`/api/sota` — the registry, callable.

One GET, one task, one dated pick. Free, CORS-open, no auth, agent-discoverable. Returns the current state-of-the-art per task as JSON, with full provenance — benchmark id, metric, paper link, model card link, and a stable snapshot id you can cache against.

CodeSOTA does not run your inference. We are the assay; you call the model at your own provider. That separation is the product — it’s why the answers are credible.

Try /api/sota →Read the methodology

§ 01 · Quickstart

Three curls.

Index — every task with at least one scored run

curl https://www.codesota.com/api/sota

OCR SOTA pick (short alias)

curl https://www.codesota.com/api/sota/ocr?tier=sota

Same call, full DB id

curl https://www.codesota.com/api/sota/document-ocr?tier=sota

TTS constraint pick

curl "https://www.codesota.com/api/sota/tts?language=pl&deployment=local&commercial_use=true&max_params=500M&metric=hardtext_entity_accuracy"

§ 02 · Response

Example payload.

What the OCR call returns. TTS also supports constraint filters for language, deployment, commercial use, streaming, cloning, latency, cost, params, RAM, sample rate, and code-switching.

GET /api/sota/ocr?tier=sota

{
  "task": "ocr",
  "task_full_id": "document-ocr",
  "task_name": "Document OCR",
  "benchmark": "omnidocbench",
  "benchmark_version": null,
  "tier": "sota",
  "as_of": "2026-04-22T00:00:00.000Z",
  "snapshot_id": "reg-2026-04-22-a1b2c3",
  "pick": {
    "model_id": "paddleocr-vl-1.5",
    "model_name": "PaddleOCR-VL 1.5",
    "model_url": "https://www.codesota.com/model/paddleocr-vl-1.5",
    "vendor": "Baidu",
    "provider_hints": null,
    "score": 94.5,
    "score_metric": "omnidocbench_composite",
    "metric_id": "composite",
    "higher_is_better": true,
    "benchmark": { "id": "omnidocbench", "name": "OmniDocBench" },
    "cost_per_1k_usd": null,
    "cost_basis": null,
    "result_date": "2026-04-22"
  },
  "runners_up": [
    { "model_id": "dots-ocr-3b", "score": 88.41, "...": "..." },
    { "model_id": "monkeyocr-pro", "score": 86.96, "...": "..." }
  ],
  "registry_url": "https://www.codesota.com/ocr",
  "methodology_url": "https://www.codesota.com/methodology",
  "retrieved_at": "2026-04-26T22:00:00.000Z"
}

GET /api/sota/tts?language=pl&deployment=local

{
  "pick": "Kokoro v1.0",
  "reason": "Kokoro v1.0 is the best available hardtext entity accuracy pick among PL language, local deployment, commercial-usable, under 500M params. Evidence tier: codesota measured.",
  "alternatives": ["XTTS v2", "F5-TTS", "Piper"],
  "evidence": {
    "benchmark": "codesota-tts-hardtext-v2",
    "verification_tier": "codesota_measured",
    "run_date": "2026-05-17"
  }
}

§ 03 · Fields

Schema reference.

Field	Type	Meaning
task	string	Short alias if one exists, otherwise the full DB id.
task_full_id	string	Always the full DB id. Stable across alias renames.
task_name	string	Human-readable task name from the registry.
benchmark	string	Dataset id of the winning benchmark for this task.
benchmark_version	string \| null	Reserved for v0.2 quarterly snapshots — currently null.
tier	"sota"	Always "sota" in v0.1. balanced/cheap return 501.
as_of	ISO 8601 string \| null	Date the winning result was scored.
snapshot_id	string	reg-YYYY-MM-DD-xxxxxx. Stable across re-fetches that don’t change the pick.
pick	object	The SOTA model entry for this task — see schema below.
pick.model_id	string	Codesota canonical model id.
pick.model_url	URL	Direct link to the model card on codesota.com.
pick.vendor	string \| null	Vendor or upstream lab if known.
pick.provider_hints	string[] \| null	Reserved — where you can run this model. v0.2.
pick.score	number	Numeric score, in the metric’s native units.
pick.score_metric	string	snake_case <benchmark>_<metric>, e.g. omnidocbench_composite.
pick.metric_id	string	Raw metric id from the registry.
pick.higher_is_better	boolean	Direction. true for accuracy/pass@k; false for CER/WER/FID/loss.
pick.benchmark	object	{ id, name } of the benchmark dataset.
pick.cost_per_1k_usd	number \| null	Reserved — unified vendor pricing. v0.2.
pick.cost_basis	string \| null	Reserved — narrates the cost model used. v0.2.
pick.result_date	YYYY-MM-DD \| null	Date of the winning result.
runners_up	pick[]	Up to 3 entries below the pick, same shape.
registry_url	URL	Curated task hub on codesota.com (e.g. /ocr).
methodology_url	URL	Always /methodology — published, dated, citable.
retrieved_at	ISO 8601	Timestamp the response was generated.

§ 04 · Task aliases

Short names accepted.

The DB carries verbose, hyphenated task ids. The API accepts both forms.

Short	Full id
ocr	document-ocr
code	code-generation
asr	speech-recognition
stt	speech-recognition
tts	text-to-speech
vqa	visual-question-answering
caption	image-captioning
t2i	text-to-image
t2v	text-to-video

§ 05 · Errors

Status codes.

HTTP	When	Body
404	Unknown task id (or alias not registered)	{ error, hint, see }
404	Task is registered but has zero scored runs	{ error, hint, see }
501	Caller asked for tier=balanced or tier=cheap (v0.1 supports only sota)	{ error, hint, see }
503	Registry database unreachable	{ error }
500	Unexpected query failure	{ error, hint }

All error responses include CORS headers, JSON body, and a hint or see field where useful. see typically points back to the index endpoint or the task hub on codesota.com.

§ 06 · Limits

What you can count on.

Caching. Cache-Control public, max-age=300, s-maxage=300. Vercel’s edge serves most hits without touching the database.
CORS. Access-Control-Allow-Origin: *. No credentials, no cookies. Browser code from any origin can call the endpoint.
Rate limit. None today. Reasonable use: a router polling every minute is fine; a crawler hitting every task every second is not. We’ll publish quotas if abuse warrants them.
Versioning. The path is permanent. Reserved fields (provider_hints, cost_per_1k_usd, cost_basis, benchmark_version) publish as null today and populate in v0.2 — additions only, never breaking.
Discoverability. Allowed in robots.txt for AI and search crawlers. Documented in llms.txt for LLM-grounded answers.

§ 07 · Positioning

Why this is not a router.

S&P doesn’t trade. UL doesn’t sell appliances. CodeSOTA doesn’t route inference. The reason every score in the registry is credible is that we have no financial stake in which model wins — adding inference markup would put exactly the wrong incentive in the loop.

/api/sota is the assay. You call the model at your own provider — OpenAI, Anthropic, Replicate, fal, self-hosted, whatever. Your latency, your billing, your circuit breakers. Our SOTA pick stays editorially independent of where you run inference.

If you want a reference router built on this endpoint, see hardparse.com — a separate product that demonstrates how to consume the registry. Note the separation: hardparse takes inference traffic, codesota does not.

/api/sota — the registry, callable.