What the comparison measures
Each model record separates the model identity from the provider route. We record the exact model or provider ID when the source publishes it, the current version label, modalities and use-case tags, free access type, allowance summary, reset period, context window, card requirement, and the source checked date.
A free provider route does not mean the model is universally free. Access may depend on a provider account, region, plan, model availability, or a temporary limit. When the source does not establish a fact, the record says Unknown.
Evidence labels
| Label | Meaning |
|---|---|
| Source verified | The linked provider or independent evaluator page directly supports the recorded claim. |
| Owner confirmed | Kem personally confirmed the fact; this is used only when that confirmation exists. |
| Needs rechecking | The source is stale, conflicting, unavailable, preview-only, or otherwise insufficient for a confident current claim. |
How benchmark evidence is handled
A benchmark observation is stored with its benchmark name, version, metric, score and unit, model version, source kind, methodology note, comparable group, caveats, and evidence label. Source kinds distinguish a provider-published result, an independent evaluation, and an owner-run result. No owner-run results are published because FreeAIList has not run those tests.
For four open-weight models, we use independent English MMLU-Pro, GPQA, and LiveCodeBench results from the Swallow LLM Leaderboard. Its 0–1 scores are shown as percentages; the source publishes its methodology and CC BY 4.0 license. GPT OSS results use the evaluator’s documented medium reasoning mode. The scores describe those evaluated checkpoints and settings, which may differ from a hosted or quantized route.
Epoch AI independently administers the OTIS Mock AIME 2024–2025 math test, with results for ten more catalog models. Its data is available under CC BY 4.0, and its test page explains the 45 questions and scoring. Each score links to the evaluated model's scorecard; some scorecards show the best evaluated setting, which may differ from a hosted free route. The separate provider-published Llama 4 Scout MMLU-Pro, ChartQA, and DocVQA results remain in their own comparison group. ChartQA concerns questions about charts; DocVQA concerns document images. Scores from different evaluation runs are never merged into one bar chart.
We do not calculate a universal intelligence, token-efficiency, or value score. A score is not a guarantee of writing quality, coding success, factuality, latency, privacy, or safety in a visitor’s workflow.
Missing, disputed, and stale values
- Missing quota, reset, price, or benchmark context stays Unknown.
- Conflicting provider or evaluation claims remain separate, with a caveat or Needs rechecking.
- Different benchmark versions, metrics, modalities, or model checkpoints are labeled incomparable rather than ranked together.
- HTTP 403 and timeout responses are manual-browser recheck signals, not automatic proof that a source is broken.
- Facts are checked against official model catalogs, documentation, pricing, billing, model cards, papers, or official announcements whenever available.
How to use the result
Use the comparison to shortlist two or four candidates, open the provider evidence, and test a small task representative of your work. Keep a fallback when the route is rate-limited or a preview changes. For a longer explanation, read How to read AI benchmark scores without being misled.