Coding is the most honestly measured field in AI: SWE-bench Verified hands every model the same 500 real GitHub issues and counts how many it actually fixes. This ranking follows the independent like-for-like runs (Epoch AI's table and public aggregators) rather than vendor-reported numbers, which mix different agent scaffolds and are not comparable. One honest caveat: the benchmark is close to saturated at the top, and Moonshot's open Kimi K3 even topped LMArena's frontend coding arena in October 2026 — the field below the top three is moving fast.
Anthropic's thoughtful AI assistant for writing, coding, and reasoning.
Anthropic's models hold the top of the independent SWE-bench Verified table and power the Claude Code agent — the default choice for serious repository work.
The world's most popular AI chatbot.
A close second on SWE-bench Verified, with OpenAI's Codex agent built around the same models for autonomous coding sessions.
Google's AI assistant woven into Search, Gmail, and Android.
Third in Epoch AI's like-for-like SWE-bench runs, and first-rate at reasoning over very large codebases thanks to its long context window.
xAI's witty AI assistant
Frontier-level coding from xAI — Grok tops the Aider Polyglot board in public aggregator tables and ships inside the Grok apps.
Alibaba's capable AI chat assistant
The strongest open-weight coding line: Qwen3-Coder posts near-frontier SWE-bench and Aider numbers and can be self-hosted.
The AI-first code editor built for pair programming with AI.
Not a model — the AI code editor most of the models above get used through. Ranked here as the leading coding product built on them.
Build software collaboratively
The product pick for shipping without a local setup: describe an app and Replit's agent builds, runs and deploys it in the browser.
Modern terminal with AI
The agentic terminal: Warp runs multiple coding agents in parallel inside a modern terminal, a different surface for the same frontier models.
Ranking basis: SWE-bench Verified (independent runs compiled by Epoch AI and public aggregators) and LMArena coding arenas. Order reflects the consensus of those public boards as of the check date. We publish ranks, not scores, because scores from different harnesses are not comparable. Data checked October 2026. Rankings are editorial, ordered from the public sources named here; Tools Hub does not accept paid placements on this list.
A benchmark of 500 real, human-verified GitHub issues from open-source projects. A model gets the repository and the issue text and is scored on whether its patch makes the project's tests pass. Because every serious lab reports it, it became the common yardstick for coding models — though the top scores are now close enough together that the benchmark is nearing saturation.
On independent like-for-like SWE-bench Verified runs, Anthropic's Claude models lead, with OpenAI's GPT line a close second and Google's Gemini third. In practice the three are close enough that your tooling — editor, agent harness, price — matters as much as the model.
Increasingly, yes. Alibaba's Qwen3-Coder and Moonshot's Kimi K3 post scores near the proprietary frontier, and Kimi K3 topped LMArena's frontend coding arena in October 2026. For self-hosting, cost control or privacy-sensitive codebases they are a serious option.