You're using hardparse in its public beta. The API is free and keyless — no card, no trial timer. When we exit beta, pricing follows CodeSOTA's three-tier framework: you pick the trade-off, we run the inference.
Every request to POST /v1/parse goes through our GPU pipeline on
open-weight VLMs. No signup. No auth. No credit card. Just ship.
Three tiers mapping onto the benchmarks we publish at
codesota.com/ocr.
/v1/parse
routes to tier: "balanced"
unless you opt into something else. Prices are approximate — we'll finalize when the Stripe dial is flipped.
Best open-weight VLM sitting within a few points of the frontier at a tiny fraction of the cost. GLM-OCR / PaddleOCR-VL on our GPUs. This is the tier you want 99% of the time — production workloads, RAG pipelines, document ingestion jobs.
Frontier quality for the last couple of points: compliance runs, eval suites, audit trails. Routes to whichever model is top of the CodeSOTA OCR leaderboard. Most workloads don't need this. When they do, one flag flips.
Smallest model that still clears the quality bar — 3–8B open checkpoints, distilled variants. When you're doing a million calls, money wins. Background OCR, bulk archive work, pre-processing before LLM calls.
Vendor APIs charge 10–150× what open-weight VLMs actually cost to run. We pass the open-model economics through, keep a fair margin, and publish the benchmarks so you can check our work.
| Vendor | Per 1K pages | Quality class | Open weights |
|---|---|---|---|
| Hardparse ● live | $0 today · ~$0.10 after | balanced · 95% of SOTA | yes |
| Google Document AI | $100–$1,500 | Good on printed docs | no |
| Azure AI Document Intelligence | $50–$1,500 | Tiers by feature | no |
| Mistral OCR | $1,000 | Closed, fast | no |
| Tesseract (self-host) | $0 + ops | Dated · loses structure | yes |
When the pipeline is stable enough that we'd be comfortable billing someone for it. No hard date — we'd rather leave it free a month too long than charge for something that's still rough. We'll email everyone who's used the API before flipping the switch, and keep a grandfathered tier for early adopters.
Yes — there will always be an anonymous free tier rate-limited by IP. We expect that limit to drop (from 100/day to maybe 25/day) when paid tiers activate, but the keyless "try it now" workflow is permanent.
Because the cost gap between the best open VLM and the best frontier model is roughly 150×, and forcing every user onto the frontier wastes money 99% of the time. This is the CodeSOTA thesis: three tiers, you pick the trade-off, we run the rest.
No — numbers above are targets, not commitments. We'll finalize when we have enough throughput data from beta to price without guessing. If the final number surprises you, we'll grandfather beta users at the target prices shown here.
Not today. The models are open-weight — GLM-OCR and PaddleOCR-VL-1.5 are on Hugging Face — so nothing stops you from running the pipeline yourself. We're working on a published Docker image for teams that need on-prem for compliance reasons. Email hi@hardparse.com if that's you.
Email hi@hardparse.com with your expected volume and any compliance constraints. Volume discounts and dedicated capacity are available today in beta for teams willing to commit.
We publish the numbers behind each model at codesota.com. If a frontier API beats the open-source tier for your task, we'll say so — and so will the leaderboard.