Code models. All time.

1,057 catalogued models + 82 score-only names
· data as of 2026-09-07

Find any recorded model, lab or family

Benchmark · all 11 available

87 / 1,139 entries scored here · partial coverage, not a global ranking

Highest recorded score

79.2%

Run 2025-12-05 · Sonar Foundation Agent · +0.4 points over Doubao-Seed-Code (78.8%)

Score source

87 scores · 87 with a run date · 0 without

Releases accelerated

13 in 2009 → 49 in 2016 → 119 in 2023. 2023 is the recorded peak; 2026 is partial, 64 to 2026-09-07.

Section 3 · this benchmark

How the recorded frontier moved

SWE-bench Verified — a 500-instance subset of SWE-bench human-confirmed as solvable, OpenAI/Princeton collaboration. Introduced 2024-08-13.

Benchmark source

best recorded submission so far — agent submissions, pooled across scaffolds; not a controlled model-only comparison
a new record
another dated score
RAG baseline (retrieval only, no agent) — a different setup, not on the line
x: leaderboard run date · y: score (%)

Every frontier step · 14

Frontier steps on SWE-bench Verified, oldest first
Run date	Score	Rise	Scaffold	Lab	Source
2024-04-02	22.4%	first	SWE-agent	lab not recorded	Score
2024-09-18	25%	+2.6	Lingma Agent	lab not recorded	Score
2024-10-02	28.8%	+3.8	Lingma Agent	lab not recorded	Score
2024-10-22	40.6%	+11.8	Tools	lab not recorded	Score
2024-10-29	53%	+12.4	OpenHands	lab not recorded	Score
2024-12-21	62.2%	+9.2	CodeStory Midwit Agent	lab not recorded	Score
2025-01-17	64.6%	+2.4	W&B Programmer O1 crosscheck5	OpenAILab source	Score
2025-05-14	66.4%	+1.8	Aime-coder v1	AnthropicLab source	Score
2025-05-19	70.6%	+4.2	TRAE	lab not recorded	Score
2025-05-22	73.2%	+2.6	Tools	lab not recorded	Score
2025-07-31	74.8%	+1.6	Harness AI	AnthropicLab source	Score
2025-08-04	76.8%	+2	EPAM AI/Run Developer Agent	lab not recorded	Score
2025-09-28	78.8%	+2	TRAE	lab not recorded	Score
2025-12-05	79.2%	+0.4	Sonar Foundation Agent	lab not recorded	Score

