Agentic model leaderboard

Ranked by Cost-to-Done: the real dollars to finish the job.

preview dataupdated 18h ago

			Standard			Recursiv-only · agentic
#	Model	Score ▾	Quality	Speed	$/1M tok	Cost-to-Done	Completion	Tool acc.	Self-corr.
●1	Claude Sonnet 4.6 Anthropic	85	86	88	$1.80	$0.21	85%	93%	71%
●2	Claude Opus 4.6 Anthropic	80	92	61	$9.00	$0.42	93%	97%	81%
●3	Gemini 3.1 Pro Google	78	89	96	$5.00	$0.39	88%	94%	74%
4	Grok 4.1 Fast xAI	74	82	142	$2.00	$0.29	79%	88%	63%
5	GPT-5.4 OpenAI	73	90	74	$8.00	$0.48	90%	95%	78%
6	Kimi K2.5 Moonshot	62	80	70	$1.20	$0.34	74%	85%	58%
7	Gemini 3 Flash Google	53	73	168	$0.30	$0.33	59%	81%	47%
8	DeepSeek V3.2 DeepSeek	43	78	58	$0.45	$0.51	68%	79%	49%
9	MiniMax M2.5 MiniMax	40	74	81	$0.55	$0.46	62%	76%	44%
10	GPT-4o mini OpenAI	6	64	121	$0.26	$0.72	41%	69%	33%

Benchmarks test models in a vacuum. We run them on real, multi-step work and rank by what it actually costs to finish. The Recursiv-only columns are ones no single-model benchmark can produce.

how we measure →Book a demo

Experiments

Every number above comes from one of these runs.

live·Jun 2, 2026

The real cost of finishing the job

Price-per-token is a lie for agents. We ran ten frontier models on the same real tasks and measured the dollars it actually took to finish. The cheapest models were often the most expensive.

2.7x

cheapest model, real cost-to-done

read →

Run it yourself

Stop guessing which model to ship.

Every number here was produced by running real agentic work on Recursiv. Book a demo and we will show you the platform these experiments run on.

Book a demo How we measure