From Ars Technica · · 3m
Open models are 4.4 months behind, at a fifth of the price
Chinese open-weights models have nearly caught the US frontier, and the math on paying up just got narrow.
In brief
Chinese open-weights models have nearly caught the US frontier, and the math on paying up just got narrow. Pay for a closed frontier model only when a deadline lands before the open models catch up, which is roughly four months out. Originally reported by Ars Technica.
The gap is 4.4 months
A new Mozilla report puts the distance between US frontier models and the best Chinese open-weights models at just 4.4 months. That is the whole lead you are paying a premium for.
Three points, a third of the price
Moonshot AI's Kimi K3 scores within three points of Anthropic's Fable 5 on the Artificial Analysis Intelligence Index, while costing 30 percent of it.
Open should be your default
Mozilla's State of Open Source AI report argues most organizations should run open models for the bulk of their work and treat closed frontier access as the exception.
Where closed still wins
Mozilla CTO Raffi Krikorian names three: expert professional work, high-intensity retrieval, and long context. The call is about the workload, not the company.
“We see the decision to pay for closed [models] as workload-specific rather than organization-specific.”
Open weights, closed recipe
You can download the weights and run them yourself, but labs still hold back training data, the pipeline, and the training code. Closed US models keep all of it proprietary.
You are buying the wrapper
Closed models work out of the box with compliance packaging, support, and someone to blame. Plenty of companies simply lack the staff to run open weights well.
DoorDash splits the work
The delivery company runs Kimi for routine jobs and saves Fable for the hard tasks that would take a human expert a long stretch. Two tiers, one workload router.
Seven hours versus twelve
METR measures a model's time horizon: how long a task a human expert would need, done reliably. The best closed model handles jobs 1.7 times longer than the best open one.
“If the open frontier can handle a seven-hour job, the closed frontier can handle a 12-hour one.”
The eight to twelve hour band
Under eight hours, either kind of model works, so send it to the cheap one. Over twelve, nothing is reliable yet. The premium only buys you that middle slice.
Same harness, same score
Labs build custom harnesses that flatter their own models. Vals AI ran everyone on a neutral one: on Terminal-Bench 2.1, Z.ai's GLM 5.2 landed within a point of Claude Opus 4.7 and 4.8, at a fifth of the cost per task.
Open runs, closed earns
Eight of OpenRouter's top 10 models by August 2026 token volume ship open weights. Yet a Linux Foundation paper found open models took just 4 percent of revenue, on data through September 2025.
The Android playbook, again
Krikorian's worry is not that the open models are Chinese, it is that they are all Chinese. Give it away, own the ecosystem. He wants US and EU labs competing in the same open lane.
“The Chinese labs are running the same playbook the Americans ran with Android.”
Who funds the commodity layer
His fix borrows from Linux: public compute like Switzerland's Apertus, neutral foundations holding the standards, and philanthropy paying for the evaluation and audit work nobody else will.
“It's hard to fully trust a model with decisions if you can't tell how it was trained or what it was evaluated against.”
What matters most
Pay for a closed frontier model only when a deadline lands before the open models catch up, which is roughly four months out.
