Codesota is the smart router for AI agents and apps. A single unified API across OCR, speech, vision, code, retrieval, and translation — with SOTA, balanced, and cheap tiers picked from the open benchmark registry. Not one vendor's models. Not one modality. Every task, measured, routed.
Three readers. One product. The same unified API serves the agent builder, the app builder, and the research team — because the registry underneath is the same registry for all of them.
An agent like OpenClaw, Hermes, or a Claude-Code-class tool needs TTS, STT, OCR, code, retrieval, and vision all at once. Today that means eight SDKs, eight billing portals, eight rate-limit ceilings, and eight model-deprecation cycles. Codesota collapses all of it into a single bearer token and a single contract.
Picking the right model per task means reading benchmark papers, reconciling leaderboards, renting GPUs, and stitching APIs. The router reads the registry for you and calls the model that wins on your chosen tier — SOTA, balanced, or cheap — so the product team can ship the feature instead of the infrastructure.
When the benchmark registry moves — a new OCR model clears OmniDocBench, a new embedder takes MTEB — the router picks it up on the next call. No drift between the number on the leaderboard and the model in production. The assay is the contract.
One request parameter, three possible answers. The tier grammar is the same for every endpoint — /v1/ocr, /v1/tts, /v1/stt — because the benchmark underneath is the arbiter, not a vendor's marketing page.
{"tier": "sota" | "balanced" | "cheap"}| Tier | What it picks | Criterion | Cost profile | Example (OCR) |
|---|---|---|---|---|
sota | Frontier model for the task | highest score on the canonical benchmark | highest $/call | GPT-5.4 · Gemini 2.5 Pro · Claude Opus 4.6 |
balanced | Best quality-per-dollar | cost-adjusted score across the benchmark | order of magnitude cheaper than SOTA | open model on GPU cloud · hosted by us |
cheap | OSS self-hosted, amortised | floor set by benchmark score, not vibes | lowest $/call, latency varies | PaddleOCR-VL-1.5 · dots.ocr · MonkeyOCR-pro |
Other platforms route — but not across modalities, not across vendors, and not on benchmark-derived tiers. A brief map:
We turned the thesis into a single endpoint. POST /v1/ocr is live today at hardparse.com. On OmniDocBench, the open-source winner is measurably better and two orders of magnitude cheaper than the closed frontier.
| # | Model | Org | Kind | Score | $/1K | Cost basis |
|---|---|---|---|---|---|---|
| 01 | PaddleOCR-VL-1.5 | Baidu | open | 94.50 | $0.09 | self-hosted, amortised |
| 02 | dots.ocr 3B | Rednote | open | 88.41 | $0.04 | self-hosted, amortised |
| 03 | MonkeyOCR-pro | OSS | open | 86.96 | $0.03 | self-hosted, amortised |
| 04 | GPT-5.4 | OpenAI | closed API | 85.80 | $15.00 | vendor API · retail |
| 05 | Gemini 2.5 Pro | closed API | 84.20 | $12.50 | vendor API · retail | |
| 06 | Mistral OCR 3 | Mistral | closed API | 83.40 | $1.00 | vendor API · retail |
PaddleOCR-VL-1.5 self-hosted cost of $0.09/1K pages is the amortised inference cost on a single A100 at a typical utilisation — COGS, not a retail price. GPT-5.4's $15/1K is OpenAI's published list price. So this is a COGS-vs-retail comparison, which flatters the delta. For a straight retail-vs-retail comparison every row shows the actual hosted price you can buy today on /ocr.
tier: "sota" on OCR today.Oil has grades. Electricity has tariffs. Shipping has class codes. Every mature market commoditises by standardising the contract, not the molecule. Intelligence is next — and the contract is the thing worth building.
OpenAI, Anthropic, Google are refineries. They ship something extraordinary, but a refinery's output is only useful once the market around it standardises how you buy, price, and substitute it. On OCR today, we can already quote three interchangeable grades against the same assay. The rest of the tasks follow the same shape.
Codesota is building the assay and the contract. The benchmark registry is the assay report. The task endpoint is the grade. Everything else is implementation detail.
What has shipped, what's in design partnership, what's ranked and waiting its turn. We publish this table as-is so readers can see the gap between the registry and the router — and help close it.
Live means the endpoint accepts production traffic today. In design means we are co-building with one or more partner teams. Ranked means the benchmark exists in the registry but no endpoint is served yet.
| Task | Modality | Registry coverage | Endpoint | Status |
|---|---|---|---|---|
| Document OCR | Vision | OmniDocBench · Mar 2026 | /v1/ocr | Live |
| Text-to-Speech | Speech | ranked | /v1/tts | In design |
| Speech-to-Text (ASR) | Speech | ranked | /v1/stt | In design |
| Visual Question Answering | Multimodal | ranked | /v1/vqa | In design |
| Translation | Text | ranked | /v1/translate | In design |
| Everything else | Mixed | ranked · not served | — | Ranked |
Public products whose shape matches the router. Listed as design targets, not customer logos. If you're building one of these, the rest of this page is for you.
Open-source desktop agent that takes real actions — files, browsing, shell. Replaces eight vendor SDKs with one router contract.
Nous Research's Hermes Agent routes through 20+ models on OpenRouter today. Codesota folds the cross-modality work (OCR, speech, vision) into the same tier contract.
Coding agent class that benefits from cheap-tier retrieval and embedded OCR when the user drops a screenshot or PDF. One bearer token for the non-LLM work.
RPA tools, DevOps agents, QA bots. All of them need the same non-LLM modalities behind a contract they can depend on for two years.
Three doors, one building. Pick the one that matches what you're shipping this quarter.
Codesota is independent, open, and built in Warsaw. The benchmark registry is free to read forever. The router is free to try on OCR. Paid tiers come after design partnership — not before.
Production-ready. Bearer token, one POST, tiered response. Start on hardparse.com.
Open hardparse.com →Propose the benchmark, the dataset, and the endpoint shape. We co-build in design partnership.
Submit a task proposal →One API, every non-LLM modality, three tiers — drop in as a design partner before we price the general tier.
Contact the team →