Which model, on which platform, at what cost.
Submit your real documents. Every feasible model answers every one, an independent judge grades them all, and the ranking is priced at your actual volume. What you get back isn't inference — it's a defensible decision, kept honest over time by drift detection.
No vendor account needed to run one. You sign up for the winner, not the shortlist.
Qwen 2.5 14B for 501 items
$1.42
| Model | Quality | At 501 |
|---|---|---|
| Qwen 2.5 14B | 4.31 ± 0.12 | $1.42 |
| Claude Haiku 4.5 | 4.24 ± 0.19 | $4.67 |
| Llama 3.1 70B | 4.02 ± 0.22 | $1.20 |
| Gemini 3.6 Flash | 3.88 ± 0.31 | $0.59 |
An actual verdict screen. When the top two overlap within their spread, the platform says too close to call instead of naming a winner.
How it works
A measured decision, not a recommendation engine
Four steps, all of them evidence. Nothing here is a projection.
- 01
Bring real items
Paste or upload a few representative documents — up to five. Several rather than one, because a single document cannot tell a model that is reliably good from one that got lucky.
- 02
Every feasible model answers
Our shortlist, or your own candidate list. Models that cannot hold your sample are excluded with the reason, not silently dropped.
- 03
An independent judge grades
Every answer, on the same rubric: are the claims true, is anything missing, does it follow instructions. Graded out-of-band, not by the model that answered.
- 04
Priced at your volume
You state how many items like these you actually run. The ranking is costed at that number, per platform, including the priciest option that fits.
The platform is allowed to decline
Profiles
One object crosses into production
A finished bake-off can become a profile: the decision, made durable. Your application sends a profile id instead of a model id, and the platform routes what was measured.
POST /v1/chat/completions
Authorization: Bearer $NEOSHUNT_KEY
{ "profile": "invoice_extract",
"messages": [{ "role": "user", "content": "…" }] }Same handlers as the UI, bearer-key auth. Routing is part of membership.
Quality drift
The chosen model quietly gets worse on your workload — re-graded against the run that picked it.
Price drift
A vendor changes its list price, or a cheaper platform appears for the same model.
Roster drift
The candidate set changes — a model is retired, or a new one would now win.
An advice layer says whether acting on a drift actually costs money — including the case where it doesn't and you should leave it alone.
Capacity
Four families, ranked side by side
The Proving Ground compares across all of them in one run, so the answer is not confined to whichever vendor you already have an account with.
GPUs we own and power, plus certified partner capacity. Named, single-operator hardware — not a spot market.
API-first vendors: Anthropic, Google, OpenRouter and the rest, at vendor list price.
Rented instances — H100, H200, RTX PRO — provisioned per engagement and billed by the hour.
Azure OpenAI, AWS Bedrock, GCP Vertex. Arrives with bring-your-own-account, so you open only the account the winner needs.
Measured, not estimated
The economics come with receipts
These are measurements taken on this platform's own hardware and rented capacity, with the certification and benchmark files behind them. They are not projections, and they are not a promise about your workload — that is what a bake-off is for.
| What was measured | Result | Conditions |
|---|---|---|
| Owned GPU, all-in | ~$0.067 / GPU-hr | including electricity |
| 70B FP8 on rented H200, batched N=128 | $0.24 / M output | 19,955 tok/s prefill = $0.048/M input |
| Same model class owned, at electricity | ~$0.01 / M output | |
| One QLoRA fine-tune (7B, 2k examples, rented H200) | $0.52 · 4m15s | end to end |
| 70B FP8 on RTX PRO 6000 ($1/hr) | 150 tok/s @ N=8 | certified un-waived · 49,152 ctx honest ceiling |
Your keys, your accounts
Bring your own vendor keys
Members vault their own vendor keys — AES-256-GCM, verified before they are stored — and production traffic routes on their accounts at zero platform charge. You are paying us for the decision, not a margin on tokens.
Vaulted
AES-256-GCM
verified before stored
Platform charge on BYOK traffic
$0.00
you pay the vendor directly
Per-token markup on credits
None
vendor list price
Sample applications
See a profile working before you wire anything up
Chat, image generation, knowledge search, batch and fine-tuning all run against a profile in the portal — production surfaces that exist to demonstrate the decision, not to be the product.
Stop guessing which model to standardise on.
Bring three real items and the volume you run. You'll have a priced, graded ranking — and a profile you can put into production behind one line of JSON.