Architecture cost review
Jev vs GPT-5.4 Nano for Labels: Which Costs Less per Million?
Jev's input rate, $0.042 per million tokens, is below even the batch input rate of every small LLM priced here. The decision comes down to what the LLM gives you for the extra money.
Direct answer
For a bare label, Jev is cheaper at every item size we priced. At 500 tokens it costs $21 per million items, against $125 on GPT-5.4 nano and $60 on Claude Haiku 5.5 or GPT-6 Luna. Even batch Luna costs $30. Pick the LLM when you need a reason or several fields, or when your test shows the accuracy gap one 77-category test found: Jev 79.0% against 83.9% and 86.2% for two OpenAI models.
Decision summary
| Decision area | What matters |
|---|---|
| 300-token item | Jev $12.60 · Haiku 5.5 / Luna $40 (batch $20) · nano $85 (batch $42.50) |
| 500-token item | Jev $21 · Haiku 5.5 / Luna $60 (batch $30) · nano $125 (batch $62.50) |
| 2,000-token item | Jev $84 · Haiku 5.5 / Luna $210 (batch $105) · nano $425 (batch $212.50) |
| Rates (OpenRouter, 3 Nov 2026) | Jev $0.042 in / $0 out · Haiku 5.5 and Luna $0.10 / $0.50 · nano $0.20 / $1.25 · batch halves each |
| Context | Jev 262,144 · nano 400,000 · Haiku 5.5 1,000,000 · Luna 1,050,000 |
| Output | Jev: a choice plus probabilities · LLMs: any text you ask for |
The arithmetic at three item sizes
Each LLM row is input tokens × input rate plus a 20-token label × output rate, per million items. A million 20-token labels is 20 million output tokens. That costs $25 on nano ($1.25 × 20) and $10 on Haiku 5.5 or Luna ($0.50 × 20). Jev's answer costs nothing.
At 300 tokens: Jev 300 × $0.042 = $12.60. Haiku 5.5 or Luna $30 + $10 = $40. Nano $60 + $25 = $85. At 500 tokens: Jev $21, Haiku or Luna $50 + $10 = $60, nano $100 + $25 = $125. At 2,000 tokens: Jev $84, Haiku or Luna $200 + $10 = $210, nano $400 + $25 = $425.
Batch halves both LLM rates. That makes Haiku 5.5 or Luna $20, $30 and $105 at the three sizes, and nano $42.50, $62.50 and $212.50. Jev still comes out lower at every size, because its $0.042 input rate is below the $0.05 batch input of the cheapest LLM.
What OpenTweet's 2x figure means here
OpenTweet's arithmetic page put GPT-5 nano at about twice Jev's cost for a bare label. On 3 November list prices and our 500-token assumption, GPT-5.4 nano comes out near six times Jev ($125 ÷ $21 = 5.95). We did not reconcile their token or price assumptions. The point stands either way: on token price alone, the decision model is cheaper.
Where the small LLM earns its higher bill
A reason the user will read. Jev returns a choice and probabilities, not a sentence. If a support agent needs "Billing, because the customer mentions a double charge", the LLM does it in one call.
Several fields at once. Pulling an order number, a date and a product name out of a message is extraction, not choice. Asking Haiku 5.5 for a 150-token JSON object on a 500-token item costs $50 + $75 = $125 per million items. Jev cannot return free-text fields at any price.
It is already in your stack. If the label rides along on a call you already make, the marginal cost is the 20 extra output tokens: $10 per million items on Luna. A separate Jev call at $21 costs more than that.
Accuracy, and what a wrong label is worth
Kai Waehner cites an independent 77-category intent test with Jev at 79.0% and two unnamed OpenAI models at 83.9% and 86.2%, and calls the gap significant. The article does not give the sample size. A 77-way choice is hard for any small model, so this is the case where an LLM is most likely to win.
If your own test shows a 4.9-point gap (79.0% against 83.9%), that is 49,000 more wrong labels per million. Haiku 5.5 costs $39 more than Jev per million 500-token items ($60 − $21). $39 ÷ 49,000 is about $0.0008. If fixing one wrong label costs you more than that, the dearer model pays for itself. For nano, the figure is $104 ÷ 49,000, about $0.002.
- Wrong label costs a person 30 seconds: use the more accurate model
- Wrong label just sends a ticket to the second-best queue: Jev's price likely wins
- Unsure: route Jev's low-probability answers to the LLM and keep the rest
A split that uses both
The cheapest setup is often a cascade. Jev labels everything, and anything below your probability cutoff goes to Haiku 5.5. If 15% of items fall below the cutoff at a 500-token size, the bill per million items is $21 for Jev plus $9 for the LLM (15% of $60), or $30.
