Best Decision Models in 2026, Ranked by Cost per Million Decisions

Each model's place is set by one number: its OpenRouter input price × 500 tokens, which is the cost of a million decisions at a 500-token state. Ties share a place. Output is free on all of them, so nothing else enters the bill.

We did not test any of these models. Where an independent test exists we name it. Where only the vendor has spoken, we say so. A model ranked third could be wrong twice as often as one ranked eighth, and nothing on the rate card would tell you.

Decider Flash V1.2 is the pick for most Jev users who want to cut the bill. It costs $0.01 per million input tokens and reads 262,144 tokens, the same window as Jev. At 10 million 500-token decisions a month, that is $50 against Jev's $210. Perplexity describes it as a 4B-parameter sibling of Decider V1.2 with the same API contract.

The catch is evidence. We found no independent accuracy test of Decider Flash, and a 4B model is small for questions with many options. Span-02 costs the same, reads text and images, and is described as open-weight. Its window is 131,072 tokens, and again, only Respan has spoken for its accuracy.

Decider V1.2 is Perplexity's 27B checkpoint at $0.02, with the 262,144-token window. If Decider Flash misses on your test set, it is the first model to try next. It costs twice Flash and still half of Jev.

Liquid's d1.1 and Inception's Mercury Decide both cost $0.02 with 65,536-token windows. d1 is still listed at the same price as d1.1, its successor, so pin d1.1 unless you have tuned thresholds on d1. The ~liquid/d1-latest alias will follow new versions on its own. Avoid it in production for that reason.

Clef Flash is $19 per million decisions, 9B, and its weights are Apache 2.0 according to Developers Digest and Classmethod. Its 24,576-token window is the catch. In one independent 117-case test it scored 82.1% against Jev's 88.6% at Jev's thresholds, and 90.6% against 90.9% with its own thresholds.

Jev 1.14 costs $21 and has the most documentation, OpenRouter cookbooks and third-party tests of anything here. It takes text and JSON only. The same 117-case test found its probabilities moved by up to 0.10 between identical calls. d1-ultra and Microsoft-Decision-1 cost the same as Jev with an eighth of the window, 32,768 tokens. d1-ultra adds image input. Drex v1.5 is a 9B model at $20 with 131,072 tokens.

Solar Decide and Solar Decide Flash cost $25 per million decisions and read 524,288 tokens, twice Jev's window. Drex v2.0 (a 35B mixture-of-experts with 3B active, per Nace.AI) and Drex Max cost $30. Drex Max reads up to four images and 262,144 tokens.

GPT-6 Luna Decisions is $50, with a 1,050,000-token window. Buy it for very long states or when your contracts already sit with OpenAI. Clef Omni at $75 and Clef at $96 are the dearest. Clef is worth it only if you need images and want open weights, and it did win the 117-case test at 92.3%. That lead was inside the test's own 95% interval.

Three hosted models also publish weights: Clef and Clef Flash under Apache 2.0, and Span-02, which Respan calls open-weight. Kev 4B, covered on our September comparison, is open-weight too. Open weights let you move to your own GPUs later without rewriting your questions.

Convai's Laya is the only model here that exists only as weights. It has no token price. Our September estimate put one always-on T4 at about $252 a month, which beats Jev only above roughly 11 million 500-token decisions a month. Against Decider Flash at $5 per million decisions, the same T4 needs about 50 million a month to break even ($252 ÷ $5).

Baseten's Joey 1 ($5), Elephant 1 ($17.50) and Mammoth 1 ($50) arrived in late October with 16,384-token windows. They may suit short states, but we found nothing beyond Baseten's own description. Together's Tev1 4B is labelled experimental. ScaleDown's Classify-1 ($20, 1,000,000 tokens) is worth a look for very long documents.

Respan's Span-01 ($10) and the free Span-01 Lite score a conversation for behaviours you define. They are not general routers, so they sit outside the ranking.