OpenAI and Anthropic Cut AI Prices Up to 80% in Four Months — Here's What Actually Drove It

Four months ago, running OpenAI's mid-tier model cost $5 for every million input tokens. Today the same tier costs $2 — and the small model that sat beside it costs a fifth of what it did in June. Between June and October 2026, OpenAI and Anthropic cut the prices of their flagship large language models by as much as 80%, in a sequence of announcements so rapid that data scientist Max Woolf, who tracked every one of them, called the pattern "weird" — and then asked the obvious question: why would two companies about to go public slash their main source of revenue?

In short: OpenAI and Anthropic have cut flagship API prices by up to 80% since June 2026, driven less by generosity than by open-weight rivals like DeepSeek and Kimi, whose cheap models now run on dozens of competing inference providers. The headline cuts are real, but the bill you pay depends on cache reads, peak-hour pricing, and subscription allowances that moved in the opposite direction.

The price cuts, month by month

Woolf's timeline, published October 25, 2026, lays out the sequence. All prices below are per million tokens, in input / output / cache-read order, as listed at the time of each announcement:

The single biggest cut is the July 30 one: GPT-5.6 Luna went from $1/$6 to $0.20/$1.20 per million tokens, an 80% reduction on both input and output. But the cumulative picture matters more than any single row. From the GPT-5.6 family's June launch to GPT-6.1 Sol in late September, the cost of using OpenAI's mid-tier model fell roughly 75% — $5/$30/$0.50 down to $2/$10/$0.10 — while the model itself got better, not worse. Anthropic's side tells the same story: Opus 5.5, launched September 22 at $4/$20, costs 40–60% less in total than Opus 5, which it replaced two months after launch. Anthropic said plainly in its announcement that Opus 5.5 "requires less compute to serve than Opus 5, and its pricing reflects that."

An independent check supports the trend. Data from Vercel's AI model gateway, cited by Woolf, shows the average cost per token across its gateway falling to 43.3% of its April 2026 level. Aggregate spending through the gateway still rose over the same period — because, in Woolf's reading, the models are getting better faster than they are getting cheaper, so people use more of them.

The hidden cost: cache reads decide your real bill

The input and output prices in the table above are the numbers that make headlines. They are not the numbers that decide what an agentic coding session costs. When you use an AI coding agent, the model re-reads the entire conversation — including all the code it has already seen — on every turn. Woolf puts the cache-read share at 96–99% of total token spending for agentic coding, with cache hit rates of 90–100% depending on the provider. In other words, the cache-read price is the price that matters, and the input price is mostly a footnote.

That reframes the whole timeline. Anthropic's September 1 launch of Fable 5.1 kept the same $10/$50 input/output price as Fable 5 from June, but cut the cache-read price from $1 to $0.25 per million tokens — a 75% reduction. Applied to a workload where cache reads are nearly all of the spend, Woolf calculates that single change cuts the cost of using the model by 70–75%. OpenAI answered three weeks later: GPT-6.1 Sol halved its cache-read price to $0.10, matching Anthropic's 5%-of-input cache rate. The two labs are now effectively tied on cache pricing at the mid-tier, and the competition has moved to a line item most users never look at.

DeepSeek understood this first. Its V4 Flash model, released in April 2026 at $0.14/$0.28 per million tokens, carried a cache-read price of $0.0028 — about 2% of its input price, against the 5–10% the US labs were charging. Woolf argues the July 80% Luna cut was aimed squarely at that gap: Luna's cache read fell to $0.02, still seven times DeepSeek's rate, but close enough to stop the bleeding. DeepSeek's updated model, V4.1 Flash 1017, now costs $0.15/$0.60 per million tokens off-peak — double at peak hours — with the same $0.003 cache-read rate, and scores 45 on the Artificial Analysis Intelligence Index against 24 for the original Flash, according to Woolf's reading of the index.

Why now: open weights set a floor the big labs cannot ignore

Woolf's central argument is that the cuts are a response to competition, not a gift. Models have become close to interchangeable for many coding tasks: switching from one provider to another is a toggle in a harness like OpenCode or Zed Agent, or a different endpoint string through OpenRouter. When switching costs approach zero, the cheapest capable model sets the price for everyone.

