Jev vs Claude HaikuOne of them can write. One of them is cheap.
Claude Haiku 4.5 is a small fast LLM that can do both halves of the job: make the judgment and write the answer. Jev only makes the judgment. In exchange it returns a probability over the options you declared, which TypeSafe says is trained to be calibrated, though it has published no calibration curves. It costs about $2 to classify 100,000 replies at 500 input tokens each, on the published rate card, instead of hundreds of dollars.
Last updated: September 2026
Jev decides. An LLM writes. OpenTweet publishes.
The short answer
If the call has to end in a sentence, use Haiku. Jev generates no text at all, so there is no version of a write-the-reply task that Jev does badly. It does not do it.
If the call ends in a branch in your code, use Jev. On TypeSafe's own eval it scores 67.8% against Haiku 4.5's 53.6%, costs $0.0004 per case against $0.0195, and returns in 0.4 seconds against 12.5. It also gives you a probability distribution instead of a sentence you have to parse, which is the thing that lets you automate above a threshold and escalate below it.
Both of those are true at once, which is why most real pipelines run both models. Jev reads every item in the firehose. Haiku only ever sees the ones that survived.
Jev vs Claude Haiku 4.5, side by side
| Measure | Jev (jev-1.13.0) | Claude Haiku 4.5 |
|---|---|---|
| Workflow accuracy, 4 workflows | 67.8% | 53.6% |
| Prompt-mode accuracy | No prompt mode exists | 18.1% |
| Cost per workflow case | $0.0004 | $0.0195 |
| Latency per workflow case | 0.4s | 12.5s |
| Input rate card | $0.042 per 1M tokens, output free | Not verified here |
| Classifying 100,000 replies at 500 tokens each | About $2.10 | About $1,950 at the eval board rate per case |
| Generates text | No. None at all | Yes |
| Image input | No. Text only | Yes |
| Returns a probability distribution | Yes, over your declared answer space | No. It can write a number, which is not the same thing |
| Schema conformance | Guaranteed. It cannot emit a value outside the schema | Good with structured output, not guaranteed |
| Questions per request | Many, evaluated in one parallel pass, bounded only by the token budget | Many, but each answer is generated token by token |
| Context budget | 64k per request, 32k for state plus longest question | Far larger |
| Counting and date arithmetic | Documented as unreliable. Do it in code | Not documented as a strength; verify on your own data |
| Conversation and tool use | No | Yes |
| Access | Waitlist, or OpenRouter, Vercel AI Gateway, Cloudflare | Public API |
Accuracy, cost per case and latency from evals.typesafe.ai, which is TypeSafe's own eval, read on 18 September 2026. Jev limits and pricing from docs.typesafe.ai/models. "Not verified here" means we did not have a first party rate card on 18 September 2026. Prompt-mode accuracy is the board's separate prompt-mode column, not the workflow column. Capability rows reflect what each vendor's own structured-output documentation returns, Anthropic's for the Haiku column, not a test we ran.
The cost math on 100,000 social replies
A realistic month for anyone running a brand account: every mention, quote and reply gets read once by a model that decides whether a human should see it.
| Line | Jev | Claude Haiku 4.5 |
|---|---|---|
| Replies to classify per month | 100,000 | 100,000 |
| Input tokens per reply | 500 | 500 |
| Total input tokens | 50,000,000 | 50,000,000 |
| Published input rate | $0.042 per 1M | Rate card not verified here |
| Output charged | None. Output is unmetered | Yes, every generated token |
| Monthly cost | About $2.10 | About $1,950 using $0.0195 per eval case |
| Wall clock at 0.4s vs 12.5s per item, serial | About 11 hours | About 14 days |
Jev figures are arithmetic on the published rate card: 100,000 replies at 500 input tokens is 50M tokens, and 50M at $0.042 per 1M is $2.10. The Haiku column uses the per case cost from TypeSafe's eval board, where a case is a whole workflow rather than a single reply, so it is an upper bound and not a like-for-like price. Serial wall clock is illustrative: both APIs are called concurrently in practice.
The number that actually changes a product decision
At $2 a month you run the classifier on everything, including the large majority of replies nobody would have read. At $1,950 a month you run it on a sample, and a sample is a different product. That is the whole argument for a decision model, and it is worth more than any accuracy delta on the board above.What each one is for
Use Jev when
- Every item in a stream needs the same small judgment: spam, rage bait, worth a reply, which bucket.
- You want to act automatically above a confidence threshold and route the rest to a person.
- The answer space is closed: a Choice of up to 255 options, a Score of 2 to 10 levels, or a yes or no Noul.
- You want twenty questions about one document answered in a single parallel pass.
- Latency sits in a user-facing loop, and 12 seconds is not available to you.
Use Claude Haiku when
- The output is language a person reads: a reply, a summary, a rewrite, an explanation.
- Any input is an image, and Jev cannot see images at all.
- The task needs several steps of reasoning chained, or tool calls, or a conversation.
- The work involves counting, arithmetic or date ordering, all of which TypeSafe documents as Jev weaknesses.
- Volume is low enough that the price difference is noise, which is roughly under 10,000 calls a month.
The one independent test, and what it actually found
Mike Taylor, head of evals at Every, ran Jev against Claude Fable 5.1 on his own writing.
Taylor sent 37 documents with 21 questions each in one request: 777 judgments in under 0.7 seconds for an estimated quarter of a cent. Across 11 experiments he made 1,709 judgments for less than a cent. Median latency per passage was 0.35 seconds against 8.83 for Fable 5.1, and cost was roughly 580x lower.
