The great AI price war: How Chinese open models are squeezing GPT-5.6 Sol and Claude Fable 5

The age when the frontiers of artificial intelligence research belonged to select American laboratories was one of great monopoly, where OpenAI and Anthropic set the trend and price standards, and others in the industry followed suit. All that has changed in July 2026, with two Chinese laboratories producing models that compete with the likes of Silicon Valley on benchmark performance but do so at a much lower cost, with perhaps most critically, access to the weights. This is not a tale of how China is catching up. This is a tale of how China has rewritten the rules of the game altogether.

Also read: OpenAI’s Astra Model solved 10 open math problems: This is how we know it’s true

The premium tier: Sol and Fable 5 play it closed

GPT-5.6 Sol by OpenAI and Claude Fable 5 by Anthropic occupy the top rows of most benchmark tables. Sol has a slight edge in the Coding Agent Index of Artificial Analysis, and both are on top in GDPval-AA v2, which is an evaluation benchmark based on real-world applications across 44 occupations.

However, both are expensive closed models. Sol costs $5 for every million input tokens and $30 for every million output tokens in the API pricing, but input cache costs only $0.50 per million tokens. Fable 5 is similarly priced at the frontier tier. None comes with model weights. You pay for access, but you do not have the model, and you never will. This used to be the price you had to pay for frontier-level AI. No longer.

The challengers: cheaper, open, and closing the gap fast

The Moonshot AI Kimi K3 came out on July 16 as, according to the organization itself, the biggest open-source model to be released: 2.8 trillion parameters, sparsity mixture of experts that activates only 16 out of 896 experts per token, and a massive 1-million-token context window. The cost of using it is $3 for every million input tokens and $15 for every million output, with a cached input at $0.30. According to GDPval-AA v2, its score is third overall, right below Fable 5 and Sol, and above Claude Opus 4.8.

Also read: 65 pct of Samsung’s flagship sales come from Tier 2 and Tier 3 markets, says JB Park on India’s foldable future

Along came DeepSeek, which approached the goal in a very different manner. The DeepSeek-V4-Flash-0731, launched on July 31, did not scale up the model. Instead, it maintained the same architecture of 284 billion total parameters and 13 billion active parameters of the previous version and only retrained the model more intensively on data related to agentic and reasoning abilities. The result is a performance score of 82.7 on Terminal-Bench 2.1 compared to 61.8 in the previous version and 7.3 to 54.4 on DeepSWE. All of this at a cost of $0.14 per million input tokens and $0.28 per million output tokens under an MIT license.

Read those two pricing numbers again. DeepSeek’s flagship agentic model costs roughly 1 percent of what Sol charges for output tokens. That is not a rounding difference. That is a different economic category altogether.

Why this matters more than the benchmark charts

It might be tempting to dismiss this as China’s desperate attempt at keeping up with price, since it is not capable of matching in raw capability. However, this does not match the data from the benchmarking exercise. First, Kimi K3 is defeating Claude Opus 4.8, which until recently was believed to have frontier-level capabilities. Second, the Flash model created by DeepSeek is defeating its larger version, V4-Pro. Both of these organizations do not appear to be selling a second-rate product. They are selling a frontier-like capability at commodity prices since they do not have a subscription-based revenue model like OpenAI and Anthropic.

There’s the accessibility too. Kimi K3’s weights are accessible. DeepSeek’s model is MIT-licensed, which means that any firm from anywhere in the world, even India, can download it, fine-tune it using their proprietary data, and deploy it on their own infrastructure, completely free from reliance on a US API, a US data-retention regime, or a US export control regime. This isn’t an abstract point. The release of GPT-5.6 required two weeks of vetting from the US government before OpenAI could release it. Anthropic’s own Fable 5 model had to be taken down for three weeks in June under export controls from the Department of Commerce. When your American frontier model can be shut down by Washington overnight, the case for an alternative writes itself.

The market share math is already shifting

It is not an academic exercise in pricing. Moonshot’s previous open release, Kimi K2.6, reached the second-most popular position among the models on OpenRouter, ahead of virtually all Western frontier labs in independent usage metrics, within months of its release. Moonshot is apparently raising up to $2 billion in funding, valued at $31.5 billion precisely because of its open-source momentum. Every developer who uses a $0.14 input model versus a $5 input model is a developer OpenAI and Anthropic do not bill.

American labs currently hold an advantage at the upper end of the capability curve. However, “currently” is doing a lot of work in this sentence, and the gap between them and the rest of the world is narrowing with each release cycle. While Sol and Fable 5 may be the definitive versions when it comes to raw capabilities, Kimi K3 and DeepSeek V4 Flash are the definitive versions when it comes to intelligence per dollar, and the latter metric is ultimately what will drive investment decisions for most companies using AI at scale.

There is no price war on the horizon. The price war is already upon us, and it is being fought by American companies.