OpenRouter is where that pressure is visible. In May 2026, Woolf counted 13 distinct providers hosting DeepSeek's V4 Flash on the platform. For the updated 1017 variant, there are now 29 — and several of them, including Cloudflare and DeepInfra, serve the model within 10% of DeepSeek's own price. A provider that tried to charge far above the market rate would simply lose the routing. Moonshot AI's Kimi K3.1 Pro, launched in late October, adds to the pressure: it scores 55 on the Artificial Analysis index against 58 for Claude Opus 5.5, at half the input/output price ($2/$10) and the same cache-read rate, with full model weights promised by November 4. Once those weights land, Woolf expects most of the 20-plus providers already hosting the older Kimi K3 to offer K3.1 Pro at roughly the same price.

IBM's chief executive made the same point from the buyer's side this week. Arvind Krishna told investors that open-weight models now handle routine work at about 10% of the cost of running frontier models on-premises, according to The Register's account of the call, summarized in the October 26 AI news digest. Krishna is, as the digest notes, talking his own book — IBM sells the hardware those on-prem deployments run on — but the direction of the claim matches the price data.

Not every price is falling. The same week's market data from TANDIU, which tracks 71 models' list prices, recorded one increase: Upstage's Solar Pro 4 rose from $0.09/$0.36 to $0.15/$0.60 per million tokens on October 20, a 66.7% rise in blended price. The same notes list the two models added that week — MiniMax M3.1 Flash at $0.15/$0.60 and Kimi K3.1 Pro at $2/$10 — both priced at the new, lower end of the market. The pattern is a floor dropping for the mainstream, not a universal collapse.

The catch: subscriptions moved the other way

Cheaper API tokens do not automatically mean cheaper AI. The same week GPT-6.1 Sol launched, OpenAI cut the usage allowance for the $200-a-month Codex tier by 50%, with the explanation that it was adjusting allowances "to reflect our increasingly efficient models." Woolf, who had been paying $100 a month for a Codex plan, describes his own response: once the model got roughly four times cheaper to run, he cut his plan's usage quota to a fifth of what it had been, ending up roughly neutral. The labs' API prices fell; the value bundled into subscriptions got quietly thinner.

That tension is the part worth watching. Both companies are reported to be preparing public listings — OpenAI in 2027 and Anthropic within a couple of months, according to reports cited by Woolf — and leaked financials show both losing billions on research and scaling. Cutting your main revenue line by up to 80% is the opposite of what a company does to polish a prospectus. One explanation, reported by Bloomberg Law and cited by Woolf with caveats, is that profit margins per token are as healthy as 80% even after the cuts — the inference business may simply have had that much room. Another is that the cuts are a land grab: cheaper tokens mean more usage, more lock-in through enterprise commitments, and more data flowing into personal agents like OpenAI's Dots and Anthropic's Claude Duet.

What the current prices actually are

Here is where the market stands today, October 26, 2026, using each vendor's published list price. Prices are per million tokens, input / output / cache read. Peak-hour and long-context surcharges apply on some models — DeepSeek, for example, doubles its V4.1 Flash rates during peak hours.

A few things stand out. First, $2/$10 is now the crowded middle: four different vendors' flagship-tier models sit on it. Second, the lowest listed prices still belong to the Chinese open-weight models — DeepSeek and MiniMax at $0.15/$0.60 — with the cheapest US-vendor tier at $0.10/$0.50 for Haiku 5.5 and GPT-6 Luna. Third, Google's Gemini 4 Argon matches the $2/$10 tier only until November 30, when its price doubles, a reminder that some of today's prices are explicitly temporary. For the full story on the newest reasoning models, see our breakdown of OpenAI's next-generation reasoning models.

Will prices go back up?

The argument that token prices are artificially low — that the labs are loss-leading like Uber and Lyft in the 2010s and will raise prices once everyone is hooked — gets a hearing in Woolf's post, and he rejects it. Uber and Lyft were a duopoly with the power to set prices; the inference market has 29 providers for a single model and a new open-weight release every few weeks. As long as any of those providers can serve a model profitably at the current price, the price has a floor set by competition, not by the two biggest labs.

