80% price cuts aren't generosity: they're a market fighting over share

Nobody cuts a price by 80% out of kindness. When a company does it to a product people already pay for, something else is pushing: a rival, a cost curve, or a plan to make the money back elsewhere. With OpenAI and Anthropic, we think it's all three.

The latest spark is a minimaxir.com post whose premise is blunt: no one cuts prices by 80% unless some externality is in play. Per Bleap's write-up, OpenAI dropped GPT-5.6 Luna by 80% on July 30, 2026, from $1/$6 to $0.20 per million input tokens and $1.20 per million output tokens. The Wall Street Journal had reported in June that OpenAI was weighing drastic cuts to compete with Anthropic, which it expected to follow with similar moves.

The externality is a rival you can't ignore

Read that WSJ headline again: OpenAI was anticipating a war for users. That's not the language of a company with a moat. It's the language of a market where the product is converging.

The pressure isn't only from each other. One Medium analysis claims DeepSeek's V4-Pro scores 80.6 on SWE-bench Verified against 80.8 for Claude Opus 4.7, at roughly one twenty-eighth of the price. That's one author's reading of price pages, and benchmarks aren't your workload. Still, if even half of it holds, premium pricing for near-identical capability becomes very hard to defend. Cut or lose the customer.

That's what commoditization looks like: differentiation moves from "who has the smartest model" to "who is cheapest at good enough." Epoch AI's March 2025 analysis had already found prices for a given performance level falling at a median of 50x per year across its benchmarks, so in our view the underlying cost curve gives these companies room to cut. Whether the latest cuts are real efficiency or margin sacrificed for share is harder to say: Epoch notes that margins may explain some of the earlier drops but says it found no clear evidence for it, so we shouldn't pretend to know the split.

Cheaper tokens, bigger bills

Here's the part the headlines skip. Computerworld, citing Gartner, says token costs may fall 95% by 2030 while inference costs for agentic workflows rise more than fivefold over the next two years. Gartner calls the assumption that savings will flow through a "token-deflation illusion."

We find that persuasive. A price per token is a unit price, and the unit you consume is changing. Agents burn 5x to 30x more tokens than a chatbot on equivalent tasks, according to the same report. An 80% discount that you offset by running five times more calls is a bill that stays flat. Cheaper tokens don't make AI cheaper; they make heavier usage affordable, and heavier usage is exactly what we suspect the providers want.

Enjoy the discount, but don't marry the vendor

So what should a developer do? Three things, in our view.

First, take the cut. Rerun your numbers; if you sized a business case a year ago, it's probably stale.

Second, assume subsidized pricing can reverse. The same market that cut prices also saw subscription terms shift: SemiAnalysis reported in October that OpenAI had halved the API-equivalent value of its $200 plan. That's a subscription, not the API, but it shows how quickly generosity gets reshaped once a company needs to reprice. Cheap today is not a contract.

Third, build so you can leave. Route requests to the cheapest model that does the job, keep prompts and evaluation sets portable, and test an alternative provider on your own data every quarter. A price war is the best time to negotiate and the worst time to hardcode one vendor's quirks into your product.

Our verdict: this is good news with an asterisk. Competition is doing what competition does, and developers are the beneficiaries for now. But a market racing to the bottom on price eventually has to find money somewhere, whether through usage growth, tiered plans or new pricing rules. Enjoy the 80%, budget as if it won't last, and keep your architecture portable.