TL;DR: Does a model that cannot write change the economics of AI? I think it does. TypeSafe’s Jev charges $0.042 per million input tokens and nothing for output, separating cheap structured decisions from the generated reasoning frontier labs sell at up to $10 in and $50 out.
Update, October 1, 2026: What OpenAI shipped at DevDay. At DevDay on September 29, OpenAI shipped what this essay said the labs would sell: persistence and priority, not tokens. OpenAI launched always-on agents (dots) and an Agents API that keeps long-running sessions, a premium speed tier called Ultrafast, and a $500-a-month plan with 25 times the usage of Plus. They also released GPT-6.1 Sol, which the company claims comes close to Astra at a fifth of Astra’s price, so the cuts again land below the flagship.
In barely two weeks, TypeSafe’s Jev went from launch to viral adoption, six open clones, and investor “fervor” that could value TypeSafe above $10 billion, according to a September 24 report by The Information.
That rapid industry recognition of Jev’s value reinforces my view that it is another tremor in the generative and agentic AI discontinuity: a shift in one layer that amplifies the others. These tremors are building the structural foundation for the Agentic Era in which intelligence is cheap, ubiquitous, and coordinated at machine scale.
My bet is that Jev is a tremor because it points to a broader architectural shift in which intelligence is being decomposed into specialized forms of computation, each invoked only where it is economically appropriate. Jev focuses on decisions: the bounded judgments software makes inside workflows. In doing so, they are becoming function calls priced toward zero. Deliberation remains the premium product for frontier labs.
That distinction changes the architecture of an AI workflow. A cheap decision model can handle routine forks. More capable models are called when the decision is uncertain, or when prose and deeper reasoning are required. The frontier does not disappear. It becomes more selective.
That matters after the backlash against “tokenmaxxing” and the push toward more efficient agentic systems. Jev is a signal that developers increasingly understand the frontier model need not adjudicate every step of a workflow. Where bounded decisions can be resolved reliably, software can buy less frontier inference while keeping control of the workflow, its thresholds, and its record.
If we carry that idea forward, my argument is that intelligence is beginning to separate into two products:
Decisions, the narrow, structured judgments that make up many of the routine judgments software asks models to make, have become a function call priced toward zero.
Deliberation, open-ended reasoning and writing, is what the frontier sells, at ten dollars in and fifty out.
For OpenAI and Anthropic, the competitive threat would not necessarily be a superior model that beats them. It would be software that buys less from them and retains control of the workflow and its record of decisions. Jev makes such a consequential architectural choice easier to implement.
Where those decisions can be resolved reliably, some pricing power moves away from the model layer and toward whoever controls the workflow and the record of what was decided.
TypeSafe itself may be copied away. The architecture it exposed may be harder to reverse.
The unresolved questions are how much work qualifies, whether the economics survive validation and escalation, and who ultimately captures the resulting value.
What is TypeSafe’s Jev decision model?
Jev reads text and answers predefined questions with a choice or score and a probability, rather than generated prose.
On September 15, San Francisco-based TypeSafe emerged from stealth with a $40 million seed round led by DCVC. Co-founder and CEO Diogo Almeida left OpenAI in 2024 and is one of the primary authors of the paper that taught language models to follow instructions, the paper that made ChatGPT possible.
The announcement of Jev caused an immediate shockwave. The launch video was watched about forty million times. The company explains that users give Jev the state of their program and a question with the possible answers declared in advance. It returns one answer and a probability. Input costs $0.042 per million tokens. Output is free, in the company’s words “too cheap to meter”.
TypeSafe calls it a System One model, after Daniel Kahneman’s fast, intuitive mode of thought. Developers submit up to about 32,000 tokens and one or more questions-a choice from a list, a score, a yes or no-and receive answers with probabilities in about a third of a second. Jev cannot choose “none of the above” unless that option is supplied. Almeida says Jev was trained exclusively on synthetic data using “reinforcement learning for calibrated decisions.”
Understanding the method behind Jev matters, because the product's value depends on it.
Where the training behind ChatGPT rewards answers people prefer, Jev’s stated objective is reliable confidence: among answers assigned 90 percent probability, nine in ten should be right. A related approach, described in “Rewarding Doubt”, rewards probability estimates while penalizing over- and underconfidence. That illustrates the idea, not Jev’s undisclosed implementation. TypeSafe has published neither the method nor the model’s weights or parameter count.
Outside analysis of the API suggests a language-model backbone repurposed for decisions, with a probability read directly from the network. Several open clones use that design. What remains is the ability to read text and answer declared questions without task-specific training. What disappears is the generation stage.
That is the economic hinge.
A language model first processes its input, then generates its answer one token at a time. This phase is called decode, where hidden reasoning also accrues. On that outside account, Jev processes the input and reads off a decision without decoding. The saving comes from removing generated text, not merely discounting it.
The trade-off is opacity. Simon Willison notes that a language model’s explanation may be unreliable, but Jev offers none: a move “even further towards black box machine learning”. For the audit-trail problem I described after Astra, that is a further loss of visibility. Almeida frames the change differently: intelligence “will be more like a database than a coworker”. You ask it questions, millions of times, cheaply.
Independent comparisons find Jev faster and cheaper, but the calibration claim is less secure. On a phishing benchmark, one broad question produced 62.6% accuracy, compared with Haiku’s 81.3%. Those results show how much accuracy depends on question structure. They do not, by themselves, establish whether Jev’s confidence scores are reliable.
Splitting the question into five and fitting a regression lifted Jev to 95%. Haiku, given the same decomposition, reached 93%. As one reviewer put it, “the accuracy and the calibration aren’t in the box.”
Should the evaluation measure calibration as well as accuracy: are high-confidence answers reliable enough to accept without escalation? The relevant cost comparison should include the cases that still require another model.
TypeSafe’s headline claims it is 193.6 times faster and 444.6 times cheaper. But its accuracy evaluation raises a separate question: the reference answers come from GPT-6 Astra and Fable 5.1, so agreement with those answers is not independent evidence of correctness.
Jev’s speed and cost are the attraction. Reliable confidence on the customer’s own questions remains the proposition to prove.