The AI Model Pricing War Is Here — And Your Margins Depend on Picking the Right Side

Three things happened in quick succession this month that changed the math for anyone building AI-powered products.

On February 11, MiniMax released M2.5—an open-source model that matches Claude Opus 4.6 on the SWE-Bench coding benchmark at 1/20th the cost. A million tokens that costs $15 with Claude Opus runs at $0.75 on MiniMax. On February 14, comparisons started circulating showing DeepSeek V3.2 delivering near-GPT-o1 reasoning quality at $0.25/M input tokens—140x cheaper than o1's $35/M. Then on February 17, Grok 4.20 Beta launched with a rapid-learning architecture that improves weekly, priced at $0.20/M.

The AI pricing floor didn't slowly erode—it fell through. And for indie hackers, this changes everything about which products are now economically viable.

Here's a real-world example. Say you're building an AI writing tool. A user generates 10 articles a month, each requiring roughly 50,000 tokens of processing. That's 500,000 tokens per user per month.

At Claude Opus prices ($15/M input), each user costs you $7.50/mo in API costs alone. If you're charging $29/mo, your gross margin on AI costs is 74%—before hosting, support, or anything else. Tight, but workable.

At Grok or MiniMax prices ($0.20/M input), that same user costs $0.10/mo. Your gross margin on AI costs jumps to 99.6%. You could cut your price in half and still be more profitable than the Claude-powered version.

This isn't just about saving money. It's about which business models become viable. Freemium, unlimited plans, high-volume tools, consumer products at $5/mo —none of these worked with 2025 pricing. They work now.

Here's what the "use the cheapest model" crowd tends to gloss over: model lock-in. xAI is pricing Grok at $0.20/M to win market share. That price is subsidized by Elon Musk's backing and xAI's ambition to dominate the market. Once you've built your product around Grok's API, prompt format, and behavior, switching costs are real.

The same applies to DeepSeek. Its pricing is aggressive. But all data routes through servers in China. For any product handling customer data, contracts, or financials, that's a compliance and trust problem waiting to happen.

The practical answer isn't to avoid cheap models—it's to build with an abstraction layer. Tools like LiteLLM, OpenRouter, or PortKey let you swap providers with a single config change. You get cheap prices today, and flexibility when the market shifts tomorrow.

The price war isn't slowing down. Google just slashed Gemini Flash pricing again. MiniMax is open-sourcing models that match frontier performance. DeepSeek V4 is expected in the coming months. Every month, the cost per token is lower than the month before.

The upside for indie builders is real: AI-powered products are becoming structurally cheaper to operate, which means more business models work, more price points are viable, and more niches are worth serving. A product serving 500 users that would have lost money at 2024 prices is profitable in 2026.

The risk is commoditization on the layer you're building on. If cheap models can do what you do, your moat had better be distribution, data, or domain expertise—not just "we wrapped GPT."