OpenAI and Anthropic cut LLM prices up to 80%. The moat is gone.

On September 22, OpenAI priced GPT-6 Sol at $2 / $10 / $0.20 per million input, output, and cache-read tokens. Seven days later it halved the cache price again, to $0.10, matching Anthropic’s cache read at 5% of input. That is a 75% cumulative cut on the GPT-6 family in a single week. Max Woolf’s timeline of the June-to-September price war lays it out release by release, and the pattern is hard to miss: two labs, trading cuts, on models that are still getting better.

The obvious read is that this is a land grab. The less obvious and more interesting read is that it is a forced move. OpenAI and Anthropic are not cutting prices because they want to. They are cutting because the thing they sell has stopped being scarce.

What actually happened, in order

The sequence matters. On June 26 OpenAI shipped GPT-5.6 Sol at $5 / $30 / $0.50 and Luna, a small model, at $1 / $6 / $0.10. On June 30 Anthropic launched Claude Sonnet 5 at $2 / $10 / $0.20, a launch price it called temporary. On July 24 came Claude Opus 5 at $5 / $25 / $0.50. Then, on July 30, OpenAI cut GPT-5.6 Luna by 80%, to $0.20 / $1.20 / $0.02.

That Luna cut is the tell. According to the writeup, without it Luna would have lost to DeepSeek V4 Pro on cost-per-task. DeepSeek’s Flash V4 0420 had landed in April at $0.14 / $0.28 / $0.0028, with a cache read at 2% of input, and went viral partly because OpenAI had no small model for simple tasks at the time. The 80% cut was not a gift to customers. It was a response to a Chinese open-weight model that had already taken the cheap end of the market.

The rest of the timeline is the same story at higher price points. Anthropic made the Sonnet 5 launch price permanent on August 10. OpenAI cut GPT-5.6 Sol by about 20% on August 21, for three months. Anthropic shipped Fable 5.1 on September 1 with a 75% cache-read cut and said it was strictly better on benchmarks than Fable 5. OpenAI answered with GPT-6 Astra on September 3 at Fable-comparable pricing. GPT-6 Sol and GPT-6 Luna arrived September 22 at half the input and output price of their predecessors. Anthropic shipped Claude Opus 5.5 the same day, explicitly stating that “Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that.”

That sentence is the whole article. The price cut is a compute-efficiency claim wearing a pricing table.

The cache read is the real price

Most coverage of token prices looks at input and output. For agents, that is the wrong number. The writeup notes that cache read costs are 96% to 99% of total token expenditure when you use an agent for coding, because the cache holds the entire prior conversation each turn, and it gets hit 90% to 100% of the time depending on the inference provider. A 75% cut in cache-read cost works out to a 70% to 75% cut in what it costs to run the model. The footnote is the bill.

This is why the September 29 GPT-6.1 Sol move matters more than the headline input price. Cutting cache read from $0.20 to $0.10 to match Anthropic’s 5%-of-input ratio is what makes the GPT-6 family cumulatively 75% cheaper. It is also why DeepSeek’s 2% cache-read ratio was such a threat: the advertised price understated the effective price.

Why the “true cost” argument falls apart

There is a durable argument that these prices are artificially low, that OpenAI and Anthropic are doing an Uber-style loss-leader, and that the bill comes due when the free ride ends. The writeup rejects it, and we think correctly, on structural grounds. Uber and Lyft were a duopoly that could set prices once the competition was gone. Inference is not that.

On OpenRouter, any provider can list a model it hosts, and requests route to any available endpoint unless the developer specifies a preference like lowest price. DeepSeek V4 Flash had 13 distinct providers in May. The 1017 variant has 29. Cloudflare and DeepInfra are within 10% of DeepSeek’s own cost on effective input pricing. Those providers have no reason to serve an unprofitable model; they would move the compute to something that pays. So “the true cost of a token” converges toward marginal cost, not away from it.

The competitive pressure is not only at the cheap end. Moonshot AI’s Kimi K3.1 Pro, announced days ago, scores 55 on the Artificial Analysis index against Opus 5.5 at 58, at half the input and output price and the same cache read, $2 / $10 / $0.20. Kimi K3 already has more than 20 hosting providers, and the weights for K3.1 Pro drop November 4. Moonshot priced it below the original K3. That is a frontier-adjacent model at commodity pricing.

What the IPO timing actually says

Both labs are reportedly close to public markets: OpenAI in 2027, Anthropic within a couple months. Cutting your primary revenue source by up to 80% before an S-1 looks like self-sabotage. It is not, if inference margins are as healthy as some reports suggest, with profit-per-token margins cited as high as 80% with caveats. The cuts are partly a bet that volume and efficiency outrun the price decline, and partly a defensive necessity. An S-1 that shows you losing share to open-weight models is worse than one that shows you cutting prices to hold it.

OpenAI and Anthropic are not cutting prices because they want to. They are cutting because the thing they sell has stopped being scarce.

The Vercel gateway data in the writeup captures the tension neatly. Cost per token has fallen to 43.3% of the April 2026 level, but aggregate spend is rising, because the models are getting better faster than they are getting cheaper. Cheaper tokens do not shrink the bill if you use more of them. They change what is worth building.

That is the part builders should internalize. The old calculus, where you ration context and pick the cheap model for grunt work, is being rewritten every few weeks. A cost model you tuned against GPT-5.6 Sol in June is wrong by September. The writeup’s own author cut a Codex Pro subscription to a fifth of its quota after GPT-6.1 Sol made the same work 75% cheaper, which is not the behavior the labs want but is exactly the behavior the pricing invites.

The moat, if there is one, is no longer the model. It is enterprise commitments and personal agentic products like OpenAI Dots and Claude Duet, the things that make switching cost something other than an API string. On raw tokens, the labs are in a commodity business with 29 sellers and a competitor shipping comparable quality at half the price on November 4. Watch what happens to Opus 5.5 pricing after the Kimi weights drop. That is the number that tells you whether the cuts are strategy or surrender.