RL environments & SOTA-with-code.
One aggregated, dated, source-tiered registry of the evals, RL environments, models, and papers that move the frontier — the place AI labs check to see which environments actually separate models, and the team that builds the verifiable-reward environments that do.
Three ways in. All backed by the same evidence. #
Whether you are choosing what to cite, deciding what to train on, or tracking the frontier, you land on the same dated, source-tiered registry underneath.
Pick an eval01Browse benchmarks →
Browse benchmarks by what they prove.
Every eval with status, saturation, and lift evidence — so you cite numbers that still separate models, not dead ones.
Improve a capability02Open the recommender →
Get the environments that lift it.
Pick a capability gap; get RL environments and datasets ranked by discriminative power — the ones most likely to move a strong model.
See what’s new03Open the feed →
The frontier feed.
Latest evals, environments, models, and papers — chronological, dated, and linked back to the registry rows they touch.
One representative result per capability area. #
A compact snapshot of the nine capability areas. The homepage only prints a model when the registry row is verified and has an inspectable source URL; otherwise the row stays pending instead of promoting a stale or weak claim.
- Results
- 9,102
- Models tracked
- 163
- Datasets indexed
- 371
- Capability areas
- 9
| Capability | Evidence | Trusted model | Metric | Score | Source | Snapshot |
|---|---|---|---|---|---|---|
| Language & Knowledge | ||||||
OCRBench v2ovis2-5-9bWildASRVQA-v2SWE-bench VerifiedClaude Opus 4.7GAIAMTEBAtari 2600MVTec-ADcurl https://www.codesota.com/api/sota/swe-bench
Cited & referenced byResearchers and analysts
cite the registry. Univ. of Surrey · AAAI 2026Tomasz Tunguz · Theory VenturesUseAIAPIAlternativeToHacker Newsr/MachineLearningSee all citations →
cite the registry.
Results become useful when they answer a question. #
Follow public questions through hypotheses and experiments to dated evidence, durable claims, negative results, and reproduced findings.
QUESTION → HYPOTHESIS → EXPERIMENT → EVIDENCE → CLAIM → FINDING
- Questions
- 7
- Experiments
- 14
- Claims
- 18
- Findings
- 6
How do we minimize language-model BPB under a 16 MB artifact budget?
The pinned leaderboard moves from the 1.2244 baseline to 1.0565 BPB. The selected #1394–#1530 milestones improve from 1.08563 to 1.07336, but they are a frontier progression rather than one literal code lineage.
Which reasoning benchmarks still separate frontier models?
In this matched snapshot ARC-AGI-1 spans 31.4 points, ARC-Challenge spans 0.8, and GSM8K spans 0. Protocol-matched reruns are still required before calling this an intrinsic benchmark property.
When should a benchmark move from leading to primary to control?
Pythia reference results suggest overlapping dynamic ranges, but the promotion thresholds need repeated measurements and confidence intervals.
Practical routes. Benchmarks as evidence. #
The top-level map is a navigation layer, not a perfect ontology. Capabilities, modalities, and vertical domains stay linked through task pages, benchmark sets, datasets, models, papers, and evidence rows.
RouteLanguage & KnowledgeMMLU-Pro · GPQA · MTEB
Reasoning, exams, retrieval, and knowledge-heavy language tasks.
Route63.40Vision & DocumentsCOCO · OCRBench · OmniDocBench
Images, detection, OCR, layout, tables, and document parsing.
RouteAudio & SpeechWildASR · VoiceBench · ESC-50
ASR, audio tagging, voice assistants, speech quality, and TTS.
RouteMultimodal MediaVQA-v2 · TextVQA · MMMU
VQA, charts, video, image-text reasoning, and media understanding.
Route87.6%Code & Software EngineeringHumanEval · LiveCodeBench · SWE-bench
Code generation, repair, repository tasks, and verified software work.
Route87.6%Agents & Tool UseGAIA · WebArena · OSWorld
Long-horizon tool use, browser work, OS tasks, and workflow execution.
RouteStructured Data & ForecastingMTEB · tabular · graph suites
Embeddings, retrieval, reranking, tabular prediction, graphs, and forecasting.
RouteRobotics, Control & RLAtari · Habitat · LIBERO
Simulation, control, games, embodied agents, and manipulation.
RouteScience, Medicine & IndustryCheXpert · MVTec-AD · MedQA
Scientific QA, medical imaging, industrial inspection, and applied AI.
A leaderboard row is not a fact until it can be inspected. #
CodeSOTA is useful only if the evidence is visible. The homepage now surfaces the provenance contract before editorial lineages and release notes.
Dated scores
Rows carry access dates and snapshot context so old frontier claims do not masquerade as current facts.
Metric direction
Every benchmark declares whether higher or lower is better before a winner is selected.
Source tiers
Paper, vendor, reproduced, and registry-maintained rows are labeled separately.
Provenance trail
Benchmark pages connect model, paper, dataset, code, and source URL where available.
The neutral measure for the environments labs train on. #
A market of RL-environment startups is selling to frontier labs — and the labs keep asking the same question: does this environment actually separate models, or is it saturated? CodeSOTA is the independent party that answers it. If you build environments, we certify yours discriminates. If you train models, we tell you which ones are worth the run.
The registry is callable. #
Agents and notebooks should not scrape leaderboards. They should call a stable, source-aware endpoint and cache the snapshot they used.
codesota/sota-api
curl https://www.codesota.com/api/sota/swe-bench
curl https://www.codesota.com/api/sota?area=vision-documents
{
"task": "swe-bench",
"metric": "resolve rate",
"direction": "higher",
"leader": {
"model": "registry top pick",
"score": "dated value",
"source": "paper | vendor | reproduced",
"snapshot_id": "2026-04-27"
}
}
Tell CodeSOTA what to watch. #
Use your words, not our taxonomy.
You do not need to know whether the thing is called SWE-bench Verified, agentic coding, open-weight models, or Polish LLMs. Write the alert you actually want. We map it to benchmark rules, show the interpretation back, and fall back to a weekly digest when the request is too broad.
Broad request detected. Defaulting to a weekly digest.
- Weekly digest of top CodeSOTA changes
Trained something #
that beats the table?
Submit a checkpoint, paper result, or correction with structured benchmark provenance. We validate the score, cross-check the source, and add the row to the registry with its date and evidence trail.