The open-model gap is not one gap.

The short answer: choose a proprietary frontier model when the cost of failure is high and maximum capability matters. Choose open weights when control, privacy, customization, or high-volume economics matters more than winning every benchmark. Most serious systems should route between both rather than turn the choice into an ideology.

Epoch AI estimates that the best open-weight models have lagged the closed frontier by an average of four months since January 2026, equivalent to an average eight-point gap on its composite ECI. The same analysis warns that public benchmarks may flatter open models because public test sets are easier to optimize against. Read the Epoch AI methodology before treating “four months” as a law of nature.

Benchmark by benchmark, the frontier looks different

Composite scores are useful for orientation, but they erase the shape of capability. A model can be close on fresh code generation, farther behind in a terminal, and competitive again on one narrow agent task. The heatmap below compares the best available score in each access class—not one hand-picked pair of models.

Near-parity on a benchmark does not mean interchangeable products. Reliability, tool schemas, context handling, multimodality, safety controls, provider uptime, and agent scaffolding all sit outside a single score. It does mean that “open models are always much weaker” is no longer a useful selection rule.

Coding gives two contradictory answers—and both can be right

Open models often look closest on code-generation tests. That is not the same as replacing a frontier coding agent. LiveCodeBench asks a model to produce solutions to fresh programming problems. Terminal-Bench asks it to keep state, use tools, inspect failures, and complete longer tasks in an environment. Repository repair introduces another layer: search, patch planning, test interpretation, and recovery.

The official SWE-bench leaderboard makes this distinction explicit. Its full table mixes agents, retrieval systems, multi-rollout methods, and review loops; its separate bash-only view exists to compare base language models under a more consistent scaffold. A higher agent score can reflect a better model, a better harness, or both. See the SWE-bench Verified comparison guidance.

The economic gap runs in the other direction

The closed frontier usually wins the capability ceiling. Open-weight APIs and self-hosted deployments can win on marginal cost, provider choice, cache economics, and the ability to optimize a stable workload. But “free weights” do not mean free inference. Hardware, engineering time, observability, batching, idle capacity, and failure recovery belong in the denominator.

GLM-5.2 is a useful example. Artificial Analysis placed it on the intelligence-versus-cost Pareto frontier, but also found that it consumed more output tokens per task than several open peers. It can be inexpensive per unit of intelligence and relatively token-inefficient at the same time. That is why production comparisons should use cost per completed task, not just price per million tokens. See the Artificial Analysis GLM-5.2 evaluation.

Most “open-source LLMs” are really open-weight LLMs

We use “open source” in the heading because that is how people search. The technical distinction matters. Open weights let you download and run model parameters. Fully open systems may also publish training code, data, evaluation harnesses, and a license compatible with the Open Source Definition. Commercial use, redistribution, derivative models, and acceptable-use restrictions vary by family.

Treat the license as a deployment requirement, not a footnote. A model can win the benchmark and still be the wrong choice if its terms, hardware footprint, or auditability do not fit the organization.

Which should you choose?

If you want the actual open-weight leaders rather than the category argument, use the live open-source LLM ranking. For a workload-specific decision, compare finalists side by side and run a private eval drawn from your own failure cases.

Methodology and limits

• We compare the highest observed open and proprietary score for each metric in the WhatLLM data loader.
• Artificial Analysis is the primary independent benchmark source; not every model has every score.
• Different reasoning levels and model variants can have different cost and performance.
• Public benchmark parity does not establish production parity or causal superiority.
• The four-month estimate is Epoch AI’s composite analysis, not a WhatLLM measurement.
• Scores change. The page exposes the fallback snapshot date so readers can judge freshness.