AI price war: OpenAI catches up with Anthropic in business spending

For most of 2026 the story of business AI had a clear leader. Anthropic built the models that developers trusted for serious coding work, Claude Code became a habit inside engineering teams, and companies poured money into Claude through every channel they could find. That lead is now gone, at least by one closely watched measure. According to data from the model marketplace OpenRouter, reported by the Wall Street Journal on 7 October 2026, business spending on OpenAI and Anthropic models was roughly even in September. In January, Anthropic still took about three quarters of that money.

The reason is not a single breakthrough model. It is price. OpenAI spent the summer shipping cheaper models and then cutting their prices again, and a large share of customers followed the savings. Anthropic answered with its own cuts at the start of October. The result is the first real AI price war between the two labs that sell the most expensive frontier models, and it lands at an awkward moment: both companies are preparing for public listings and need to show investors that their revenue is durable.

This article walks through what the data shows, what it does not show, why customers moved, how deep the price cuts went, and what a price war means for developers, for the two companies and for everyone who pays for AI in the end. The core numbers come from the WSJ report, with additional detail from summaries by Gate, Post Cutoff, Magica and OnTimeBrief.

What the OpenRouter numbers actually measure

OpenRouter is a startup that sits between developers and model providers. Instead of signing separate contracts with OpenAI, Anthropic, Google and others, a company sends its requests to OpenRouter, picks a model, and pays one bill. That makes OpenRouter a useful window into how companies really spend, because switching from one model to another is often a one line change in a configuration file. When prices move, the money on OpenRouter moves quickly.

The analysis behind the WSJ story covered about 120,000 companies that use both OpenAI and Anthropic models through the platform. OpenRouter said the group is mostly made up of AI native startups, plus some older technology companies and large enterprises. Among those customers, the split of spending between the two labs went from about 75 to 25 in Anthropic's favor in January to close to 50 to 50 in September.

Weekly data tells a sharper story. Post Cutoff, citing a chart first highlighted by OfficeChai, notes that in the week of 7 September OpenAI's share of the combined spend passed 50 percent for the first time since February 2024. GPT 6 Astra alone accounted for roughly a fifth of that combined spend. Separate data from the corporate card company Ramp reportedly showed OpenAI briefly ahead in mid September before Anthropic narrowly took the lead back.

What the data does not show

It is worth being careful here, because headlines like "OpenAI catches Anthropic" can easily overstate the case. The OpenRouter figures cover only API spending routed through OpenRouter. They do not include direct enterprise contracts with either company, they do not include ChatGPT or Claude subscriptions, and they do not include cloud marketplaces like Amazon Bedrock, Google Vertex or Microsoft Azure, where a lot of large company spending lives.

The sample also only counts companies that use both labs. A company that runs everything on Claude through a direct contract simply does not appear. And spending is not the same as usage: if OpenAI cuts prices by 80 percent and a customer moves half its work over, the dollar figures can look very different from the token counts. Magica points out that a fair follow up would need a consistent customer group and both spending and usage disclosed side by side.

So the honest summary is narrower than the headline: among startups and tech companies that already use both providers and buy through a router, OpenAI won back a large share of the budget this year. That is still a meaningful signal, because these are exactly the customers who switch fastest and who often predict where larger buyers go next.

How OpenAI turned it around

The turning point was the GPT 5.6 family that OpenAI released over the summer, made up of three tiers called Sol, Terra and Luna. Instead of pushing only the most capable model, OpenAI positioned the lineup around cost: a strong top model and two cheaper tiers that handle the bulk of everyday work well enough.

Shortly after launch, OpenAI cut prices again. According to the WSJ, GPT 5.6 Luna became 80 percent cheaper and GPT 5.6 Terra 20 percent cheaper. Then came GPT 6, with Sol and Luna versions that Post Cutoff reports were priced at about half of the matching GPT 5.6 models. For a company that sends millions of routine requests per day, such as summarizing support tickets, classifying documents or extracting fields from invoices, those cuts change the monthly bill dramatically.

