Jev vs GPT-6 vs Claude Sonnet 5.5: Do You Need a Model That Returns Scores Instead of Sentences?

TypeSafe's Jev is a model that returns a chosen answer plus probabilities instead of text, priced at $0.042 per million input tokens with free output, according to the company. As of 14 October 2026, that is 2.4 times cheaper on input than the lowest-priced model in OpenAI's GPT-6 family (GPT-6 Luna at $0.10) and about 48 times cheaper than Claude Sonnet 5.5 at $2.

The catch is scope. Jev cannot write a sentence, and its accuracy depends heavily on the task. This comparison is for teams running classification, routing or triage at volume.

What does Jev return that a chat model does not?

Jev returns a pre-defined answer with probabilities. You write a question and the allowed outputs, and the model picks from that fixed set. TypeSafe's docs describe three question types that can be mixed in one call:

• Choice: picks one option from a list and returns choice, probabilities and confidence.
• Score: rates a state on a rubric and returns score, probabilities and confidence.
• Noul: answers whether a statement is true, as a single value between 0 and 1.

The company's launch post says possible outputs are defined in advance and that Jev "gives up string generation." It supports up to 255 options per question, and for larger sets it scores each option and then chooses. TypeSafe says the model never makes type errors, because the output must match the schema.

That is a narrower claim than "no hallucinations". The launch post puts Jev's hallucination rate at 0% because schema matching is guaranteed, and says that number is not empirical. Matching a schema does not mean the answer is right.

How do Jev, GPT-6 and Claude Sonnet 5.5 compare?

GPT-6 is a family, not one model, so the table lists the three models closest to this job. Prices are per million tokens on standard tiers, as of 14 October 2026. Jev's price is the company's.

OpenAI also lists gpt-6-astra at $10 input and $50 output and gpt-6.1-sol at $2 and $10. Sources: TypeSafe, OpenAI, Anthropic. Last checked: 14 October 2026.

What does a classification job cost on each?

The gap is large at volume, even though a single call costs fractions of a cent. Take one million calls, each with 1,000 input tokens and a 10-token answer. This is our arithmetic from the list prices above, not a measured bill.

Batch pricing halves the OpenAI and Anthropic figures, which narrows but does not close the gap with Jev. The comparison also flatters Jev if you need reasoning: a cheap model that needs a second pass costs more than its list price suggests.

TypeSafe's launch post says it cannot prove the price isn't subsidized, and that it expects prices to go down, not up. Treat $0.042 as an early-access rate.

Is Jev as accurate as GPT-6 and Sonnet 5.5?

On some tasks yes, on others no. The strongest public evidence is Vals AI's evaluation, which compared Jev 1.13.0 with eleven other systems including GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna, Claude Sonnet 5.5, Claude Opus 5.5 and Claude Haiku 4.5. The page does not name a funder.

Vals says its LegalBench slice is small, that it is not representative of the full benchmark, and that LegalBench is public and probably in the LLMs' training data. Latency was measured from one machine on a residential connection.

On cost, Vals reports Jev at $0.02 per 1,000 claim-verification cases. Its stated cost ratios against frontier models do not reconcile exactly with its own table, so read them as "orders of magnitude cheaper", not as a precise multiple.

Can you trust Jev's confidence scores?

Only after you check them on your own data. Vals found Jev best calibrated on one task and worst on another. A separate summary in The Daily Brief cites a third-party study of 900 synthetic support tickets that measured a calibration error of 0.107, and Vals also reports that Jev's separate confidence field adds no information beyond its probability distribution.

Calibration matters because routing systems act on thresholds. In Vals's test, a threshold set on 30% of the data and applied to the remaining 265 items let Jev automate 95% of cases, but at a 1.6% error rate against a 1% budget. Six of twelve systems exceeded the budget.

InfoWorld's analysts say the same thing in plain terms. Stephanie Walter of HyperFrame Research said developers must define questions, outputs, thresholds and escalation paths up front, and validate accuracy and calibration on their own data. Paul Chada of Doozer AI noted that a probability does not explain why a decision was made, which may be a problem in regulated industries.

What do TypeSafe's own benchmark claims leave out?

TypeSafe's headline is "193.6x faster, 444.6x cheaper." The launch post says these come from workflow evals run by its own model-capabilities team. The reference answer is the average of GPT-6 Astra and Claude Fable 5.1, which the company says biases results toward those labs. It also says some bias could exist because its team built the workflows, and that speed evals were generally run from laptops on the West Coast.

These are company claims. The independent numbers above are more modest and more mixed.

One third-party phishing test summarised by The Daily Brief shows how task framing changes the result. Asked a single question about 2,000 emails, Jev scored 62.6% against 81.3% for Claude Haiku 4.5. When five narrower questions were combined by a logistic regression fitted on 1,000 labelled emails, Jev reached 95.0% against 93.2% for Haiku, a difference that was not statistically significant, and a two-line regex scored 91.8%. That summary is second-hand, so treat it as a pointer, not a ruling.

The lesson matches TypeSafe's own docs, which advise atomic, gut-check questions and combining answers in code.

Who is behind Jev, and can you buy it today?

TypeSafe's founder is Diogo Almeida, described by InfoWorld as a former OpenAI researcher and RLHF co-inventor. In an announcement on its site, the company says it raised $870 million at a $7.5 billion valuation in a Series A led by Andreessen Horowitz, with Sequoia Capital and DCVC also named. FourWeekMBA reported the round on 9 October 2026.

Availability is limited. As of 14 October 2026, Jev is in early access, and InfoWorld quoted Broadcom site reliability engineer Advait Patel describing it as a hosted service in a single region. He flagged data residency, SLAs and vendor lock-in as early-stage risks.

Which should you use: Jev, GPT-6 or Sonnet 5.5?

Pick by job, not by price alone.

A sensible pattern is a cheap scorer in front and a larger model behind it. Jev or Luna handles the clear cases, and anything below your confidence threshold goes to Sonnet 5.5 or GPT-6 Sol. If you already compare model costs for coding tools, the same budgeting logic applies as in our Cursor vs Windsurf vs Claude Code pricing breakdown.

What's still unclear?

• Whether the price lasts. TypeSafe says it cannot prove the price isn't subsidized.
• Whether results transfer. Jev tied frontier models on one task and finished last on another. No public test covers your data.
• Whether the vendor numbers hold. The 193.6x and 444.6x figures use a reference built from competing models and were run by TypeSafe.
• How independent the independent test is. Vals AI does not disclose who commissioned or funded its evaluation, and its page includes some figures that do not reconcile.
• Weakest point of the case for Jev. A cheaper general model, GPT-6 Luna, already matched it on claim verification, so the real saving may be smaller than the headline ratio once you measure on your own workload.

Last checked: 14 October 2026.