That costs $9 more than Jev alone and $30 less than Haiku 5.5 alone. The LLM sees only the hard cases. Set the cutoff from a labelled sample, not from the first week's gut feel.
A year of labels at 10 million a month
Scale the 500-token figures by ten for 10 million labels a month. Jev costs $210 a month, or $2,520 a year. Claude Haiku 5.5 or GPT-6 Luna cost $600 a month, or $7,200 a year, and $300 a month on batch. GPT-5.4 nano costs $1,250 a month, or $15,000 a year, and $625 a month on batch.
So moving from nano to Jev saves $12,480 a year at that volume, and from Haiku 5.5 $4,680. Set that against the cost of the extra wrong labels from the accuracy section before you move.
What changes in your code
With an LLM, the label set lives in the prompt and a parser reads the reply. With Jev, the label set becomes the list of options in a typed question, and the answer arrives as a choice with a probability per option. The parser and its retry-on-bad-format branch go away.
Jev can only pick from the options you list. If your LLM sometimes answers "none of these", add that as an option, or those cases will be forced into the nearest label. Then run your labelled set again and read the probability on each wrong answer. That tells you where to set the cutoff for sending a case to the LLM.
Key takeaways
- →For a bare label, Jev is cheaper at every item size we priced. At 500 tokens it costs $21 per million items, against $125 on GPT-5.4 nano and $60 on Claude Haiku 5.5 or GPT-6 Luna. Even batch Luna costs $30. Pick the LLM when you need a reason or several fields, or when your test shows the accuracy gap one 77-category test found: Jev 79.0% against 83.9% and 86.2% for two OpenAI models.
- →If you only need one label and a wrong label costs under about a tenth of a cent, move to Jev; otherwise run both on your labelled set and price the extra errors.
- →The 77-category test did not name its OpenAI models or sample size, so treat its gap as a reason to test, not a measured fact about nano.
What users report
Users split on whether a decision model beats a cheap small LLM: Jev wins on latency, but some find small LLMs accurate and cheap enough when speed doesn't matter.
Paraphrased from public posts, with the date and a link to each original. These are other people's accounts, not our tests, and we have not verified each one.
One user reported that Nano was unreliable on open-ended tasks but cheaper than larger models when the task was explicit and output rigidly structured, and usually faster.
Value · GPT-5.4 Nano · Hacker News, Sep 2026
A side-project developer said sorting a list into about 20 categories took 30-60 seconds on DeepSeek V4 Flash, while Jev did it in under a second at the same cost.
Switching · Jev · Hacker News, Sep 2026
A developer replaced a 5.6 Luna ranking sub-agent with Jev to filter retrieval results, reporting 5-10 seconds saved per round, no precision loss, and a slightly lower bill.
Switching · Jev · Hacker News, Oct 2026
One user said Luna still beat Jev on accuracy for their product; latency didn't matter to them and extra Luna tokens cost a negligible amount.
Quality · Jev · Hacker News, Oct 2026
A machine learning engineer said plain LLMs had already been cheap and serviceable zero-shot classifiers, so the excitement around Jev frustrated them.
Value · Decision models · Hacker News, Oct 2026
How this page was prepared
Prices on this page come from OpenRouter's live models API, read on 3 November 2026. Cost per decision is state tokens multiplied by the input rate, because output is free on every decision model listed; small-LLM costs add the stated label length at that model's output rate. OpenRouter's credit fee is shown separately where it matters. We ran no private benchmark for these pages: accuracy and latency figures are credited to the vendor or third party that measured them, and vendor figures are labelled as vendor claims.
Frequently asked questions
Is Jev cheaper than GPT-5 nano?
Yes, for a single label. At a 500-token item with a 20-token label, GPT-5.4 nano costs $125 per million items and Jev $21. Batch nano is $62.50.
Is Claude Haiku 5.5 cheaper than Jev?
No. Haiku 5.5 costs $0.10 per million input tokens and $0.50 output, so $60 per million 500-token labels, or $30 on batch. Jev costs $21.
Is Jev as accurate as a small LLM?
Not always. One independent 77-category test had Jev at 79.0% against 83.9% and 86.2% for two OpenAI models. On fewer, clearer options the gap may vanish. Test on your own cases.
Can Jev give me the reason for its label?
No. It returns choices and probabilities only. Ask an LLM for a reason only when someone will read it.
When does batch pricing change the answer?
It halves the LLM bill but leaves it above Jev at every size we priced, since batch input for Haiku 5.5 and Luna is $0.05 per million against Jev's $0.042. Batch also means waiting for results.
How much would I save in a year by moving labels from nano to Jev?
At 10 million 500-token labels a month, nano costs $15,000 a year and Jev $2,520, a saving of $12,480. On batch nano the saving is $7,500 - $2,520 = $4,980.