VENICE ATLAS · RANKING · FACTUALITY
Factuality ranking
Models for
factuality
Give accurate answers and avoid unsupported claims.
Compare performance, or switch to value for money to see how cost changes things.
Higher ratings mean better performance on this category’s tests. Individual test results appear alongside the overall rating.
| Model | Actions | ||||
|---|---|---|---|---|---|
| #1 | Qwen 3 235B A22B Instruct 2507Text · private | 20.46 | 54% | — | ↗ |
| #2 | Gemini 3 Flash PreviewText · anonymized | 17.95 | 69% | — | ↗ |
| #3 | GPT-6.1 SolText · anonymized | 15.49 | — | 74% | ↗ |
| #4 | InklingText · private | 12.19 | — | 44% | ↗ |
| #5 | DeepSeek V4 Flash 0423Text · private | 11.70 | 29% | — | ↗ |
| #6 | Google Gemma 3 27B InstructText · private | 8.87 | 10% | — | ↗ |
| #7 | GLM 5V TurboText · anonymized | 5.43 | — | — | ↗ |
| #8 | Mistral Small 3.2 24B InstructText · private | 4.90 | 12% | — | ↗ |
Find the balance
between performance and cost
Look for models that perform well without costing too much. Two models with similar scores may have very different costs.
For a similar amount of text sent and received: 50% input, 50% output.
Higher = better score. Further left = lower price.Each point is a model. Tap a point or choose a model below to see its numbers.
No models to plot for this selection. Try another usage profile or open the model table.
At a similar score, the model further left costs less. At a similar price, the higher model scores better.
How to compare performance and cost
Performance here means the model’s score on this category’s tests. Cost is an estimate per million tokens: the pieces of text you send and receive. Choose a usage profile that resembles your work.
Balanced averages input and output prices equally. For example, $2 input and $8 output gives an estimated $5 per million tokens. Longer prompts give input a weight of 80%; longer answers give output 80%; repeated context gives cache reads 80%, with 10% each for new input and output.
In the value-for-money ranking, a higher index means a better performance-to-cost trade-off within that usage profile. It is a comparison guide, not your final bill. Read the methodology.
Use these results as a guide
Rankings help narrow
the shortlist
Use a ranking to focus your search, then test the shortlisted models with your own workload.
Generated media models live in their original catalogs because no single LLM Stats score covers image, video, music, or voice endpoints well.