to input and output prices. For how to mix cheap and expensive models, see our model routing explainer.

What's confirmed, reported and uncertain

• Confirmed (vendor pages): every price and date in the chart and table above; OpenAI's efficiency explanation for the July 30 cuts; Anthropic's "less compute to serve" statement; the cancelled Sonnet 5 increase; the November 21 end of GPT-5.6 Sol's promotion "at least."
• Reported: OpenAI's June deliberations over price cuts (Bloomberg, citing The Wall Street Journal; the WSJ original wasn't read).
• Argument, not fact: Woolf's view that competition is the main driver and that prices will trend to marginal cost; his cache-share and Vercel figures, which are his own.
• Uncertain: the labs' actual serving costs and margins, and whether GPT-5.6 Sol's promotional price ends after November 21.