Amazon gives the decider away, and the agent's cheapest calls come off the meter
AWS shipped Strands Decider 2B free this week, the same week OpenAI previewed its own decision model and TypeSafe's raise talks surfaced. The cheapest calls in an agent's loop just got a sticker price of zero.
Vincent Jiang · 3 min read
A homebrew clone became Amazon's free product
Marc Brooker, an AWS distinguished engineer, saw TypeSafe's Jev and built his own take on it 12. The homebrew decider briefly topped the Jevbench ranking for models its size 1, so Amazon cleaned it up and released it on 1 October as Strands Decider 2B: roughly 2 billion parameters on a Qwen3.5-2B base, open under Apache 2.0, with no hosted API and no per-call charge 2. It reads an agent's state and returns a scored choice plus a confidence number 12.
The agent's frequent steps are deflating off the meter
The steps it wants are the frequent ones: route this request, select that tool, review that action 2. AWS measured an earlier checkpoint of the model, version 18, at a median 106 milliseconds per decision on a local RTX 3090; the released build's own latency is unpublished 2.
Jev, launched 15 September, prices input at $0.042 per million tokens with output tokens free, against the $0.20 to $10 input range TypeSafe lists for LLMs 3. Reviewing every action in an agent run cost $2.94 on Jev and $372 on a frontier model in one hackathon test 4. No per-token price list was built for that.
A million input tokens: $10 at the frontier top, 4 cents on Jev, free from AWS
Data
| Value | |
|---|---|
| Frontier LLM input | $10/M |
| Cheapest LLM input | $0.2/M |
| TypeSafe Jev | $0.04/M |
| Strands Decider 2B | free |
Free weights still burn someone's GPU
Amazon's own charts put the model second among public models its size, behind Mapika's decider, and first only among those publishing a complete training recipe; Jev itself is absent, so no head-to-head exists 2. AWS gave no per-request operating estimate, and TypeSafe, which keeps Jev's weights closed, charges $42 for a billion input tokens 23. Free is the sticker, not the cost.
A thousand dollars builds a clone; the board still favors Jev
OpenAI previewed its Decisions API at DevDay on 29 September, with Sam Altman promising extreme speed by narrowing the lab's Luna model to a predefined choice 4. Dozens of clones shipped in the two weeks after Jev 12, and Brooker puts the cost of building one at "hundreds or thousands of dollars" 1. The scoreboard still disagrees: Jev 1.13 has won 21 of 25 StarCraft runs, the open OpenJev stands 0–9, and Jev's cheapest verified win cost $0.016 6.
Fast and cheap is easy; the moat is the data
"If you want it really fast and cheap, use dice," TypeSafe CEO Diogo Almeida says. "Intelligence is the hard part," and his moat is the synthetic data behind the model 4. The current clones "seem more like ML people wanting to implement a cool architecture," he says 1.
The money is still arguing about that moat. A $40 million seed at a $200 million valuation landed on 15 September 5, and the talks now reported by The Information are for $1 billion or more above a $10 billion mark, fifty times the seed valuation if it closes there 7. Even credit is contested: when the raise surfaced, swyx of Latent Space complained that a YouTube video had its logo cropped out with no attribution, adding "pls support smol creators" 89.
Free decisions are demand, not disruption
Jev is named for William Stanley Jevons, on TypeSafe's thesis that "every order of magnitude drop in the cost of intelligence unlocks orders of magnitude more use cases" 3. If that holds, Amazon's free decider sells more compute, not less: more agents making more decisions on hardware someone still rents.
The tell comes on two dates. The first is the price OpenAI publishes for its Decisions API, still in limited preview with no published rate 4. The second is whether TypeSafe's round closes above $10 billion 7. A decision priced like a token keeps the meter alive; a decision priced near zero confirms the bet.
Deepdive
AI-generated from this story and its cited sources. Not investment advice.