Then he planted seven defects. Fable 5.1 caught all seven. Jev caught six, missing the same unexplained action defect in all three runs. That is the honest shape of the trade: the cheap model is very nearly as good and occasionally is not. His own framing was that he would want a more thorough accuracy check before production, and that it works well as an early warning system.
Note that Fable 5.1 is a large model, not Haiku, so that test is not a Haiku head-to-head. No independent Jev against Haiku benchmark has been published that we could find. Source: every.to.
What the two of them return
The response shape is the real difference, more than the price.
{
"model": "jev-latest",
"answers": {
"department": {
"type": "choice",
"choice": "billing",
"probabilities": { "billing": 0.84, "technical": 0.159, "sales": 0.001 },
"confidence": 0.596
},
"frustration": {
"type": "score",
"score": 1.035,
"legend": { "0": "Calm", "1": "Frustrated but civil", "2": "Very angry" },
"confidence": 0.842
},
"is_urgent": { "type": "noul", "noul": 0.999 }
},
"usage": { "input_tokens": 312, "output_tokens": 48 }
}Three things to read off that. The Choice came back at 0.84 for billing but only 0.596 confidence, so a conservative pipeline sends it to a human. The Score is a probability weighted mean, which means 1.035 ranks and thresholds but does not measure a magnitude. And the Noul returns noul alone, with no confidence field. That last one is the detail most third-party articles get wrong, and getting it right is how you know a page was written from the response schema rather than the press release.
What a Jev response actually looks like
The free will it go viral scorer asks Jev seven questions about a draft: one Score for reach strength across five rubric levels, three Scores for hook, clarity and specificity, and three Nouls for AI slop, engagement bait and toxicity. TypeSafe publishes a 70ms to 500ms end-to-end range, and the seven questions fit in roughly 430 input tokens, which is about $0.000018 at the published $0.042 per million. No signup, no key.Frequently asked questions
Should I use Jev or Claude Haiku for classification?
Jev, if classification is all you need and the volume is real. On TypeSafe's own four-workflow eval Jev scores 67.8% against Haiku 4.5 at 53.6%, at $0.0004 per case against $0.0195, in 0.4 seconds against 12.5 seconds. Haiku, if the same call also has to produce a sentence, read an image, or do anything Jev structurally cannot. Jev generates no text at all, so a pipeline that ends in written output needs Haiku or another LLM somewhere in it regardless.
How much does it cost to classify 100,000 social replies?
About $2.10 with Jev, at 500 input tokens per reply and $0.042 per 1M input tokens with output unmetered. On TypeSafe's eval board the same 100,000 cases at Haiku 4.5's $0.0195 per case would be about $1,950, though an eval case is a whole workflow rather than one reply, so treat that as the ceiling rather than the like-for-like number. Either way the gap is two to three orders of magnitude, and at that size it decides whether you run the filter on everything or only on a sample.
Is Jev more accurate than Claude Haiku?
On TypeSafe's own board, yes, by a wide margin: 67.8% against 53.6% on workflow accuracy, and 18.1% for Haiku in prompt mode against a Jev score that has no prompt-mode equivalent. TypeSafe built that eval, there is no human ground truth in it, and the reference answer is the average of two other frontier models. Treat it as directional. The one independent test published so far is Mike Taylor's at Every, where Jev caught 6 of 7 planted defects and Claude Fable 5.1 caught 7 of 7. Fable is a much larger model than Haiku, so that result is not a Haiku comparison.
What does Jev return that Haiku does not?
A probability distribution over the answer space you declared. A Choice returns the chosen option, the probability of every option, and a confidence. A Score returns a probability weighted mean plus the rubric legend. A Noul returns a single float from 0 to 1 and nothing else, with no confidence field, which is the detail most third-party write-ups get wrong. An LLM can be asked to state a confidence, but the number it writes is generated text, not a measured distribution.
Can Jev replace Claude Haiku entirely?
No. Haiku writes, summarises, answers, reads images and holds a conversation. Jev does none of that. It also cannot count reliably, cannot order dates, and takes text only. If your workload is a mix of judgment and writing, the cheap answer is not to pick one, it is to let Jev filter the stream and let Haiku write only for the items that survived the filter.
Does Jev have a free tier like some Haiku access routes?
No free tier is documented. Pricing published on docs.typesafe.ai/models is $0.042 per 1M input tokens with output unmetered, and direct access is waitlist gated. You can call Jev today without the waitlist through OpenRouter, the Vercel AI Gateway or Cloudflare, but you still pay per token.
Keep exploring
The rest of the Jev cluster, plus the part of the pipeline that actually sends the post.
What Jev is, in one page
System One, the three primitives, and the limits TypeSafe documents itself.
Jev alternatives
Seven options compared on cost per 1,000 decisions, latency and what each one returns.
Jev vs GPT
The four-workflow benchmark table, and what each model cannot do.
Jev vs Gemini Flash
The multimodal axis. Jev is text only and that rules it out for image-heavy feeds.
Will it go viral, free scorer
A live Jev scorer. Seven questions, no signup.
Reply faster on X
A niche feed of posts worth answering, with AI drafts you approve before they send.
Classify with the cheap model. Write with the good one. Publish with us.
One POST to /api/v1/posts reaches X, Bluesky and LinkedIn, whichever model wrote the text. From $11.99 a month.
- 7-day free trial
- No X developer account needed
- REST API and MCP server on every plan