Open models crushed prices. Anthropic and OpenAI now compete on cost

Thesis: the MIT license and inference optimizations made price the main battleground, not benchmarks. That is good news for users and bad news for a business model built on scarce access.

Anthropic and OpenAI shipped their new models, Opus 5.5 and GPT-6 in the Sol and Luna variants, with the same message: faster and cheaper. Not "an order of magnitude smarter," not "closer to AGI," but cheaper. That is no accident. Open models have started setting the price ceiling. Paid access now has to defend itself with features rather than with its own existence.

The best evidence is DeepSeek V4.1 Flash. Released on 10 September 2026, it uses a MoE architecture with 552 billion parameters, of which 8 billion are activated in prefill and 16 billion in decoding. The model natively handles images up to 384 tokens. Its key-value cache takes 890 bytes per token, roughly four times less than in the previous V4 version, and it holds a context of one million tokens. All of it under the MIT license. That last piece of the puzzle is underrated. The MIT license means any operator can host the model on its own infrastructure, with no negotiations, no fees and no permission required.

On top of that comes a layer of optimization that lowers cost without changing the model's weights. The Chinese team 是石科技 presented the Meta-Infer engine. Software optimization alone, without touching the hardware, raised DeepSeek's inference throughput almost sevenfold. One and a half accelerators of the 6000D class were said to outperform a single B300. The company manages more than 20,000 PFLOPS of capacity and ten data centers. The bottleneck is the hardware abstraction layer. When operators reach for generic kernels, throughput drops by an order of magnitude. This is engineering work with a return counted in money, not in points on a leaderboard.

The scale of the business confirms the direction. DeepSeek is to exceed 1 billion dollars in annualized revenue, up from below 500 million a few months earlier, with a second funding round worth 7.5 billion dollars, a target valuation of 500 billion yuan and preparations for a listing in Shanghai. More than 70 percent of the company's compute goes to training new models, leaving less than 30 percent to serving the ones already out. Management says raising API prices did not cost it customers. It is a portrait of a vendor that does not have to win on price, because it wins on the availability of the weights.

The consequence for the market is unambiguous: the era in which the most expensive model was the default choice is ending. Organizations are starting to treat models as a portfolio of assets: cheap open weights for bulk tasks, expensive frontier models for a narrow, critical fraction of the work. Paid vendors can respond in only one way: with even deeper cuts, or with features an open weight cannot provide, such as auditing, guarantees or integration. "More expensive because better" has stopped being an argument, and that is the healthiest change in this industry in years.