· AI
Anthropic Cut Its Cheapest Model 90 Percent. Its Flagship Tier Tells a Different Story.
The cheap tier of this price war is real competition settling a price. The expensive tier is still mostly running on someone else's money.
Anthropic released Claude Haiku 5.5 on October 7 and priced it at $0.10 per million input tokens and $0.50 per million output tokens for prompts under 100,000 tokens, down from Haiku 4.5's $1 and $5, a roughly 90 percent cut at the low end that Anthropic says averages out to 75 percent cheaper to run across real workloads. That price lands exactly on top of OpenAI's cheapest model, GPT-6 Luna, which OpenAI had already cut to the same $0.10 and $0.50 two weeks earlier. Sam Altman's only comment on the move, posted to X, was that the new models are "half the price per token," with "even less per task" once efficiency gains compound. Two labs, two cuts, one identical number at the bottom of the menu.
My position: the cheap tier of this fight is genuine competition doing exactly what competition is supposed to do, and small businesses buying AI by the million tokens should take the savings and route their routine work there without much hesitation. The flagship tier is a separate question, and I would not build a five-year plan on today's price for it.
The steelman for skepticism, stated fully. A reasonable critic would point out that neither of these companies is pricing from a position of profit. Anthropic's leaked IPO prospectus, reviewed by Reuters and reported by Wolf Street, shows revenue of $4.6 billion in 2025 against an operating loss of $8.1 billion and compute spending of $7.3 billion, nearly triple the year before. The company has also disclosed plans to commit $518 billion to infrastructure in the coming years, funded by investors betting on a target valuation north of $2 trillion, not by a business that currently earns a profit on the tokens it sells. A skeptic can fairly argue that calling this "the market setting prices" gives the market too much credit when venture capital and a looming IPO are doing the actual subsidizing, and that the bill comes due the moment public shareholders demand margin discipline instead of growth.
That case is honest, and I do not think it is wrong about the flagship models. Where it misses is treating the whole price list as one number. Haiku and Luna sit at the bottom of each company's lineup, which is precisely the segment where open-weight models from Meta, Mistral, DeepSeek, and Alibaba's Qwen already compete on close to equal footing, and where this blog has covered those models taking real token volume at a fraction of frontier pricing. A lab cannot subsidize its way out of a price floor that open-weight competitors set independently of its own balance sheet. OpenAI's own explanation for its cut, improvements in caching and inference rather than a promotional discount, points the same way: part of this specific price move is a real efficiency gain, not a loss leader. The $8.1 billion hole is concentrated in training the next frontier model and the infrastructure behind it, which is exactly the cost structure attached to Opus and Sol tier pricing, not Haiku and Luna.
So the honest read is not "ignore the subsidy" and it is not "the whole price war is fake." It is that the subsidy and the competition are sitting at different levels of the same price list, and you should treat them differently. Run the math on your own usage: a workload doing 50 million input and output tokens a month at Haiku 5.5's new rate costs about $30 combined, against roughly $300 a month eighteen months ago at Haiku 4.5 pricing before this round of cuts. That gap is durable because two independent labs and a field of open-weight alternatives all landed near the same floor without colluding on it. The price of whatever flagship model your team is building against next year is a different bet, and I would not assume it only moves down.
The practical move is to route the routine 90 percent of your token volume, the drafts, the summaries, the classification and extraction work, to whichever commodity-tier model is cheapest this quarter, and treat that choice as disposable since the floor keeps dropping regardless of which lab wins any given month. Keep the expensive model for the narrow slice of work that actually needs it, and do not sign a long contract assuming its price holds. If you are not sure which of your AI costs sit in which bucket, that split is worth doing before the next invoice, and it is most of what picking a model without overpaying actually means in practice.
Sources
References used in this article. Links also appear alongside the relevant claims.
Let's make it happen.
You don't need a polished brief. A couple of lines about where you want your company to go is plenty, and we'll come back with what we'd tackle first.
We scope the work around your goals and timeline before quoting anything, so you know exactly what you're getting.