Cost estimate / fine-tuned open-weight models

What would your AI workload cost on a fine-tuned model?

Describe one workload. We match it to a model we have trained and tested, then compare your current API bill with serving that model on a dedicated GPU. Every figure comes from our published tests, and the assumptions are listed below the results.

An estimate, not a quote. Delivered cost depends on your traffic, availability needs and where the model runs, and is measured in a pilot.

01The workload
02What you use today

Token counts are those measured on our public test set.

03Your requirements

02Models behind the estimate

Four fine-tuned models, tested against frontier APIs.

Each model is a small LoRA adapter on an open-weight Qwen3 base, served with vLLM on one GPU. The three text models are published on Hugging Face with their full results.

Gyde fine-tuned models used by the estimator
ModelTypeMeasured resultServing
complaint-classifier-us-cfpb Qwen3-8B + LoRA adapter. Text to text. Gyde 55%; Sonnet 44.1%; Haiku 37.8%; GPT-5.4 33.9%; GPT-5.4 reasoning 38.7%. Issue accuracy, 82 labels, 1,999 real complaints. About 36,000 complaints per hour on one NVIDIA A30, JarvisLabs India at $0.41/h. US model card, October 2026.
complaint-classifier-in-rbi Qwen3-8B + LoRA adapter. Text to text, English and Hindi. Gyde 96.3%; Sonnet 99.4%; GPT-5.4 99.1%. Category accuracy, 463 synthetic complaints. About 36,000 complaints per hour on one NVIDIA A30, JarvisLabs India at $0.41/h. India preview model card, October 2026.
vehicle-complaint-triage-us-nhtsa Qwen3-8B + LoRA adapter. Text to text. Gyde 87.6%; Sonnet 86.3%; Haiku 85.6%. Component accuracy, 2,000 newest real complaints. About 100,000 complaints per hour on one NVIDIA A30, JarvisLabs India at $0.41/h. Vehicle triage model card, November 2026.
Document extraction (private preview) Qwen3-VL-8B + LoRA adapter. Multimodal: page images to JSON. Digital 86.2% / 85%; scans 47.5% / 71.7%; photos 25% / 45%. Documents fully correct, statements / invoices, synthetic set. Phone-photo statements with row-level fixes: 90.6% fully right, against 77.1% for the largest model alone (synthetic, 24 unseen layouts). About 6,711 pages per hour on one NVIDIA H200 NVL spot, AceCloud India at $2.06/h. Internal test on synthetic documents, October 2026.

03Assumptions

How the estimate is built.

Reviewed 2026-11-05. The figures are list prices and measured test speeds, not invoices.

01

Your current cost

For the published tasks we use the input and output tokens each closed model used on the public test set. Closed models need the full label list in every prompt, so they use far more input tokens than a fine-tuned model. That list is the same on every request, so we price it at the provider's cached-input rate wherever the shared part is long enough for that provider to cache it (Anthropic from 512 tokens on current models, OpenAI from 1,024, Gemini from 4,096). For other models, or if you change the token counts or paste a sample, we use standard list prices from Anthropic, OpenAI and Google, checked 2026-11-05, with half price for batch. Announced price changes apply from their date. If you enter your monthly spend, we use that instead.

02

Gyde's cost

Real-time traffic needs a GPU switched on for the whole window: 730 hours a month for 24/7, or 220 hours for business hours. We add GPUs when peak traffic, taken as 2x the average, exceeds one GPU's measured speed. A nightly batch pays only for processing time plus about 160 seconds of start-up per run. GPU prices are India list prices converted at ₹95 to $1.

03

What is not included

Standby capacity, failed starts, storage, engineering time and data labelling are not in the estimate. A custom task also needs training, at about ₹250-430 of GPU time per run. Results on public or synthetic test sets do not transfer to your data; a pilot measures that.

Gyde Inference pilot

Turn the estimate into a measured result.

A pilot tests the model on your own cases, under your data rules, and records what each accepted result actually costs.

Discuss a workload Do not submit sensitive records through the marketing form.