That does not mean your bill can only fall. The Codex allowance cut shows where the money can be recovered: subscriptions, usage caps, and premium tiers. The labs are also building moats that have nothing to do with token prices — enterprise contracts, personal agents, and the ecosystems around them. If you are buying tokens by the million, the market has rarely been better. If you are buying a subscription, read the allowance, not the model name.

FAQ

Did OpenAI and Anthropic really cut prices by 80%?

For one model, yes: GPT-5.6 Luna fell 80% on July 30, from $1/$6 to $0.20/$1.20 per million input/output tokens. Other cuts were smaller — 20% for GPT-5.6 Sol, 40–60% in total for Claude Opus 5.5 versus Opus 5, and a cumulative 75% for the GPT-5.6 Sol family by the time GPT-6.1 Sol arrived. The "up to 80%" figure is the largest single cut, not the average.

Why are AI model prices falling so fast?

Three forces, according to Woolf's analysis: open-weight models like DeepSeek and Kimi that undercut the big labs on price; inference infrastructure getting cheaper to run, which Anthropic cited directly for Opus 5.5; and near-zero switching costs, since models are interchangeable through harnesses and OpenRouter. Each new release has to beat the last one's price to win usage.

Are the cheaper models worse than the expensive ones they replaced?

On the benchmarks the labs publish, no — the cheaper replacements score higher than the models they superseded. GPT-6.1 Sol is reported by OpenAI to match GPT-6 Astra's performance at a fifth of the price, and Opus 5.5 outperforms Fable 5.1 on Anthropic's own charts. Those are vendor-reported figures, and independent benchmarks like the Artificial Analysis index broadly agree on the direction: Kimi K3.1 Pro at $2/$10 sits within three index points of Opus 5.5 at $4/$20.

What does the price collapse mean for ChatGPT and Claude subscriptions?

API price cuts do not automatically change subscription prices. In fact, OpenAI cut the Codex $200-tier usage allowance by 50% at the same time GPT-6.1 Sol launched, arguing that more efficient models justify lower allowances. If you subscribe, the allowance and the reset rules matter more than the per-token price.

Will prices keep falling?

Nothing guarantees it. The competitive pressure is real — 29 providers host DeepSeek's latest model on OpenRouter, and Kimi's weights land November 4 — but both labs are also heading toward IPOs and losing billions on research, which could push them toward monetization through subscriptions and enterprise deals instead. The safest reading is that token prices have a competitive floor, while the total cost of using AI can still rise through allowances and tiers.

Key takeaways

• OpenAI and Anthropic cut flagship API prices by up to 80% between June and October 2026, with GPT-5.6 Luna's July 30 cut from $1/$6 to $0.20/$1.20 the largest single reduction.
• Cache-read pricing, not input pricing, decides what agentic coding costs — cache reads are 96–99% of token spend, and the Fable 5.1 cache cut alone reduced total cost by 70–75%.
• Open-weight rivals set the floor: DeepSeek's V4.1 Flash 1017 costs $0.15/$0.60 off-peak with a $0.003 cache rate, and 29 providers host it on OpenRouter.
• The $2/$10 tier is now the crowded middle, shared by GPT-6.1 Sol, Claude Sonnet 5.5, Kimi K3.1 Pro, and Gemini 4 Argon — though Argon's price doubles after November 30.
• Subscriptions did not follow the cuts down: OpenAI halved the Codex $200-tier allowance when GPT-6.1 Sol launched, so check allowances, not just token prices.

The price war of 2026 is not a sale. It is the market discovering what a token is actually worth when anyone can serve one — and as of October 26, the answer keeps coming back lower than the last time anyone asked.

Sources: Max Woolf, "OpenAI and Anthropic cut their LLM prices by up to 80% within a few months. Why?" (minimaxir.com, October 25, 2026); OpenAI API pricing; Anthropic API pricing; DeepSeek API pricing; AI News, October 26, 2026 (ai0.news); TANDIU weekly data notes, October 19–25, 2026. Historical prices are as reported in Woolf's timeline; current list prices were checked against the vendors' pricing pages on October 26, 2026.