Open models tighten the screws on Anthropic

Two new foundation models landed within days of each other, and both could sharpen the pressure on Anthropic. According to the source, Moonshot Labs' Kimi K3 and Alibaba’s Qwen 3.8 are both close to Anthropic’s Fable 5 in performance, and both are expected to release their model weights publicly in the coming weeks.

That matters because the economics of foundation models are brutal. Training demands heavy spending on researchers, compute, and electricity. After launch, the biggest ongoing cost is inference — serving model responses to users. For companies that lease infrastructure rather than own it, those costs scale directly with usage, which limits margin expansion.

The source frames three broad approaches:

• Lease data centers and power, as it says Anthropic, Knowledge Atlas and Moonshot Labs do
• Own data centers but buy power, which it attributes to Meta and Alibaba
• Own both generation and data centers, as it says SpaceX does

The core argument is straightforward: the more of the stack a company owns, the more inference costs become fixed rather than variable. That creates room for higher margins, and also opens other businesses, such as hosting open models or leasing compute capacity.

Recommended reading

Anthropic’s $1.5 billion author settlement gets court approval

Why Anthropic looks exposed

The source argues that Anthropic is in a particularly weak position if open and low-cost rivals keep improving. It says the company has leaned heavily on regulation and recursive self-improvement, while also keeping a strong focus on model safety and self-censorship in Fable and Mythos.

It also points to pricing pressure. Citing Figure 1, the source says Fable 5 is nearly 3× as expensive per completed task as OpenAI or open models. If buyers start optimizing harder for price, that gap could become difficult to defend.

The source also sees product risk. Anthropic has backed products such as Claude Code and Cowork, but says tools in that layer face intense competition from projects including OpenCode, OpenClaw, Hermes, and other harness startups. By contrast, it argues OpenAI has stronger options in consumer product design, publishing, voice, hardware, and potentially infrastructure ownership.

Kimi K3 and Qwen 3.8 add to a pattern

The timeline is tight: Kimi K3 was released on July 16, Qwen 3.8 was announced on July 19, and GLM 5.2 arrived in mid-June. The source argues this is bigger than the “DeepSeek moment” of 2025" because it suggests repeated catch-up by multiple labs, not a one-off surprise.

If that pattern holds, the biggest winners may be companies that pair competitive models with owned infrastructure or stickier products. The most exposed, the source argues, are model-only providers such as Knowledge Atlas, Moonshot Labs, and especially Anthropic.