Decision models stop charging for output and undercut LLMs on routine agent calls

I.Five decision models in seventeen days

On 15 September, TypeSafe AI, a start-up led by former OpenAI researcher Diogo Almeida, released Jev, which it calls a “System One” model. Jev takes a state and a typed question and returns a choice, a score or a yes/no answer, each with a probability. It writes no text. TypeSafe priced it at $0.042 per million input tokens and billed output at zero, describing it as too cheap to meter.

Others moved fast. Fastino Labs shipped the open-weight GLiNER2.5-Decide on 24 September and the hosted GLiDE on 30 September, according to MarkTechPost. Upstage listed Solar Decide on OpenRouter on 28 September at $0.10 per million input tokens, currently discounted to $0.05, with output free. On 1 October, Cloudflare released Clef and Clef-flash under Apache 2.0, and AWS released the 1.9-billion-parameter Strands Decider 2B as open weights.

One closed product became a product category in under three weeks. What the launches share is a billing model: pay for what the model reads, not for what it writes.

II.The bill moves from what a model writes to what it reads

LLM pricing is built around output. TypeSafe’s launch post notes that existing models charge between $0.20 and $10 per million input tokens, with output typically about five times dearer. Anthropic lists Claude Fable 5.1 at $10 input and $50 output. A decision model emits a label and a number, so the output line on the invoice disappears.

The second shift is the interface. Cloudflare built Clef to be fully compatible with Jev’s API, so customers can swap one for the other. Upstage’s listing says Solar Decide supports the same /v1/systemone schema. Within a fortnight, a start-up’s request format became the shared plug for a category, which is how commodities form.

Demand signals arrived early. Vercel reported on 18 September that nearly 13% of paid AI Gateway teams tried Jev within 24 hours, twice as many as any previous model launch on the gateway. Arize’s Laurie Voss told TechTarget on 22 September that the big labs would follow quickly and that Jev would find it lacks much of a moat. Cloudflare and Amazon proved the point nine days later.

III.The honest technical read: a better-packaged classifier

What changed is narrower than the launch language suggests. Decision models cannot write text, code or summaries, and they need a predefined schema. Jev cannot answer outside its schema, but, as one early review put it, it can still be confidently wrong. The genuine advance is ease of use: zero-shot typed decisions through a hosted API, with calibrated probabilities, and no training run.

That is the skeptic’s strongest point. Glean’s search team tested Jev on four enterprise tasks and concluded that it does not change the fact that classifiers already existed. On query classification Jev agreed with the production baseline 66.8% of the time, against 74.5% for a fine-tuned model. It underperformed production systems on reranking. It won elsewhere: an 8.1-fold speed-up on model routing, and 2.4% disagreement on citation judging against 17.9% for LLM alternatives.

The headline numbers are vendor-graded. TypeSafe’s claim of results up to 193.6 times faster and 444.6 times cheaper compares Jev with frontier models taking 3 to 329 seconds per task, which includes long chains of reasoning a classifier never runs. TypeSafe concedes it cannot prove its price is not subsidised. The Register notes Cloudflare’s scores have not been reproduced on the official Decision Index.

The benchmarks also disagree with each other. Cloudflare’s tables, summarised by Beri.net, show Clef ahead of Jev on intent detection (94.20 against 79.74 on BANKING77) but behind on When2Call, the test of whether an agent should act at all (72.37 against 80.97). Clef-flash drops to 66.77 on out-of-scope detection where Clef scores 97.43. Routing is solved; judgment is not.

IV.The economic channels

Prices. Our estimate, using list prices and a 300-token routing call: Jev costs $12.60 per million decisions, Clef $72, a small LLM about $70 and Claude Fable 5.1 about $3,500. Take an agent platform running one million tasks a month with 20 control calls each. Routing those through a frontier model costs about $70,000 a month; through Jev, about $250. The saving is real for teams that default to their best model for everything. For teams already routing to small models, it is closer to $1,100 a month. Clef, priced at nearly six times Jev, saves almost nothing over a small LLM.

Market structure. Open weights under Apache 2.0 put a ceiling on what any closed decision model can charge, since anyone with the hardware can host Clef or Strands. But hardware is the catch: The Register puts Clef-flash at 41GB-plus of GPU memory and Clef at 85GB-plus. Strands Decider runs on a consumer RTX 3090, but no hosted provider serves it yet. Hosting, not weights, decides the price.

Productivity. Latency compounds inside agent loops. Cloudflare says Clef classified websites in its threat-intelligence workflow in 2.2 seconds against 4.7 for the fastest LLM it tested. A loop with ten gating decisions saves seconds per task, which matters more for interactive agents than the cost line.

Capital and labour. The direct labour effect is small; this is infrastructure, not a worker substitute. The capital effect runs against frontier labs: the cheapest, highest-volume slice of their API traffic is exactly the slice decision models target. Engineering effort shifts to writing schemas and setting confidence thresholds, which is where the probability output earns its keep.

V.Second-order: calibration becomes the product

The first consequence is architectural. A decision model that returns a probability lets software escalate the uncertain cases. Fastino already builds this in, with GLiDE adding extra reasoning on uncertain cases. Reported results show Strands answers given at 0.9 confidence or higher were right about 95% of the time on unseen short tasks. Expensive LLMs, and humans, move to the tail of hard calls. Ned Bellavance told TechTarget that human-in-the-loop workflows will still use an LLM.

The second is distribution versus demand. Elena Brooks argued on 24 September that Vercel’s 13% shows reach through an existing gateway, not paid, repeated use. Jev is still in waitlisted early access. Nobody has published retention.

The third is security. Strands Decider’s bundled server binds to localhost with no authentication, MarkTechPost reports. A model that gates agent actions is a policy layer. Shipped open, it inherits every deployment mistake its users make.

VI.Precedent: the model cascade, rebranded

Routing cheap traffic to a small model and escalating the rest is an old pattern in machine learning, and intent benchmarks such as BANKING77, which Cloudflare cites, predate the LLM boom. What is new is packaging: a hosted, prompt-only API, a shared schema, and a billing model that zeroes output. Glean’s summary fits: Jev makes a good zero-shot classifier much easier to use. Ease of adoption, not a new capability, is what moved 13% of a gateway in a day.

VII.How it plays out: three scenarios

Probabilities are our analytical judgment, not forecasts from a model, and they sum to 100%.