Best AI model for each job, decided by the same tests

15 models share first place — joint 1st of the 22 models on the Combined board in the last 24 hours — from DeepSeek V4.1 Flash (95 ± 2.2) to Claude Sonnet 5 (89 ± 2.2), none of them measurably behind the leader. As of 26 Oct 2026, 10:51 UTC. The Combined board: coding 50%, reasoning 25%, tool use 25%, ranked with statistical ties. The board re-ranks after every run, so the answer on this page follows the leaderboard.

Coding: 5 models share first place — joint 1st of the 22 models on the Coding board in the last 24 hours — from DeepSeek V4.1 Flash (96 ± 0.4) to GPT-5.5 (94 ± 1.0), none of them measurably behind the leader. Reasoning: 14 models share first place — joint 1st of the 22 models on the Reasoning board in the last 24 hours — from Claude Fable 5.1 (100 ± 2.5) to DeepSeek V4 Pro (95 ± 2.5), none of them measurably behind the leader. Agents and tool calling: 13 models share first place — joint 1st of the 22 models on the Tool use board in the last 24 hours — from DeepSeek V4 Pro (91 ± 2.5) to GPT-5.6 Sol (85 ± 2.5), none of them measurably behind the leader. Each page shows the five best-placed models with statistical ties and the full board behind a free sign-in.

GPT-6 Luna is the best value: about $0.34 per 1,000 correct coding answers over the last 30 days, every attempt billed, the lowest among the 13 models within five points of the best on the Coding board in the last 24 hours. GPT-5.6 Luna is the fastest: a median response time of 3.3 s per coding attempt over the last 30 days, the lowest among the 13 models within five points of the best on the Coding board in the last 24 hours. Both count the same 7 coding tasks, every attempt billed.

The Combined board: coding 50%, reasoning 25%, tool use 25%, ranked with statistical ties. Every model gets the same tasks on the same schedule — coding every four hours, reasoning and tool use daily, with an hourly canary probe — and is graded by executing what it produces. Every score carries a standard error from its own run-to-run repeatability. Walking down a board, a model joins the group above it unless that group’s leader is ahead by more than 1.96 times their combined standard error, so places such as 1, 1, 1, 4 are normal and a shared place is written joint. A model graded on fewer tasks than the rest keeps its position and is never called tied. The rule is on the methodology page.

Yes. The Smart Router is one OpenAI-compatible endpoint with auto-* routing modes: auto-best follows the Combined board, auto-coding, auto-reasoning and auto-tooling follow their boards, the value modes take the most points per measured dollar of one run among models within five points of the best, and auto-fastest-quality the fastest within the same bar. Routing is drift-aware by default, uses your own provider keys and adds zero markup on tokens; Free includes 1,000 routed requests a month. The modes are described in the router documentation.