ModelsOpen-Weight vs Closed Models: What the Gap Costs You Now
Closed models still lead on aggregate benchmarks, but the gap narrowed sharply through 2026. What open weights buy, what they still cost, and how to decide.
The interesting question is no longer whether open-weight models are competitive. It is how much of the remaining gap you actually consume, and what you give up to close it.
As of August 2026 the top of the closed field still leads. Claude Fable 5 Max scored 1,815 on GDPval-AA v2, with GPT-5.6 Sol Max at 1,747.8. On the open side, Kimi K3 sits around 57 on the Artificial Analysis Intelligence Index and ranked first on Arena Frontend Code at 1,679 in blind developer voting — ahead of Fable 5 on that specific measure.
That last detail is the shape of the current situation in one sentence. Closed models lead on aggregate. Open models already win on specific axes.
What open weights genuinely buy
Deployment freedom. You can run the model on your own hardware. For data residency requirements, air-gapped environments, or regulated workloads where tokens cannot leave your network, this is not a preference — it is the entire decision, and it eliminates closed models outright.
Price. DeepSeek V4 Flash lists around $0.14 per million input tokens. No closed frontier model is close. Even mid-tier open models like GLM-5.2 undercut closed equivalents substantially.
No deprecation risk. A closed model can be retired, re-tuned, or quietly changed underneath you. Weights on your disk cannot. If you have pinned behaviour that took months to tune, that permanence has real value. Pinning model versions covers the general problem.
Inspection. You can examine the model, quantise it, fine-tune it, and measure it without rate limits or terms restricting evaluation.
What they still cost
Peak capability on long-horizon work. The remaining closed-model lead concentrates in extended unattended reasoning. On short tasks the gap is small; across forty tool calls it is not.
Operational burden, if you self-host. Kimi K3's weights are roughly 1.6TB and need multi-node serving. That is an infrastructure project with ongoing maintenance, not a configuration change. Most teams who use open models reach them through an API and never touch this — which means they get the price benefit without the deployment benefit.
Licence complexity. "Open weights" is not one thing. GLM-5.2 and DeepSeek V4 are MIT with no conditions. Kimi K3 requires a separate agreement for model-as-a-service businesses above a revenue threshold. Read the licence for the specific model — open weights vs open source covers why the distinction bites exactly when you succeed.
The comparison most teams should actually run
Not "which is better" but "where does the cheaper option stop being good enough".
Take twenty tasks spanning your real range. Run them through a strong open model and a closed frontier model. Record completion rate, turns, and whether you would ship the output unmodified.
What you are looking for is the boundary. In most workloads it exists and it is sharp: the open model handles everything up to some complexity threshold and degrades past it. Once you know where that line sits, you route — cheap model by default, escalate past the line.
That architecture captures most of the closed model's capability at a fraction of the cost, and it is the setup most cost-conscious teams converge on. Model routing and fallbacks covers building it.
Where the field is heading
The gap has narrowed on a consistent pattern: open models reach the previous frontier roughly a year or so after it was set, and the interval has been shrinking rather than growing.
Two consequences for planning. First, do not architect around a specific model — architect around an OpenAI-compatible interface so switching is configuration. Second, re-run your evaluation quarterly rather than treating a model decision as settled. In this field a conclusion from six months ago is not evidence about today.
A decision rule
- Regulatory or residency constraints — open weights, self-hosted. Nothing else qualifies.
- High volume, bounded tasks — open weights via API. The price difference is large and the capability difference is small.
- Long unattended agent runs where failure is expensive — closed frontier, or the strongest open model, and measure rather than assume.
- Mixed — route. Default open, escalate on a mechanical trigger.
The one position that is hard to defend in 2026 is using a closed frontier model for everything by default. The capability gap no longer justifies it across a whole workload, and the cost difference compounds daily.
Common questions
Are open-weight models good enough to replace closed ones?
For most bounded tasks, yes. The remaining closed-model lead concentrates in long unattended reasoning across many steps. Kimi K3 already ranked first on Arena Frontend Code, ahead of Claude Fable 5.
Does using an open model mean self-hosting?
No, and most teams do not. They reach open models through APIs, which captures the price advantage without the operational burden — but also without the data-residency benefit that self-hosting provides.
Are all open-weight models permissively licensed?
No. GLM-5.2 and DeepSeek V4 are MIT with no conditions. Kimi K3 ships under Moonshot's own terms with a revenue threshold for inference businesses. Read the licence for the specific model.
