If you need the cheapest LLM API for classification, scoring, routing or grading work inside software, Jev wins, and it is not close: TypeSafe AI lists Jev at $0.042 per million input tokens with output tokens free, against $10 per million input tokens for Claude Fable 5.1, a 238x gap on input (TypeSafe launch post; The Register). If you need something to read, write or reason in prose, Fable 5.1 wins by default, because Jev cannot do it at all. Jev returns typed decisions — choices, scores, probabilities — and never generates text.
That is the whole trade. One model is a component; the other is a writer.
TL;DR
- Price: Jev $0.042 per million input tokens, output free; Claude Fable 5.1 $10 per million input (TypeSafe, The Register).
- Speed: Jev quotes 70ms–500ms end-to-end latency (TypeSafe); the one independent head-to-head measured a 0.35s median per passage for Jev against 8.83s for Fable 5.1 on high effort (Every).
- Accuracy: in that same test Fable caught 7 of 7 planted defects, Jev 6 of 7 (Every).
- Buy Jev for: eval harnesses, content moderation, retrieval reranking, agent step selection, form triage.
- Buy Fable 5.1 for: drafting, summarising, code, multi-step reasoning, anything a human reads directly.
- Caveats: early access is waitlisted, there is no paper, no weights and no parameter count, and the headline multipliers are vendor figures.
- Last verified: 16 September 2026.
What is Jev, and why does it refuse to chat?
Jev is the first model in what TypeSafe AI calls its "System One" family, launched on 15 September 2026 (TypeSafe). Instead of a token stream, the API returns a value that matches a schema you define: a label, a rank, a float, a boolean. There is no prose mode to fall back on.
The refusal is the design. Most production LLM traffic is not conversation, it is a decision dressed as a conversation — you ask a frontier model to pick a category, it writes a paragraph, and you parse the paragraph. Jev deletes the paragraph and bills you only for what you sent. Output tokens are free because there effectively are none to sell; TypeSafe describes them as too cheap to meter.
The naming is a deliberate joke about the Jevons paradox: make a decision cheap enough and applications will make vastly more of them (StartupHub.ai). The company was founded in 2024 in San Francisco, raised roughly $40 million in a DCVC-led seed, and its founder Diogo Almeida co-authored the 2022 InstructGPT paper (arXiv:2203.02155) that underpins instruction-following in ChatGPT. The training method is described as RLCD, Reinforcement Learning for Calibrated Decisions, tuned for calibrated confidence rather than human preference (TypeSafe).
Which is actually cheaper in a real workload?
On list price, Jev's input cost is 238x lower than Fable 5.1's (The Register). The effective gap is wider, because a chat model bills you for the explanation you throw away.
Two published data points give a sense of scale. Every's head of evals ran 777 judgments across 37 documents by 21 questions in under 0.7 seconds, at an estimated quarter of a cent, then 1,709 judgments across 11 experiments for under a cent (Every). Separately, a demo has Jev playing Doom from structured text state at roughly 10 calls per second for about $7 per hour (The Register).
The practical read: workloads you previously sampled — grading 5% of support tickets, spot-checking a fraction of generations — become affordable at 100%. That is a different kind of saving from shaving a bill. If your bottleneck is one long-form draft a day, none of this helps you, and the comparison in our GPT-6 Astra vs Claude Fable 5.1 verdict is the more relevant one.
Where does Fable 5.1 still win?
Accuracy, and generality. In the only independent head-to-head published so far, Every tested both models on four writing checks across 12 synthetic passages, six clean and six carrying planted defects. Jev's median was 0.35s per passage against 8.83s for Fable 5.1 at high effort, roughly 25x faster with an estimated cost around 580x lower — but Fable found all 7 planted defects and Jev found 6 (Every).
TypeSafe's own workflow evals point the same way: Jev scored 67.8% aggregate against 74.1% for its best comparator, winning cost and latency and losing accuracy (TypeSafe; analysis at OrcaRouter). Treat the "zero hallucinations" line carefully: it means the output is guaranteed to match your schema, not that the value inside is right. A confidently wrong enum is still valid JSON.
Speed versus accuracy is a measurable trade rather than a slogan. Across three trials each on an identical seven-constraint article-planning task, Gemini 3.8 Flash (High) and Claude Opus 4.6 (Thinking) both scored 17 of 17 on machine-checked constraint adherence; median wall time 23 seconds for Gemini vs 67 seconds for Opus (n=6, measured 2026-09-16). Sometimes the faster model gives up nothing. With Jev the giving-up is explicit: you lose a defect in seven, and you lose prose entirely.
How does Jev compare with other cheap API options?
Jev is not competing with the discount frontier models, it is competing with the practice of using a frontier model as a classifier. Open-weight and low-cost hosted options still cover far more ground, as we set out in the cheapest AI API and subscription options for 2026 and in the DeepSeek V4 Pro cost-quality analysis. If free tiers matter more than latency, GLM 5.3 Flash against Claude Opus 4.8 is the closer fight, and if raw responsiveness is the goal, our fastest LLM comparison covers the field Jev has now joined from an unusual direction.
What should you do this week?
Pick one decision your product already makes with a chat model, where the answer is a label or a number and no user ever reads the reasoning. Route that single path to Jev, keep Fable 5.1 as the fallback for low-confidence cases, and measure disagreement between the two for a fortnight. That gives you the cost saving with a safety net, and it tells you whether the missing seventh defect matters in your data.
Do not migrate anything user-facing yet. Access is waitlisted, there are no published weights or parameter counts, and TypeSafe concedes its own homepage multipliers sit at the higher end of real-world gains (TypeSafe).
FAQ
Q: Is Jev the cheapest LLM API right now?
A: For decision-shaped calls, yes on list price: $0.042 per million input tokens with free output, against $10 per million input for Claude Fable 5.1 (TypeSafe). For text generation it is not an option at any price, because it does not generate text.
Q: Can Jev replace Claude Fable 5.1 entirely?
A: No. Jev returns typed values only, so drafting, summarising, coding and open-ended reasoning stay with Fable 5.1 or a comparable frontier model.
Q: How much accuracy do you give up?
A: In the only independent head-to-head published so far, Jev caught 6 of 7 planted defects and Fable 5.1 caught 7 of 7 (Every). TypeSafe's own evals show 67.8% aggregate against 74.1% for its best comparator.
Q: Does "zero hallucinations" mean Jev is never wrong?
A: No. It means the response is guaranteed to conform to your schema. The value can still be incorrect.
Q: Can I use Jev today?
A: Early access is waitlisted, and at launch there was no paper, no released weights and no published parameter count.
Q: What is RLCD?
A: Reinforcement Learning for Calibrated Decisions, the training approach TypeSafe describes for Jev, which optimises for calibrated confidence rather than human preference (TypeSafe).
Last verified: 16 September 2026. Pricing and access terms for both models change without notice; check the vendors' own pricing pages before committing.
Corrections log: No corrections yet. Spotted an error? Tell us and we will amend the article and note the change here.
This article was produced with AI assistance and reviewed against the primary sources linked inline.
Top comments (0)