The AI price war just went from background noise to front-page strategy.

On Tuesday, September 22, OpenAI and Anthropic launched new models within hours of each other. Both launches centered on the same selling point: cheaper AI. OpenAI priced its GPT-6 Sol and GPT-6 Luna models 50% below the promotional API rates of the GPT-5.6 models they replace. Anthropic said its Claude Opus 5.5 costs about 40% less to run than Opus 5 on typical workloads.

Four days later, the market reaction started rolling in. Amjad Masad, chief executive of the AI coding platform Replit, told reporters his company is now using less open-source AI than it did earlier this year. The reason? The closed-source labs keep cutting prices.

Open source used to be the obvious cheap option. That assumption is now cracking.

What OpenAI and Anthropic actually launched

OpenAI’s September 22 announcement brought two new GPT-6 siblings. GPT-6 Sol targets professional work like coding and long-running agent tasks. GPT-6 Luna handles high-volume jobs such as summarization. Per OpenAI’s own pricing table, Sol runs $2 per million input tokens and $10 per million output tokens. Luna sits at $0.10 input and $0.50 output.

The timing was pointed. OpenAI’s developer account posted the news with characteristic swagger: “GPT-6 Sol and Luna just landed in Astra’s orbit. Both launch today with API prices 50% lower than GPT-5.6. Build with Sol. Scale with Luna. To production and beyond.”

Hours later, Anthropic answered with Claude Opus 5.5, priced at $4 per million input tokens and $20 per million output tokens — a 20% list-price cut from Opus 5. Anthropic’s headline number is bigger than the list price suggests: the company says typical workloads cost about 40% less because Opus 5.5 uses fewer tokens per task, cache reads dropped 60% to $0.20 per million tokens, and output generation runs over 30% faster.

On paper, that’s two frontier labs cutting prices on the same afternoon.

Replit’s CEO calls it like he sees it

Masad’s comments land because Replit sits in the crossfire. The company lets people build apps by describing them in plain English, and it rents models from OpenAI, Anthropic, and Google rather than training its own. In August it launched a free tier running on OpenAI’s GPT-5.6 Luna. Now Masad says OpenAI and Anthropic have been aggressively lowering prices for their smaller, faster models — enough that Replit’s mix has shifted away from open-source models this year.

He’s not celebrating the discount. Masad has warned that the labs can subsidize their own coding tools and price them below cost to compete with startups like Replit. When the companies that sell you the models also sell products that compete with yours, cheaper APIs come with strings attached.

That tension is the real story here. For developers, falling token prices are genuinely good news. For startups building on those same models, the discount can feel like a squeeze.

AI price war: efficiency gains or a price grab?

Depends who you ask.

The charitable read: better engineering. OpenAI credits improved prompt caching and inference efficiency for the GPT-6 cuts, and notes 90% discounts on cached input-token reads for both Sol and Luna. Anthropic points to a leaner Opus 5.5 that burns fewer tokens and serves cached context far more cheaply. Cheaper serving passed on to customers is a normal industry maturation story.

The cynical read: market-share warfare ahead of going public. Observers have noted the cuts could reflect a grab for developer lock-in rather than pure efficiency gains. The timing invites that reading. Anthropic is expected to list its shares within weeks, while OpenAI has reportedly delayed its own listing until next year. Cheap models win developer mindshare; mindshare wins the growth narrative investors pay for.

Then there’s China. The labs also face pressure from low-cost Chinese models, which have been undercutting Western pricing for over a year. Nobody at OpenAI or Anthropic will say they’re pricing against Chinese competitors, but the pressure doesn’t need to be named to be real.

Why this matters

If you’re a developer or a startup founder, this is the best pricing environment AI has ever offered. Token costs that used to throttle experimentation now barely register for most workloads. That changes what gets built: longer agent runs, bigger contexts, more ambitious prototypes that would have been too expensive a year ago.

But the AI price war cuts two ways. Open-source models built their reputation on being the cheap, self-hostable alternative. If GPT-6 Luna costs ten cents per million input tokens, the cost argument for running your own model gets thin. Masad’s shift at Replit is the canary: when even the believers start rebalancing toward closed models, the open-source ecosystem has to compete on something other than price. Independence, data control, and freedom from one company’s roadmap become the pitch.

Watch where this goes next. Claude Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks, and OpenAI isn’t likely to let a rival have the last word on price. The labs are racing downward, and the open-source community is watching to see if there’s a floor — or if the floor is the point.

FAQ

What is the AI price war between OpenAI and Anthropic? On September 22, 2026, OpenAI launched GPT-6 Sol and Luna at 50% lower API prices, and Anthropic launched Claude Opus 5.5 about 40% cheaper on typical workloads — both within hours of each other. The near-simultaneous launches, each built around lower prices, kicked off direct price competition between the two leading labs.

How much do GPT-6 Sol, GPT-6 Luna, and Claude Opus 5.5 cost? GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens. GPT-6 Luna costs $0.10 input and $0.50 output. Claude Opus 5.5 costs $4 input and $20 output, with cache reads 60% cheaper than Opus 5 at $0.20 per million tokens.

Why is Replit using less open-source AI? Replit CEO Amjad Masad said closed-source labs have aggressively cut prices for their smaller, faster models, which has eroded the cost advantage open-source models traditionally held. Replit now rents more capacity from OpenAI, Anthropic, and Google instead.

Will AI model prices keep falling? Unclear. The labs credit efficiency gains like better caching and leaner models, but analysts also see a fight for developer mindshare ahead of stock listings — Anthropic is expected to float within weeks — and pressure from low-cost Chinese models.

Sources: Reuters, TechCrunch, TechDefused, TechRepublic