The Decision Model Gold Rush
Short summary
A new category of AI models that pick answers instead of generating them launched in September 2026 and immediately spawned dozens of competitors. The article cuts through the noise by laying out practical criteria for choosing which decision model to use in production rather than tracking which specific model is "best."
- Jev launched September 15 and hit massive adoption (13% of Vercel's paid customers integrated it within a day), but within two weeks Amazon, OpenAI, Cloudflare, and others shipped competing models with different tradeoffs.
- DecideBench, an independent benchmark testing 400 multiple-choice decisions across eight task families, shows Jev at 98% accuracy for $32 per million tasks, but smaller open models like Laya are 35-59% accurate and much faster—accuracy-speed tradeoffs matter depending on your use case.
- The real decision criteria are deployment location (hosted vs. self-hosted), input types (text-only vs. image support), latency requirements, cost at your actual volume, and API compatibility for future switching—not which model has the best benchmark score.
Full summary
A new category of AI models called decision models exploded onto the market in September 2026. These models don't generate text—they pick between options, assign scores, or return yes/no answers with confidence levels. Jev, built by TypeSafe, launched on September 15 and hit a $200 million valuation on a $40 million seed raise. The adoption was absurd: it cleared a 140,000-person waitlist in 36 hours, and 13% of Vercel's paid AI Gateway customers integrated it within a day. By October 1, competitors had already shipped. Amazon released Strands Decider 2B (open-source, built on Qwen3.5-2B), OpenAI announced a Decisions API on GPT-6 Luna with image support, Cloudflare shipped Clef as a drop-in replacement, and smaller players like Fastino added GLiDE and GLiNER2.5-Decide. GitHub sprouted a "jev-alternative" tag with half a dozen clones. Jev got a Wikipedia page before most startups get their first customer.
The real problem with comparing these models is that benchmarks are inherently biased toward whoever runs them. Jev and Fastino both cherry-pick tests that favor their own products. An independent benchmark called DecideBench exists and shows Jev scoring 98% accuracy at $32 per million tasks—the cheapest option above 95% accuracy. Small encoder models like Laya run faster and cheaper but only hit 35-59% accuracy. A surprise winner was Imajev-4B, an open self-hosted model that handles images (which Jev can't) and scored 95% on DecideBench.
Picking the right model depends on six concrete factors: where it runs (hosted vs. self-hosted), what input types it accepts (text, JSON, images), how fast you need results, whether its calibration holds up on your actual data, total cost at your volume, and switching costs later. For low-stakes work like ticket triage—a few thousand decisions daily where mistakes cost a minute of human time—Jev is the obvious choice. For high-frequency, latency-critical work like real-time game loops or trading systems where milliseconds mean money, the encoder models look tempting until you realize 35-59% accuracy is just wrong quickly. The article includes a comparison table but admits it'll be stale in a month.