Build1 publisher3 min readPublished
Octomind wired five agent decision points to a classifier that can only veto
Version 0.54.0 sends skill routing, output scoring and verifier double-checks to Jev, a typed model that returns a calibrated probability and no text, at $0.00001 to $0.0001 an answer on Octomind's cloud.
The Engineer · Build desk
What happened
- Octomind 0.54.0 stopped sending its agent's unseen one-word decisions to a full chat model and started sending them to Jev, a TypeSafe AI model that returns a probability and no text.
- Jev now sits at five call sites: skill routing, scoring long command output chunk by chunk, confirming verifier-flagged gaps, deciding what enters memory, and deciding when a planning step can be skipped.
- Octomind measured each answer from its cloud at 0.5 to 1.3 seconds and $0.00001 to $0.0001, against TypeSafe's published 70 to 500 ms and $0.042 per million input tokens.
- The routing ships on by default on every Octomind plan including Free, with no settings toggle, and the pre-0.54 chat-model path remains the fallback when Jev is slow or unreachable.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- decision Anyone copying the pattern has to sort their call sites into veto-only and approval before wiring in a classifier, because the property that made Octomind comfortable defaulting it on is that every wrong answer costs redone work.
- exposure Free-plan users now send control-plane state, including a diff and test output, to a third-party inference service on every routed decision, and settings offer no way to switch it off.
- cost Savings estimates have to start from your own payload sizes: at the same per-token rate, a 9,500-token call site costs about 40 times a 240-token one.
- contradiction TypeSafe's homepage multipliers sit above what its own blog and Octomind's measured latency support, so the vendor's headline figures cannot carry a procurement case on their own.
At $0.042 per million input tokens, the price of an answer is set by how much state you paste into the question [4]. Octomind's observed band of $0.00001 to $0.0001 per answer works out to roughly 240 input tokens at the bottom and roughly 2,400 at the top [1]. Output is free because a probability is the whole answer [4].
The one cost figure in the post from outside Octomind's own runs lands somewhere else. A developer ran about 5,000 requests for $2 [7], which averages $0.0004 an answer, four times the top of Octomind's range [2]. At the same input price, that is about 9,500 tokens of state per call [3].
TypeSafe quotes 70 to 500 ms per answer [4]. Octomind measured 0.5 to 1.3 seconds from its cloud [5]. Its fastest answers match the vendor's slowest, and its slowest run 2.6 times that ceiling [4]. For the 70 ms end to appear in your agent, the state would have to be small and the network hop short.
The quality evidence is one eval on someone else's workload. Malte Ubl ran Jev against a classifier eval he had previously served with Gemini 2.5 Flash Lite, and it saturated the eval on quality at 6x the speed [6]. An eval a model saturates has stopped discriminating, so that result establishes a floor under Jev's quality on that task, and only a floor. It transfers to your code if the decision there is already a closed-set classification with a fixed answer space [3].
Shipping this on by default is defensible because of the direction of the failures. On the checks Jev stands in for, it can only hold work back; the verifier owns approval [12]. A false positive costs a redone check. When Jev is slow, down or unreachable, the agent runs the pre-0.54 chat-model path [13]. Octomind wrote: "give your cheap, fast model veto power, never approval power." [16]
The "0% tool-call errors" claim is true by construction: inside a closed, typed answer space the model can pick a wrong option, but every option it can pick is well formed [9]. Octomind wrote: "Wrong is recoverable. Malformed is a pager at 3 a.m." [10] That removes the old call site's tail: a few hundred tokens of reasoning and a parse step hoping the model had not wrapped "yes" in a paragraph [19].
The vendor's own headline numbers are weaker than that design argument. TypeSafe's homepage shows 193.6x and 444.6x multipliers, its own blog hedges them as the high end of real-world gains, and Octomind's post tells readers to treat them as marketing [8].
Per-workload, the number you can actually get is local. Typing /info in an Octomind session reports how many chat-model calls Jev replaced in that run [15]; multiply that count by your old per-call chat cost and you have your own saving. The illustrative payload in the post sends a diff, the test output and the flagged gap as state [17], so that is what leaves your process on every one of those calls. The chat model keeps the work that needs generation: writing code, explaining, reasoning through open problems [18].
What to watch
- Whether TypeSafe publishes per-region latency percentiles, or Octomind's 0.5 to 1.3 second band narrows, which would show how much of the 70 ms floor is a lab condition.
- Whether operators report Jev vetoing correct work often enough that the redone checks cost more than the chat-model calls it replaced.
- Whether the default-on routing gains a toggle once customers ask where the diffs and test output in the state payload go.