Mozilla's inaugural State of Open Source AI report says the performance gap between Anthropic's closed frontier model Fable 5 and Moonshot AI's open-weight Kimi K3 has narrowed to about four months, at a fifth of the cost.
Four months. That's roughly how far ahead Fable 5 sits over Kimi K3 on raw capability, according to Mozilla's report, released in mid-July. Kimi K3 scores 57 on the Artificial Analysis Intelligence Index. That's just three points behind Fable 5, at about 30 percent of the price. A year ago, a gap that small between an open-weight model and a frontier one from Anthropic or OpenAI would have been unthinkable. For any founder wondering whether frontier AI is worth the money, that gap is starting to look like the whole answer.
The report, built on Mozilla's own benchmark analysis and a survey of more than 950 developers, backs up that number. It points to a broader trend. GPT-4-class inference now costs roughly $0.40 per million tokens, down from $20 three years ago. That's a 50x drop. Open-weight competition is the reason. When Moonshot AI, DeepSeek, and a growing field of Chinese labs started shipping models within striking distance of GPT-4 and Claude, every provider with a paid API had to reprice or lose customers.
Here's the part that should worry the labs charging premium prices. Open models now handle roughly a third of real-world AI usage, per Mozilla's report, but they capture only 4 percent of the revenue in the market. Usage has decoupled from spend. Developers are running open weights in production and paying frontier prices for a shrinking slice of their workloads.
Not nothing, though. Kimi K3 ranks third on the Artificial Analysis Intelligence Index, behind Fable 5 and OpenAI's GPT-5.6 Sol, and Mozilla's researchers found open models still trail on advanced reasoning, long-context retrieval, and agentic tasks, the multi-step, tool-using work that's hardest to fake. Coding and instruction-following are close to solved. An agent that has to plan a task, call five tools in sequence, and recover from its own mistakes is still a different story. That's the gap that hasn't closed.
Harvey Built Its Own Legal AI Model Instead of Renting One From OpenAI Harvey, the legal AI startup valued near $15.5 billion, has launched Tenet, its first proprietary model, built on a customized Kimi K3 base and post-trained on attorney-generated case files. The model ships inside a broader Harvey II relaunch that also introduces a Memory feature for law firms. - why legal AI startups build their own models - how law firms fine-tune AI with case files
That's the real decision founders are making right now, whether they realize it or not. If the workload is drafting, classifying, summarizing, or writing code to spec, Kimi K3 or a similar open model probably does the job at a third of the cost. If it's an autonomous agent making judgment calls with real consequences, the 5x premium still buys something. Frankly, most startup workloads fall into the first category, not the second. And a lot of AI budgets haven't caught up to that fact yet.
The ceiling is rising while the floor collapses #
There's a wrinkle worth watching. While the cheap end of the market has cratered, the top end hasn't. OpenAI's GPT-5.6 Sol now costs $5 per million input tokens and $30 per million output tokens, roughly double what frontier pricing looked like back in January. The market is splitting in two. Commodity-grade intelligence gets cheaper every quarter, and true frontier capability gets more expensive, not less. Anthropic, OpenAI, and Google are betting that the sliver of work only a frontier model can do will keep paying for itself even as the rest of the market heads toward zero margin.
China is where the pressure is building fastest. Mozilla's report puts open-source AI adoption across China and East Asia at 89 percent, and Chinese open-weight models went from under 2 percent of weekly token traffic on OpenRouter in late 2024 to more than 45 percent by this April. That's not a niche trend. That's half the market.
One caveat before anyone moves their whole stack to open weights. Mozilla's survey found only 53 percent of teams running open models actually get them into production, versus 63 percent for teams running closed ones. The intelligence gap has closed. The deployment gap hasn't.
Also read: AIUC Wants To Insure Your AI Agents Before They Go Rogue • Todd Blanche Says the DOJ Won't Regulate AI Companies Through Prosecution • Meta One Puts a Price Tag on AI Power and Social Reach Up to $49.99
This article is posted in AI News, check it out for more related stories.
Moonshot's 2.8 Trillion Parameter Kimi K3 Just Ran on an Ordinary MacBook Pro An open source project called WASTE ran Moonshot AI's full, unpruned 2.78 trillion parameter Kimi K3 model on a 64GB MacBook Pro, streaming most of its weights off an SSD. It worked, but at only 0.3 tokens per second, and it echoes a similar hack run on the smaller Kimi K2 back in March. - large language model runs on MacBook - open weight model inference locally
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.