Why customers say they switched

Price was the main reason, but not the only one. Companies quoted in the coverage, including the developer tool maker Retool and the finance platform Ramp, pointed to lower running costs. Some also raised concerns about a 30 day data retention policy connected to Anthropic's Fable 5 model, which made it harder for them to meet their own privacy commitments.

A third reason is less dramatic but just as important: diversification. Many teams learned during Anthropic's capacity crunch in the spring, when Claude Code demand caused frequent outages, that depending on one supplier is risky. Routing tools make it easy to keep two or three providers ready and move work to whoever is cheapest, fastest or simply online at that moment.

Not everyone moved everything. One founder quoted in the coverage said his startup still prefers Anthropic's Fable 5.1 for the hardest frontier work while looking for cheaper terms on everything else. That points to a market splitting in two: a premium layer where quality decides, and a large bulk layer where the lowest acceptable price wins.

Anthropic strikes back

Anthropic did not sit still. In late September it began rolling out its Claude 5.5 series with a strong focus on efficiency. On 7 October it launched Claude Haiku 5.5 at 0.10 US dollars per million input tokens and 0.50 dollars per million output tokens. That is about 90 percent below Haiku 4.5 for most requests and puts its input price level with GPT 6 Luna. Anthropic also halved cache read prices for Sonnet 5.5 and added monthly API credits for some Max and Team plans.

In other words, the September numbers describe a world just before Anthropic's counterattack. The October and November data will show whether those cuts pull customers back or whether OpenAI's gains stick. Because switching costs on routers are so low, it would not be surprising to see the share swing again within weeks.

Why the IPO timing matters

All of this happens against a financial backdrop that raises the stakes. According to the reporting, Anthropic is working toward a stock market listing as soon as November, and OpenAI is expected to follow next year. Both companies carry valuations reported above a trillion dollars and both spend enormous sums on computing power.

API market share is one of the clearest signals investors can read about enterprise demand. A lab that can show it is winning business customers looks like a durable company. But a price war cuts both ways. Lower prices win share, yet they also shrink the revenue per request at a time when both labs are still losing large amounts of money. Several readers of the original story pointed out exactly that tension: the more aggressive the cuts, the harder it becomes to argue that the business model works on its own.

What it means for developers and companies

For anyone building with AI, the short term news is simple and good: models are getting cheaper fast, and competition between the two strongest labs is now driven by price as much as by benchmark scores. A few practical takeaways follow from that.

• Do not lock in. If your code talks to one provider directly, consider a thin abstraction or a router so you can move work when prices change.
• Split your workloads. Use the expensive frontier model only where it clearly beats the cheap tier, and send routine jobs to the cheapest model that passes your own tests.
• Re test regularly. With price cuts arriving every few weeks, a cost comparison from the summer is already out of date.
• Read the data terms. Retention and training policies differ between providers and even between models, and they now influence buying decisions directly.

The bigger picture

A year ago the AI race was mostly framed as a contest of intelligence: who has the smartest model. The September numbers show that for a large part of the market, the contest is turning into something more ordinary, the same kind of competition that shaped cloud computing a decade ago. Good enough performance at the lowest price wins the bulk of the work, while the very best models fight over a smaller premium segment.

That is a healthy development for customers and a difficult one for the labs. OpenAI proved that aggressive pricing can move billions of tokens in a few months. Anthropic has now shown it is willing to match those prices. The next few months of OpenRouter, Ramp and cloud marketplace data will tell us whether either company can turn share into profit, or whether the price war simply keeps pushing costs down for everyone. For now, the most important lesson is that loyalty in the AI API market is thin, and the price tag matters more than many expected.

Sources: Wall Street Journal (paywalled), Gate summary, Post Cutoff, Magica, OnTimeBrief.