AI Pulse · Frontier Labs Watch

OpenAI dropped GPT-5.6 Luna pricing by roughly 80 percent last week. Two days later, DeepSeek fired back with V4-Flash-0731 at fourteen cents per million input tokens. That sits about 30 percent below the new Luna tier and seven times cheaper than Claude Haiku 4.5. This is not a promotional discount. It is a pricing war, and the frontier labs are now defending token margins against open-weight Chinese models that score close to them on agentic benchmarks.

For anyone running a business with AI baked into the product or workflow, this changes the math fast. If you built around a model that cost you five hundred dollars a month in tokens, you might now run the same workload for seventy. That opens up use cases that were marginal a week ago. Customer support bots, document summarisation at scale, real-time data extraction, all of it gets cheaper to run and easier to justify.

The catch is stability. Pricing this volatile means you cannot lock in a three-year cost model and forget about it. You need to watch which models your systems call, what they cost today, and whether a competitor just undercut that by half. This is exactly the kind of thing we build into an AI command centre at the Omni platform, so you are not manually tracking token rates across four providers every week.

The broader point is that frontier performance is no longer tied to frontier pricing. DeepSeek is open-weight, which means you can run it on your own infrastructure if the API price still feels too high. The Western labs are responding by cutting prices, not by locking down models. That is good for buyers, but it also means the cost advantage you have today might evaporate in a fortnight if you are not paying attention.