Benchmarks
Methods shared by scenario boards. A benchmark older than 90 days is downgraded and is not primary evidence.
LMArena WebDev2026-10-08 · FrontendFrontend web work: landing pages, interactive components, small games, and full pages, including agent flows that need several steps and tool calls.LMArena Image-to-WebDev2026-10-09 · FrontendGenerate a web page from an image or screenshot. The source page also includes agentic coding flows that need several steps and tool use.Android Bench2026-10-06 · MobileLong Android changes across modules: a whole feature, not a single edit.Terminal-Bench 4.02026-10-06 · Terminal agentsAgent tasks in a terminal: install a toolchain, edit config, run commands, and finish the automation.The Slop Index2026-10-11 · Docs and copyHow chatbot-like a model's writing is across email, blog/essays, tweets, and Discord chat. Higher scores mean less chatbot-like writing.sudo L72026-10-09 · General coding abilityGeneral coding ability: making changes in real codebases, spotting risk, verifying dangerous cases, and explaining the call. 60 tasks written by staff-level engineers, mostly from private production repositories, plus open source. On this site it is a general board and appears in both the Web frontend and Android scenarios.Supabase Evals2026-10-10 · Framework evalModels do auth, database, RLS, storage, and debugging work on Supabase. 23 tasks, 3 runs each, 69 checks.Next.js Agent Evals2026-09-25 · Framework evalModels do real Next.js tasks. The current board is the agents that are still rerun. The older Previously measured rows are not folded into this column.Convex Evals2026-10-08 · Framework evalModels write Convex code and the harness runs it. The current version has 112 tasks. The baseline is No Guidelines: no supplied guidelines and no web access.ReactBench2026-10-10 · Ecosystem evalWrite features and fix issues in real React repositories. v1.2 has 39 tasks, 37 repositories, and 76 configurations. Million runs it. The React team does not, and neither does Meta.DeepSWE v1.12026-09-22 · General long-horizon engineeringGeneral long-horizon engineering: feature-sized changes in real open-source repos. 113 tasks across Go, Python, TypeScript, JavaScript, Rust, and more, spanning 91 active repositories. The agent commits in an isolated environment and is graded in a pristine container. On this site it is a general board and appears in the Web frontend, Android, and CLI / DevOps scenarios.SWE-bench Verified2026-02-26 · RetiredPass rate on real GitHub issues.Aider Polyglot2025-11-20 · RetiredPass rate on polyglot code edits.