100.0%
Accuracy on 294 held-out tickets, against 95.9 % for Claude Sonnet and 90.5 % for Claude Haiku
Your model100.0 %
Claude Sonnet95.9 %
Claude Haiku90.5 %
Your own model,honed on your data
Describe it in plain words. HONED writes the data, trains a small open model and tests it against what you use today.
Support intent router
Live, version 3
Compare with
Accuracy
100.0 %
+4.1 points over Claude Sonnet 5.5
4.2× cheaper$0.21 against $0.89 per 1000 calls
1.5× faster0.53 s against 0.79 s per task
| Model | Accuracy | Macro F1 | Time per task | Cost / 1000 calls |
|---|---|---|---|---|
| Your model0.53 s, $0.21 per 1000 | 100.0 % | 100.0 % | 0.53 s | $0.21 |
| Claude Sonnet 5.50.79 s, $0.89 per 1000 | 95.9 % | 95.8 % | 0.79 s | $0.89 |
| Claude Haiku 5.50.58 s, $0.042 per 1000 | 90.5 % | 90.6 % | 0.58 s | $0.042 |
| GLM 5.3 Flash (untrained)0.48 s, $0.22 per 1000 | 88.4 % | 88.1 % | 0.48 s | $0.22 |
Public Bitext support set: clean template text, so 100 % is a ceiling. Each model was tested on the same 294 held-out tickets.
Every step, in the open
Nothing happens behind a spinner. You see the questions, the data, the training and every answer it gives.
The models you pay for now
Pick up to five from the catalogue. Your model is also tested against its untrained base, every time.
From one sentence to a model you can call
Type it, say it, or attach CSV, JSONL or a PDF. Then answer a few questions about how it should behave.
A support model that reads one customer ticket and picks one of our 14 intents: refund, shipping delay, billing error, cancellation and the rest. Unsure means escalate.
tickets-2026.jsonlContinue
Measured on a public task
One task, one eval: support intent routing on the public Bitext dataset, run by the founder on HONED. Its tickets are clean template text, so 100 % is a ceiling. Your task will score differently. That is why your first eval runs on free credits.
4.2×
cheaper per call than Claude Sonnet, at list price
1.5×
faster than Claude Sonnet, at the median response time
One request
Any HTTP client with an API key. Send the input, get the output back. No SDK, no GPU, no infrastructure.
curl …/v1/models/support-router/predict
-d '{"input": "charged twice for 48213"}'
{ "output": "billing_error", "ms": 530 }Free beta, $20 in credits
Train a model, test it on your own task and use it in your app. Every new account gets $20 in free credits for training, evaluation and calls. No credit card or subscription.
$20in free credits
- About 500 training examples, written for your task
- A LoRA trained on the open model that fits your goal
- A test on 100+ held-out examples against up to five models
- Accuracy, macro F1, speed and cost per 1000 calls for each
- Retraining, Try, API access and batch runs
Cost / 1000 calls
Your model$0.21
Claude Sonnet$0.89
Provider list prices. Credits pay these costs with no markup, so $20 covers about one full training run plus API calls.
Questions
What does the free beta include?
You describe the model and answer a few questions. HONED writes about 500 examples for you to review, trains a model and tests it on at least 100 held-out examples against the frontier models you picked. Every new account gets $20 in free credits for training, evaluation, Try, API calls and batch runs.
What happens when my credits run out?
New training, Try, batch runs and API calls pause, and your models and results stay. Invite three friends: when each of them trains a first model, you get $20 more. A training run never starts without enough credits for it.
Do I need my own data?
No. A plain description is enough, and HONED writes the examples. If you have data, attach it as CSV or JSONL (up to 20 MB), or a PDF the model should know (up to 10 MB).
We fine-tune on OpenAI today. Can we move?
Yes. OpenAI takes no new self-serve fine-tuning jobs from January 2027. HONED imports the OpenAI JSONL files you already trained on.
Which models can I compare against?
Up to five from the AI Gateway catalogue, such as Claude, Gemini, GPT and DeepSeek. Your model is always tested against its untrained base too. None of these models writes or grades your data.
Which model does HONED train?
A LoRA on one open-weights model picked for your goal: Inkling Small for most accurate, gpt-oss 120B for balanced and gpt-oss 20B for lowest cost.
How is the result measured?
On held-out examples the model never saw in training. You get accuracy, macro F1 (accuracy per category, averaged), time per task and cost per 1000 calls at list prices. Every answer is shown next to the expected one.
How do I call the model?
With one HTTPS request and an API key from your workspace: send the input as JSON and the output comes back. Each key allows 60 requests a minute.
How long does it take?
With the lowest-cost goal, four measured runs took 13 to 26 minutes from upload to result. Larger models and larger datasets take longer. It runs in the cloud, so you can close the tab, and the page shows its progress live.
Start with
a sentence
Sign up with Google or an email code and get $20 in free credits. Then describe your model.