The ToolsGuide Score
Every text, vision, coding and agent model gets a ToolsGuide Score out of 1,000. The scale is fixed: 1,000 is a flawless model, one that would get every question right on every test in the category's battery. No model is close yet: today's best models score in the 600s and 700s. When models get better their scores rise, and the scale itself never moves.
How it's built
Each category has a fixed battery of hard tests (listed below). Model makers publish results on different tests, so from every result a model has, we estimate how it would do on the whole battery, allowing for how hard each test is. That lets models that took different tests be compared fairly.
People's votes from blind head-to-head comparisons are converted onto the same scale, using the models that have both votes and test results. The two are then blended. A model with fewer than two test results is scored on votes alone, and leaderboards show both parts so you can see what each score rests on.
Text · votes 50%, tests 50%
ARC-AGI v2 · Humanity's Last Exam · LiveBench · MRCR v2 (8-needle)
Vision · votes 50%, tests 50%
ARC-AGI v2 · Humanity's Last Exam · MMMU-Pro
Coding · votes 40%, tests 60%
DeepSWE 1.1 · FrontierCode 1.1 · NL2Repo · Program Bench · SWE-Bench Pro · Terminal-Bench 4.0
Agents · votes 30%, tests 70%
Agents' Last Exam · AutomationBench · BrowseComp · Terminal-Bench 4.0 · Toolathlon
Preference score for image and video
No image or video generation tests have published results yet, so these models get a Preference score instead: how often voters would pick the model over a typical one, out of 1,000. 500 is a typical model; 1,000 is a model voters would always pick. It measures what people prefer, not correctness, so it can't be compared with the ToolsGuide Score. Where two independent sets of votes cover the same models and agree, they're combined.
Scores by task
The "Leaders by task" tables and picks such as "Best for coding" use votes on that kind of prompt alone, converted onto the same fixed scale.
Benchmark scores
The related benchmark charts show each test's own result, out of 100%. Every result carries a status: verified, independent (measured by someone other than the model's maker) or self-reported (published by the maker). Most results are self-reported.
Prices
Prices are in US dollars: per million input tokens for text, vision, coding and agent models, per image for image models and per minute for video. When a model is sold by several providers, we show the lowest listed price.
Combining sources
Votes, test results and prices come from different publishers. We link the records for the same model by its name, with effort settings such as "(High)" set aside; 479 ranked models are linked this way. Product tiers such as "Max" or "Flash" count as different models.
Keeping the scale fixed
The battery and the numbers that convert results onto the scale stay the same from week to week, so a model's score only changes when its own results do. When models start to max out a test in the battery, we move to a harder battery under a new method version and say so here. Current version: tg-score-2.
Updates
The data is refreshed every week. The date at the top of each page shows the latest update.