Best AI model for 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.

The Tool use board is the 24-hour view: the average of each model’s runs over the last day, ranked with statistical ties, so one odd attempt cannot move a place. The table shows the five best-placed models; the full board, with every model and the 7-day and 30-day views, is on the leaderboard.

The Tool use board: nine tasks, each in a real sandboxed machine — either the job is done at the end or it is not. 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 the board re-ranks after every run.

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.

On the Tool use board in the last 24 hours, 13 models share first place: the measurement cannot separate them, and the one listed first could be listed last after the next run without anything having changed.

The rule, with every weight, is on the methodology page.

Seven measures, by weight: task completion 30, tool selection 20, parameter accuracy 15, efficiency 15, error handling 10, context awareness 5 and safety compliance 5.

The methodology is public — every weight, statistical test and drift constant. The test tasks are private, so no lab can train on them. The full method is on the methodology page.

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.

A function-calling model is judged on the same nine tasks. Tool selection carries 20% of the score — the right tool for each step — parameter accuracy 15% — called with the correct arguments — error handling 10% — recovering when a call fails — and safety compliance 5% — avoiding destructive operations. Task completion, 30%, asks whether the job is done at the end; efficiency, 15%, counts unnecessary calls; context awareness, 5%, whether earlier output is carried forward.

Each model sees its own tool calls and their results in its provider’s native tool format, so the score is the model’s tool calling and not a transcript format. The Smart Router’s auto-tooling mode (auto-agentic is the same mode) sends tool-calling requests to the model placing highest on this board.

Sometimes, and the board is built to show when a change is real. A shared place is the board’s way of saying the measurement cannot separate those models, so the one listed first among them can be listed last after the next run without anything having changed; a model moves to a place of its own only when the gap is beyond the measurement noise. In the last 24 hours, 13 models share first place on the Tool use board.

The boards re-rank after every run — coding every four hours, reasoning and tool use daily, with an hourly canary probe — and the 24-hour view averages the runs of the last day, so one odd attempt cannot move a place; the 7-day and 30-day views on the leaderboard smooth it further.

A model that slips for good is a different question: every suite has its own Page-Hinkley (a CUSUM-family test) change-point detector that compares each model with its own past, and the current answer, model by model, is on Is AI getting worse?. If the question is whether ChatGPT or Claude is getting worse, or a model you use has been nerfed, the family pages answer it with the same data (is ChatGPT getting worse, is Claude getting worse); the method is on AI drift detection.

On the Tool use board in the last 24 hours, the best-placed Claude model is Claude Fable 5.1 (joint 1st of 22, 90 ± 2.5) and the best-placed GPT model is GPT-6 Sol (joint 1st of 22, 89 ± 2.5). They share a place, so neither family leads on agents and tool calling: a statistical tie.

Both on one page: Claude Fable 5.1 vs GPT-6 Sol. Every Claude model is on Anthropic models and every GPT model on OpenAI models; the other four families — Google (Gemini), DeepSeek, Moonshot AI (Kimi) and Z.ai (GLM) — are on the same board above.

The Tool use board ranks ability, not price. What a model costs in use depends on its list price per million tokens and on how many attempts it needs: a model that writes four times as much, or fails and retries, costs more per correct answer than its list price suggests.

Among the models within five points of the best on the Coding board in the last 24 hours, the lowest cost per correct answer over the last 30 days is GPT-6 Luna at about $0.31 per 1,000 correct coding answers, with a pass rate of 92%. The ranking is on best value AI model, and the fastest of the same group, GPT-5.6 Luna at a median of 3.3 s, on fastest AI model.

The head-to-head pages put list price, cost per correct answer and response time beside the scores.

• DeepSeek V4 Pro vs GLM-5.3
• DeepSeek V4 Pro vs GPT-5.6 Luna
• Claude Fable 5.1 vs DeepSeek V4 Pro
• DeepSeek V4 Pro vs Gemini 3.8 Flash
• DeepSeek V4 Pro vs Kimi K3.1 Pro
• Claude Fable 5.1 vs GPT-6 Sol

Set model to auto-tooling on the Smart Router’s OpenAI-compatible endpoint and each request goes to the model placing highest on the Tool use board. auto-agentic picks the same mode under the name most agent frameworks use. The router ranks on the same benchmark results as this page and re-decides as they move; 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.

Every auto-* mode is described in the router documentation.

DeepSeek V4 Pro has a page of its own, with its place on every board and its history. The compare hub lists every head-to-head page, best AI model answers every job on one page, and the methodology has every weight and constant.