Economy is now 1 credit per call (≈¥0.17), using GPT-OSS 20B low. Standard remains GPT-OSS 120B medium at 2cr. Economy is an explicit choice for human-reviewed tagging and preprocessing.
We froze 12 new synthetic scenarios and compared five configurations with the same general prompt, rotating serial call order, no retries. These are upstream API times from Fly, separate from the older public-gateway timings below.
12 new inputs
Correct
Median
Estimated cost/call
120B medium
36/36
0.827s
¥0.0395
120B low
36/36
0.535s
¥0.0262
20B low
36/36
0.302s
¥0.0110
Qwen 3.8 none, Strict
35/36
0.233s
¥0.0584
Jev
35/36
0.488s
Not verified here
Qwen and Jev routed a documentation typo containing a quoted refund phrase to billing; the authored label was technical. This is rubric-sensitive, not a universal error or superiority claim.
A separate holdout caught a 20B mistake
On eight additional pre-fixed inputs, Standard scored 24/24 and Economy 23/24. Economy missed “soon” urgency in a message mixing a quotation and a current request. No prompt changes followed the holdout. Medians were 0.874s / 0.295s; estimated provider costs ¥0.0396 / ¥0.0104, respectively. This supports a lower-cost option, not automatic replacement of every Standard request.
Choose by task
Economy, 1cr: human-reviewed tags and draft classification.
Standard, 2cr: negation and exceptions; more correct judgments in this small holdout.
Jev, 2cr: probability-based review design. Thresholds and calibration need evaluation on actual data.
Ordinary code: exact arithmetic and deadlines, with no inference API charge.
Estimated provider costs use usage tokens, published rates and ¥150/USD, excluding authentication, billing, hosting, failures and margin. Retail price is not a technical cost floor. Up to 20 related questions about one state can share a request; this is not an unrestricted batch of 20 separate inquiries.
Correction: Qwen 3.8 also supports Strict JSON, confirmed by all 12 successful requests here. Our earlier JSON-object-mode run does not establish a unique strict-output advantage for 120B.
Small synthetic related scenarios and correlated judgments, not general accuracy, latency guarantees or calibrated confidence. 20B also made errors in earlier prompt configurations. No model substitution. API timing.provider_ms measures the provider segment; Inbox’s header shows browser end-to-end time.
On the same 24 inputs and 72 judgments, both teai Decide and Jev answered every question correctly. Median API response time was 1.79s for Decide and 1.04s for Jev. Jev was faster in this run. Both cost 2 credits per call through teai. Output semantics are another important difference.
01 / A keyword is not a request
“Refund me urgently” belongs with billing. “The error screen shows ‘refund urgently’; fix the display, no rush” belongs with technical support. Similar words, different requests.
Inbox determines the department, urgency and presence of a current refund request in one call. It sorts results by priority, shows a suggested next step and preserves the original API response. Export the queue as JSON for further integration.
No key is needed to view saved measured records. Sign in or enter a personal API key to evaluate your own text with either engine. Entered keys and the queue stay in page memory and disappear when you reload.
02 / Same gateway, same questions
On September 21, 2026, we sent 24 hand-written Japanese/English support messages through https://teai.io/v1/decisions from one Mac. Three questions and expected labels were frozen before the run. Engine order alternated by input, with one sequential call per engine/input: 48 calls, no retries.
Measurement
teai Decide
Jev
Correct judgments
72 / 72
72 / 72
All-correct inputs
24 / 24
24 / 24
HTTP success
24 / 24
24 / 24
Median response time
1.7931s
1.0442s
p95 (nearest-rank)
2.3091s
1.6100s
teai price / call
2cr ≈ ¥0.33
2cr ≈ ¥0.33
Total / 24 calls
48cr
48cr
These are teai retail prices, not upstream costs or direct-provider prices. Yen estimates use ¥1/6 per credit. Timing includes DNS, TLS, network, gateway authentication/billing, upstream processing and JSON decoding on this client. It is not isolated model inference time or browser interaction time.
Decide served openai/gpt-oss-120b via Groq, with medium reasoning, general-v2 and Strict JSON Schema. Jev served typesafe/jev-1.13-20260917 via AI/ML API.
03 / Different outputs, different application logic
Decide returns a category key, an integer level or a boolean such as {"yes":true}. It has no probability or confidence output. Jev also returns probability information; its noul field is numeric.
Before measurement, we fixed the application mapping: Jev noul > 0.5 means yes, and scores round to the nearest integer, ties up. Original values are retained. Jev’s probability output can inform review routing, but we did not evaluate calibration. Decide’s boolean must not be interpreted as 0% or 100% confidence.
04 / Watch the working interface
The captioned demonstration opens saved measurements and sends a real request to both production APIs from the pre-release interface. API waiting time is retained. Individual video timings are separate from the 24-input medians above.
Captioned interface recording and measured comparison. No audio. Pre-release UI, real production API calls.
05 / What this does—and does not—establish
Both engines handled these examples of negation, quoted text, resolved past requests and embedded instructions correctly.
This is a small synthetic set. Twenty of the 24 inputs were previously used in development; the three judgments per input are correlated. It does not establish general 100% accuracy or statistical equivalence.
One environment and one call per input/engine cannot establish performance across regions, traffic levels or real customer workloads.
Detecting a refund request is not deciding refund eligibility. Inbox neither issues payments nor sends replies; a person reviews the next action.
Earlier, we selected GPT-OSS 120B after comparing five Groq models, emphasizing strict JSON output and estimated cost. The roughly sixfold cost advantage over Qwen from that study does not apply to Jev. This new run provides a clear reason to consider Jev’s speed.
Inspect and reproduce
Download all inputs, questions, labels, 48 responses and summaries (JSON). Only synthetic examples are included, with no API keys or account information. Send each cases[].state with the fixed questions object to POST /v1/decisions, using a bearer API key. Set model to teai/decision-120b or typesafe/jev. Each successful call costs 2cr.