Somebody Just Made AI's Dumbest Decisions Cost Nothing

On 15 September, one of the people who built ChatGPT shipped a frontier model that cannot write. Diogo Almeida helped build the chatbot and the RLHF training method behind it, then spent two years in stealth at TypeSafe AI building the inversion of his own invention ¹. His verdict on the industry he helped create is blunt: "We have lightning in a bottle, and yet it is not useful," because models optimize for human language while "computers speak a different language" ¹. Jev, his new model, never speaks at all.

The model answers instead of writing

Jev takes a block of program state and a set of typed questions, and returns answers with calibrated probabilities: a choice among up to 255 options, a score on a scale, or a yes-no probability ². Every question is evaluated in one parallel pass, with no token-by-token decoding, nothing to parse, and no way to invent a fifth tool in a four-tool menu ²³. TypeSafe reports decisions in 70–500 milliseconds, input at $0.042 per million tokens, output free ². It calls the class System One models, after Kahneman's fast thinking, and trains them with reinforcement learning for calibrated decisions, so stated confidence tracks accuracy and code can act above a threshold and escalate below it ².

Agent loops were paying chat prices for traffic-light calls

An agent's life is mostly not reasoning. It is choosing: which tool next, did that click land, is this command safe to run. LangChain states the loop's cost problem plainly: every decision is another model call↗ ⁴. Within two days of launch it shipped the repair as middleware, a router that sends each request to a cheap or a powerful model, and a guardrail that checks every tool call before it executes ⁴. Vercel had already moved the OpenAI Luna classifier that reviews commands for safety onto Jev, and reports answers five to 18 times faster, with greater accuracy ¹. At Browserbase, browser agents now make their move-by-move choices on it for fractions of a cent ⁴.

The bottleneck was verification, not intelligence

The telling number is not a benchmark. A developer pointed Jev at 9,081 product-matching pairs his own engine had parked for human review, and cleared the queue in 13 minutes for 32 cents; he had shelved that stage in June because frontier pricing made it not worth building ⁵.

Automation did not stall because models could not do routine work. It stalled because honestly checking every small step cost more than the step was worth, so only the expensive steps got automated.

TypeSafe's own four-workflow evaluation puts figures on the swap↗: Jev matches GPT-5.6 Terra's accuracy, 67.8 against 67.9 percent agreement with reference answers, at $0.0004 per case ³. The headline multiples, up to 193 times faster and 445 times cheaper ²⁶, are vendor-reported against references averaged from other frontier models, and TypeSafe itself says the peaks sit at the high end of real-world gains ².

UiPath's rules are the exposed asset

UiPath built a $1.94 billion recurring-revenue business on deterministic, rules-based automation, growing 12.5 percent ⁷. Its own operating chief says customers now want "both deterministic and probabilistic AI capabilities," insists the platform stays model-agnostic, and concedes that advancing language models could commoditize simpler automations ⁷. A calibrated decision at fractions of a cent is exactly that commoditization, applied to the decision point inside every workflow UiPath orchestrates. The company can absorb it through the model-agnostic Agent Builder it already sells, or watch its moat become a commodity. Both paths are open. The exposure is not symmetrical.

Platforms that bill by outcome pocket the margin

Salesforce has reported Agentforce recurring revenue above $1.5 billion, up more than 240 percent↗, while billing customers by conversation, resolution and credits, and paying the token bill underneath: more than 19 trillion tokens consumed, and one executive notes a single engineer "can generate like $100,000 a month bill" ⁸. Every routing, guardrail and verification call moved to a near-free layer is margin on a product already sold, and the same arithmetic waits for ServiceNow and Microsoft's Copilot stack, the same question this site asked when Salesforce opened its own agent marketplace to Anthropic: who keeps the software budget once the agent layer underneath it gets cheap. This is why the edge moved in days: Cloudflare lists typesafe/jev in its AI docs, and Vercel put it in its AI Gateway on 16 September ⁵.

Less bandwidth per decision, and more decisions

Be honest about the data-center question. Because every answer arrives in one pass, there is no decoding loop streaming memory per output token, so per decision this looks lighter on memory bandwidth than the chat model it displaces for that task ³. That is the wrong ledger anyway. The point of a cheap layer is to make more agent activity worth running, and every one of those agents still calls a frontier model for its hard fraction. TypeSafe named the model after William Stanley Jevons on purpose: cheaper coal meant more coal burned ²⁵. Total demand on the reasoning tier, and on the memory suppliers behind it, could rise even as unit cost falls. That is the argument; six days of data cannot settle it.

Private, six days old, and the architecture to learn now

What is unsettled: the evals are vendor-run, reference labels are other models' answers rather than ground truth, output is a probability with no rationale, Jev can still misclassify, and TypeSafe admits it cannot prove its pricing is not subsidized ²³⁵⁶. TypeSafe itself is private, seed-funded at $40 million led by DCVC and valued around $200 million on one report↗ ⁹, so there is no ticker that owns this shift and any list claiming one is speculation.

Watch the middle of the stack

The value goes to whoever closes the distance between a capability and someone's workday, and this week that work is closing the price of honest small checks. Watch the middle of the stack: platforms that fold the layer into products they already sell keep the margin, and rules-based vendors that refuse it become the commodity. For anyone building, the cascade, cheap decisions under expensive reasoning, is the architecture to learn now, six-day-old caveats and all. The people who learn it are the ones whose work changes next.