A fresh model comparison for coding and document work
The leading candidates changed for coding and document work. Here's what moved and how to compare it with your current setup.
This report compares evidence observed on 2026-10-12 with 2026-10-19. It records changes to our decision guides, not every movement in every leaderboard.
A new release isn't required: newly published tests or revisions can change a comparison. These are changes in the evidence we observed, not proof that a model itself improved.
Coding for a small team (open weights)
The lead moved from Kimi K3 to DeepSeek V4.1 Flash. These are the highest displayed index scores in this mix; tied models share the lead.
If you use Kimi K3 for this work, DeepSeek V4.1 Flash is worth comparing: it scores 7.45 index points higher on 6 shared tests in the current evidence. That difference isn't a success percentage or a statistical confidence claim.
Compare a candidate with your current model on a recent bug and a feature that crosses files. Keep the prompts, tools and time budget the same. Check whether the tests pass, how much repair you do, and the actual cost before changing your default.
See the current guide or compare with your model.
Reading and extracting from documents
The lead moved from GPT-6 Astra to GPT-6.1 Sol. These are the highest displayed index scores in this mix; tied models share the lead.
If you use GPT-6 Astra for this work, GPT-6.1 Sol is worth comparing: it scores 0.16 index points higher on 4 shared tests in the current evidence. That difference isn't a success percentage or a statistical confidence claim.
Use a small set of your own documents with checked answers, including a difficult scan and a long file. Compare missing fields, wrong values and citations with your current model. Include retries and review time in the cost.
See the current guide or compare with your model.
Delayed evidence
Some inputs are retained from their last successful refresh: livebench. Their collection failure is not evidence that a model got better or worse. Check the source timestamps in the download.
Evidence and limits
The chart and downloadable evidence are frozen to this report. Live guide scores can change later. Each guide uses its declared capability weights and the same ranking method as For you. Published test settings and coverage vary; missing results are never filled in.
Download the observed evidence. It includes the previous and current leaders, their test configurations and source links, and the shared-test comparisons.
Reports are checked weekly and published when a guide's leaders change. Quiet weeks have no edition. Follow the reports by RSS or get the weekly email.
This report uses observed evidence snapshots. It suggests what to test, not a guaranteed improvement on your work.