Forty-two dollars per billion input tokens. That's $0.042 per million. GPT-4o mini runs about $0.15. GPT-4o runs $2.50. And Jev's output tokens — for now — are free.
That pricing sheet, published by TypeSafe alongside a $40 million seed round and a product that has been live for days, is the first thing that should stop you. Not the price. The ratio. A system priced at roughly 1/60th of a frontier model is telling you, in its own unit economics, that it is not doing the same job.
The headline claim is louder: Jev does not hallucinate. Zero. Ever.
That line is already moving through enterprise procurement and crypto trading desks. It is also definitional arbitrage, not an alignment breakthrough. The mechanism, per TypeSafe's own documentation, is that Jev does not write sentences. It selects among options, hard-capped at 255. If a system never produces free-form natural language, it cannot satisfy the academic definition of hallucination — which is the generation of fluent, factually wrong content. A system that never speaks cannot lie. That is a reframing, not a cure.
Context first. TypeSafe is the company. Almeida is the founder, and his calling card is authorship on the InstructGPT paper — the work that formalized RLHF and became the substrate for the ChatGPT stack. His pitch opens by conceding the failure modes of that approach: mode dropping, overconfidence, systemic unreliability. The concession is honest, and it's also structurally convenient. Negate the predecessor, then sell the successor.
And here is where the crypto reader needs to pay attention, because we have seen this exact choreography before. A founder dismantles the prior paradigm in public, pivots to a narrower primitive, attaches a superlative claim to it, and raises on the narrative before a single customer has been retained. ERC-20 rush vibes. Proceed with caution.
Now the mechanism. I've audited enough constrained-decoding layers to read the signature on sight: selecting from up to 255 options is classification scoring, a logit-scoring head, or a bounded decoder. That is a different technical species from a generative LLM. The most probable architecture sits a strong language model underneath, doing semantic comprehension, with a constrained output layer on top. If that reading is right, Jev's ceiling is set by the model it wraps, not by anything TypeSafe invented. The company has disclosed neither its base model, its accuracy, nor its calibration. Absent that, "better than competitors" compares an entire stack against an entire stack with the output mouth taped shut.
The mainstream anti-hallucination playbook — retrieval grounding, citation tracing, uncertainty quantification, tool-call verification — is nowhere in the comparison set. That omission is not accidental. Each of those methods keeps the generative surface intact and hardens it; Jev's method removes the surface. When you benchmark against a rival you didn't bring to the fight, you are not benchmarking. You are positioning.
Then there is the question nobody is asking loudly enough: what is Jev's accuracy on the 255-way selection itself? A confidence-calibrated classifier that can decline to answer is useful. A classifier that always answers, never discloses its uncertainty, and is marketed as incapable of error is a liability with a compliance certificate attached. TypeSafe has published none of this — no accuracy figure, no calibration curve, no refusal rate, no confusion matrix. The claim is free. The number would cost them.
The economics track the same story. If Jev's real unit of work is a short option label, $0.042 per million tokens is not aggressive — it's roughly the cost of a classifier call, dressed in generative-modality pricing. Free output is a classic acquisition subsidy, and the only question that matters is what happens when it ends.
Read the subsidy as a countdown, not a gift. If the output is a label, output cost collapses and the subsidy is cosmetic. If the output is LLM-generated explanation text, the subsidy is a burn rate, and 2026 brings either a price inversion or a capability downgrade — and early adopters eat both.
One more number. $40 million at seed. The 2024 median AI seed is a rounding error against that. The outliers — Mistral at roughly $105M, SSI at roughly $1B — carry founder pedigrees that price the team, not the product. TypeSafe's product maturity, days old and access-limited, does not independently support the figure. What supports it is narrative. And the narrative has a soft spot.
The "ChatGPT co-creator" framing does not survive contact with the paper. InstructGPT listed 20 authors; Almeida is fourth, and appears in a footnote among nine "primary contributors" — not as team lead. RLHF as an idea predates that paper by roughly five years, tracing to a different set of researchers. He contributed to a landmark. He did not create ChatGPT. That gap is credit inflation — bounded, not fabricated, but priced into a $40M round as if it were literal.
Then there's the EU AI Act angle. High-risk AI systems carry transparency and accuracy-disclosure obligations, and a capability claim made in a fundraising context sits close to the line on misrepresentation. "Zero hallucination" may also be the most auditable selling point Jev has — but only if a third party verifies it. Self-certified reliability is not reliability. It's a press release.
One disclosure owed to you: the reporting that broke this critique comes from a crypto-native outlet with a documented allergy to Silicon Valley breakthrough narratives. That skepticism is a bias, and I'll flag it — the same article that lands the technical hit also dismisses constrained decision-making too quickly. The claim is inflated, and the underlying approach has value.
Here is the contrarian read, and it cuts against the easy dunk. Everyone is dismissing Jev because the "zero hallucination" claim is a game. Fine. But constrained decision-making is genuinely valuable in high-compliance environments — medical triage, credit risk, content moderation, legal classification — precisely because a bounded output is auditable. "Cannot invent" is a real audit property. The mistake is calling it "cannot err." Choosing the wrong option among 255 is still an error; it's just been renamed into invisibility. And that rename is the actual hazard: a user who trusts "zero hallucination" stops double-checking decisions that were never hallucination-limited in the first place.
Which brings this back to the desks. If you're running an agent that routes capital, classifies on-chain events, or scores counterparty risk, a bounded-output decision layer is the right primitive. What you cannot do is treat "zero hallucination" as "zero error" and shrink your human review. The failure mode has moved from fabrication to misclassification — quieter, harder to catch, and now wearing a badge that says it can't happen.
Gas spike detected. Run — not because the code is broken, but because the marketing has outrun the contract. Uniswap V2 moved the needle on UX, not on truth; this moves the needle on definitions, not on reliability.
Watch three things. The investor list, still undisclosed — it decides whether this is a16z-backed conviction or FOMO capital with a shorter fuse. Any published accuracy or calibration data — one real number ends the debate. And whether "zero hallucination" becomes a category-wide tagline, because once a definitional trick is marketable, every competitor ships the same sentence, and the buyers pay for the confusion.