Architecture cost review
Jev vs Kev 4B vs Solar Decide: Which System One Model to Buy
Decision models went from one product to a category in two weeks. Prices are so close that context window, openness and question fit matter more than the rate card.
Direct answer
Jev is the default: $0.042 per million input tokens, a 32,000-token window and the most documentation and cookbooks. Pick Solar Decide when one question must read a very long state, since it has a 512K window at $0.05 per million. Pick Kev 4B when you want open weights at Jev's price. Pick Laya when you must self-host. Span-01 is a narrower behaviour scorer, not a general router.
Decision summary
| Decision area | What matters |
|---|---|
| Jev 1.13 | $0.042/M input · free output · 32K window · hosted |
| Kev 4B | $0.042/M input · free output · 8,192 window · open weights (LoRA on Qwen3.5-4B-Base) |
| Solar Decide | $0.05/M input · free output · 512K window · hosted (Solar Mini 4) |
| Span-01 | ~$0.02/M input · scores conversations for behaviours you define · free Lite tier |
| Laya | Free Apache-2.0 weights · 512–1,024 default window · self-hosted |
A new category with near-identical prices
TypeSafe's Jev launched the decision-model category on 15 September 2026. By the end of the month, OpenRouter's own briefing listed company for it. Upstage's Solar Decide and Jared Palmer's open-weight Kev 4B answer typed questions over the same System One contract. Respan's Span-01 scores a conversation for behaviours you define. Outside OpenRouter, Convai's Laya is an Apache-2.0 model you run yourself.
The hosted prices sit between about $0.02 and $0.05 per million input tokens, with output free everywhere. At those rates the token bill is rarely the deciding factor. Context window, accuracy on your questions, and whether you need weights you control decide the real cost.
Cost per million decisions
At a 500-token state, Jev and Kev 4B both cost $21 per million decisions in tokens, and Solar Decide $25. Span-01 lists at about $0.02 per million input tokens, so the same volume is about $10, for a narrower scoring job. Add OpenRouter's 5.5% Standard or 8% Business credit fee to each.
Long states change the order. A 20,000-token state fits Jev's window, but at $0.84 per thousand decisions. Kev 4B's 8,192-token window cannot hold it without truncation or summarisation. Solar Decide's 512K window reads it directly at $1 per thousand, and also reads states far beyond Jev's limit.
- Jev: $21 per million 500-token decisions
- Kev 4B: $21 per million 500-token decisions
- Solar Decide: $25 per million 500-token decisions
- Span-01: ~$10 per million 500-token scoring calls
Choosing by state length
Most decision points need a few hundred tokens: a ticket subject, a single message, a pending tool call, a generated draft and the passage it should be checked against. All the hosted models handle that, and so does Laya. Choose on accuracy and tooling.
Questions over a whole contract, a long conversation or a repository summary need context. Jev covers up to 32,000 tokens. Solar Decide's 512K window is the only hosted choice in this group that reads very long states in one call. That can save a summarisation step costing more than the decision itself.
Open weights versus hosted
Kev 4B is published as a LoRA adapter and pointer head on Qwen3.5-4B-Base, so you can run it yourself or call it on OpenRouter at Jev's price. Laya is open weights only, with no hosted rate on OpenRouter. Jev and Solar Decide are hosted.
Open weights matter when you need an exit plan, residency guarantees or very high volume. A hosted model is cheaper to start with, because the provider absorbs serving, scaling and updates. Our Laya pricing page works through the hardware break-even.
Documentation and tooling are part of the price
Jev has the deepest ecosystem today. OpenRouter hosts a documentation hub, a tutorial and a Decisions API reference. Cookbooks cover tool-call gating, a verified cascade, classification at scale and coding-agent permission prompts, and TypeSafe ships official TypeScript and Python SDKs. Every response reports usage.cost.
That reduces integration time, which often costs more than a year of tokens at these rates. For the other models, confirm request format, probability outputs and limits on the model page before committing. Keep your own code behind a small adapter so you can switch providers without rewriting decision logic.
How to choose in an afternoon
Label 300 to 500 real decisions from your system. Run them through Jev and one alternative chosen by your constraint — Solar Decide for long states, Kev 4B or Laya for open weights. Record accuracy, calibration at your threshold, latency and usage.cost.
Pick the model with the lowest cost per correct decision, not the lowest token rate, and keep the runner-up wired as a fallback. Repeat the test monthly. Every one of these products shipped in the last few weeks and will change quickly.
Where Laya and other open options fit
Laya, from Convai Innovations, is the open-weight model in this field without a hosted OpenRouter price. Its 421M English checkpoint answers in about 32.8 milliseconds on a T4 and runs in under 1 GB of memory, but its context is only 512 tokens by default. Its base checkpoints score below a majority-class baseline until fine-tuned. Kev 4B is the other open-weight option: larger, with an 8,192-token window, and available hosted at Jev's price while you decide whether to run it yourself.
A sensible path for a cost-sensitive team is to start hosted on Jev or Kev 4B, log labels, and move only the highest-volume short questions to a self-hosted model once volume justifies a GPU. Our Jev vs Laya comparison and Laya pricing page work through the break-even at different state sizes.
Whichever model you choose, keep the request shape behind your own small interface — state, questions, answer, probability, cost. Swapping providers then becomes a configuration change rather than a rewrite.
Key takeaways
- →Jev is the default: $0.042 per million input tokens, a 32,000-token window and the most documentation and cookbooks. Pick Solar Decide when one question must read a very long state, since it has a 512K window at $0.05 per million. Pick Kev 4B when you want open weights at Jev's price. Pick Laya when you must self-host. Span-01 is a narrower behaviour scorer, not a general router.
- →Standardise your request code on one primitive set, benchmark two models on a few hundred labelled decisions, and keep the second as a fallback.
- →Prices are OpenRouter listings and third-party summaries on 1 October 2026 and can change; accuracy varies by task and should be measured on your own labelled sample.
How this page was prepared
This September 2026 cluster uses OpenRouter's live models API, OpenRouter and TypeSafe documentation, and the Laya model cards, all checked on 1 October 2026. Vendor benchmark claims are attributed, third-party benchmarks are labelled as such, and every cost example states its token assumptions. We did not run a private benchmark for these pages.
Frequently asked questions
What is the cheapest decision model?
Among general hosted options, Jev and Kev 4B both list $0.042 per million input tokens with free output, and Solar Decide $0.05. Span-01 lists lower, around $0.02, for behaviour scoring. Laya has no token price but needs your own hardware.
Which decision model handles the longest input?
Upstage's Solar Decide lists a 512K-token window. Jev reads 32,000 tokens, Kev 4B 8,192, and Laya 512 to 1,024 by default.
Is Kev 4B open source?
Kev 4B is published as open weights — a LoRA adapter and pointer head on Qwen3.5-4B-Base — and is also hosted on OpenRouter.
Are decision models a replacement for chat models?
No. They return typed answers and probabilities, not text. Use them for routing, classification, verification and gating, and call a chat model when you need prose.