Every Major AI Price Cut (February 2026)

AI model pricing changes constantly. Vendors cut prices, launch cheaper alternatives, and shift their lineups — often without announcement. What some are calling an AI pricing war is playing out across every major provider in 2026. If you're running models in production, you need to know when the math changes.

This page tracks every major AI API pricing change in February 2026 across OpenAI, Anthropic, Google, DeepSeek, and others. We pull from our database of 412 models across 43 vendors, synced daily. All per-request costs assume 5,000 input tokens and 1,000 output tokens — a typical API call.

February 2026

The month the Opus price ceiling dropped and new budget frontiers arrived.

February was the biggest shake-up of 2026 so far. Anthropic's Opus went from being the most expensive frontier model (at $15/$75) to a mid-tier price point. Google dropped the highest-IQ model in our database at a price that undercuts both Opus and GPT-5. And Xiaomi set a new floor for frontier intelligence at a tenth of a cent per request.

The cumulative effect

Each vendor's best value frontier model as of February 2026, ranked by cost per intelligence point. This table shows the result of all pricing moves combined.

The top three value leaders are all Chinese companies: Xiaomi, DeepSeek, and MiniMax. The gap between the best (Xiaomi at $0.000019/IQ) and worst (xAI at $0.000723/IQ) is 38x. Even within Western vendors, there's a 7x gap between OpenAI's best value ($0.000079) and Anthropic's ($0.000580).

What the Opus cut means in dollars

The Opus 67% price reduction was the single largest price cut from a Western vendor in 2026. Here's what it looks like at production scale.

At 10,000 requests per day, the Opus generation upgrade saves $30,000 per month. That's $360,000 per year — just from being on the latest version of the same vendor's model.

But the savings get more dramatic if you compare across vendors. MiMo-V2-Flash (IQ 41.5) at the same 10K daily volume costs $240/month. That's $44,760 less than new Opus and $14,760 less than even MiMo's cost is a fraction of any Western frontier model.

What this means in production

Price cuts only save you money if you actually switch. And switching has costs:

• Testing. You need to validate that the new model handles your specific workload without quality regressions. Benchmarks measure general capability, not your use case.
• Migration effort. Different models have different API quirks, function calling behaviors, and output formats. Budget engineering time for the transition.
• Monitoring. After switching, you need to track quality metrics to catch regressions that benchmarks didn't predict.

The teams that benefit most from price cuts are the ones that already track their costs per model and per customer. When a new pricing option appears, they can calculate the savings against their actual usage — not a hypothetical. The MarginDash cost simulator does exactly this: reprice your real token usage against any model in the database. Sign up free to try it with your own data.

The 2026 AI pricing war in context

The AI pricing war in 2026 isn't a single event — it's a pattern of compounding pressure from three directions at once.

Chinese vendors set the floor. Xiaomi, DeepSeek, and MiniMax now offer frontier-tier intelligence (IQ 40+) for under $0.003 per request. These prices aren't promotional — they reflect structural cost advantages from efficient architectures (mixture-of-experts, distillation) and aggressive go-to-market strategies. Western vendors can't ignore a 30–225x price gap on comparable benchmarks.

Incumbents respond with generational leaps. Instead of cutting list prices on existing models, Anthropic shipped Claude 4.5 Opus at $5/$25 — a 67% cut from the previous generation. This is both a direct price cut and a quality upgrade: you get more intelligence for fewer dollars.

Google pushes from the top. Gemini 3.1 Pro at IQ 57.2 and $2/$12 is both the smartest and one of the cheaper frontier models. It puts pressure on Anthropic and OpenAI from above: why pay more for lower IQ when $2/$12 gets you IQ 57.2?

The net effect: the cost of frontier intelligence is dropping fast, but it's happening through new model launches rather than retroactive price cuts. If you're not evaluating new models as they launch, you're overpaying by default. We track every model launch and pricing change in our database of 412 models — see The Cheapest AI Models That Are Actually Good for the full ranking across all quality tiers.

Methodology

All pricing from the MarginDash model database: 412 models across 43 vendors, synced daily. Intelligence Index scores from Artificial Analysis. Cost per request calculated at 5,000 input tokens + 1,000 output tokens. List prices only — batch, cached-input, and volume discounts not included.

See our methodology page for the full breakdown of data sources and calculation methods.

Last updated: February 15, 2026