All scores with a run date · 87

Every SWE-bench Verified score with a run date, oldest first
Run date	Score	Setup	Record	Source
2023-10-10	4.4%	RAG baseline (retrieval only, no agent)	—	Score
2023-10-10	1.4%	RAG baseline (retrieval only, no agent)	—	Score
2023-10-10	1.2%	RAG baseline (retrieval only, no agent)	—	Score
2023-10-10	0.4%	RAG baseline (retrieval only, no agent)	—	Score
2024-04-02	22.4%	agent submissions · SWE-agent	new record	Score
2024-04-02	15.8%	agent submissions · SWE-agent	—	Score
2024-09-18	25%	agent submissions · Lingma Agent	new record	Score
2024-09-18	10.2%	agent submissions · Lingma Agent	—	Score
2024-10-02	28.8%	agent submissions · Lingma Agent	new record	Score
2024-10-02	18.2%	agent submissions · Lingma Agent	—	Score
2024-10-22	40.6%	agent submissions · Tools	new record	Score
2024-10-28	38.8%	agent submissions · Agentless-1.5	—	Score
2024-10-29	53%	agent submissions · OpenHands	new record	Score
2024-12-12	52.2%	agent submissions · Google Jules	—	Score
2024-12-21	62.2%	agent submissions · CodeStory Midwit Agent	new record	Score
2025-01-17	64.6%	agent submissions · W&B Programmer O1 crosscheck5	new record	Score
2025-01-18	44.2%	agent submissions · CodeShellAgent	—	Score
2025-01-22	51.6%	agent submissions · AutoCodeRover-v2.1	—	Score
2025-02-06	63.4%	agent submissions · AgentScope	—	Score
2025-02-14	42.4%	agent submissions · Agentless Lite	—	Score
2025-02-25	62.4%	agent submissions · SWE-agent	—	Score
2025-02-26	41.2%	agent submissions · Agentless Mini	—	Score
2025-03-06	32.8%	agent submissions · SWE-Fixer	—	Score
2025-05-03	64.6%	agent submissions · PatchPilot-v1.1	—	Score
2025-05-11	40.2%	agent submissions · SWE-agent	—	Score
2025-05-14	66.4%	agent submissions · Aime-coder v1	new record	Score
2025-05-19	70.6%	agent submissions · TRAE	new record	Score
2025-05-20	46.8%	agent submissions · OpenHands	—	Score
2025-05-22	73.2%	agent submissions · Tools	new record	Score
2025-05-27	42.4%	agent submissions · Amazon Nova Premier 1.0	—	Score
2025-05-28	46%	agent submissions · PatchPilot	—	Score
2025-06-16	47%	agent submissions · Skywork-SWE-32B	—	Score
2025-06-16	38%	agent submissions · Skywork-SWE-32B	—	Score
2025-06-27	23.2%	agent submissions · MCTS-Refine-7B	—	Score
2025-06-29	58.8%	agent submissions · DeepSWE-Preview	—	Score
2025-06-29	42.2%	agent submissions · R2E-Gym	—	Score
2025-07-20	23.94%	agent submissions · mini-SWE-agent	—	Score
2025-07-20	21.62%	agent submissions · mini-SWE-agent	—	Score
2025-07-20	21.04%	agent submissions · mini-SWE-agent	—	Score
2025-07-20	9.06%	agent submissions · mini-SWE-agent	—	Score
2025-07-25	38%	agent submissions · SWE-agent	—	Score
2025-07-26	58.4%	agent submissions · mini-SWE-agent	—	Score
2025-07-26	53.6%	agent submissions · mini-SWE-agent	—	Score
2025-07-26	39.58%	agent submissions · mini-SWE-agent	—	Score
2025-07-26	28.73%	agent submissions · mini-SWE-agent	—	Score
2025-07-26	13.52%	agent submissions · mini-SWE-agent	—	Score
2025-07-28	64.2%	agent submissions · Undisclosed	—	Score
2025-07-31	74.8%	agent submissions · Harness AI	new record	Score
2025-08-02	55.4%	agent submissions · mini-SWE-agent	—	Score
2025-08-03	9%	agent submissions · mini-SWE-agent	—	Score
2025-08-04	76.8%	agent submissions · EPAM AI/Run Developer Agent	new record	Score
2025-08-04	53.4%	agent submissions · CodeSweep - SWE-agent	—	Score
2025-08-05	69.6%	agent submissions · OpenHands	—	Score
2025-08-06	42%	agent submissions · SWE-Exp	—	Score
2025-08-07	65%	agent submissions · mini-SWE-agent	—	Score
2025-08-07	59.8%	agent submissions · mini-SWE-agent	—	Score
2025-08-07	34.8%	agent submissions · mini-SWE-agent	—	Score
2025-08-07	26%	agent submissions · mini-SWE-agent	—	Score
2025-08-22	54.2%	agent submissions · mini-SWE-agent	—	Score
2025-09-01	60.4%	agent submissions · EntroPO + R2E	—	Score
2025-09-28	78.8%	agent submissions · TRAE	new record	Score
2025-09-30	68.2%	agent submissions · Undisclosed	—	Score
2025-10-14	71.2%	agent submissions · Lingxi v1.5	—	Score
2025-10-15	74.4%	agent submissions · Prometheus-v1.2.1	—	Score
2025-11-03	74.8%	agent submissions · Sonar Foundation Agent	—	Score
2025-11-10	53.6%	agent submissions · FrogBoss-32B-2510	—	Score
2025-11-10	45%	agent submissions · FrogMini-14B-2510	—	Score
2025-11-20	77.4%	agent submissions · live-SWE-agent	—	Score
2025-11-20	66%	agent submissions · mini-SWE-agent	—	Score
2025-11-24	66%	agent submissions · mini-SWE-agent	—	Score
2025-11-24	61%	agent submissions · mini-SWE-agent	—	Score
2025-12-01	60%	agent submissions · mini-SWE-agent	—	Score
2025-12-01	55.4%	agent submissions · mini-SWE-agent	—	Score
2025-12-05	79.2%	agent submissions · Sonar Foundation Agent	new record	Score
2025-12-09	56.4%	agent submissions · mini-SWE-agent	—	Score
2025-12-09	53.8%	agent submissions · mini-SWE-agent	—	Score
2025-12-10	63.4%	agent submissions · mini-SWE-agent	—	Score
2026-02-17	75.8%	agent submissions · mini-SWE-agent	—	Score
2026-02-17	75.8%	agent submissions · mini-SWE-agent	—	Score
2026-02-17	75.6%	agent submissions · mini-SWE-agent	—	Score
2026-02-17	72.8%	agent submissions · mini-SWE-agent	—	Score
2026-02-17	72.8%	agent submissions · mini-SWE-agent	—	Score
2026-02-17	70.8%	agent submissions · mini-SWE-agent	—	Score
2026-02-17	70%	agent submissions · mini-SWE-agent	—	Score
2026-02-17	66.6%	agent submissions · mini-SWE-agent	—	Score
2026-02-19	72.8%	agent submissions · mini-SWE-agent	—	Score
2026-02-26	69.6%	agent submissions · mini-SWE-agent	—	Score

No run date on record · 0none — every score here carries a leaderboard run date; 37 of them have no model date in the source export

Nothing to list: all 87 scores are plotted above.

Section 4 · every model

Explore every year

YearSort rowsLabFamilyWeights

1,057 catalogue models across 77 calendar years, 1950–2026, and 82 score-only names. Scores are SWE-bench Verified.

2026 64 models

2025 106 models

2024 98 models

2023 119 models

2022 91 models

2021 79 models

2020 50 models

2019 68 models

2018 37 models

2017 55 models

2016 49 models

2015 29 models

2014 33 models

2013 19 models

2012 13 models

2011 11 models

2010 10 models

2009 13 models

2008 6 models

2007 6 models

2006 9 models

1950–2005 · 92 models Open early history

82 score-only names · on a leaderboard, not in the Epoch catalogue Open

Section 5 · the whole dataset

What this view covers

1,057 entries from Epoch AI’s notable-models catalogue — a maintained sample, not every model ever trained — and 27 of them have a score. 82 more names exist only as leaderboard scores. 63 of the 214 scores are undated in the source export. Closed-model benchmark results (GPQA, MATH and the rest for GPT, Claude, Gemini), LMArena Elo and ARC-AGI are not covered by this snapshot.

Data as of 2026-09-07 · fetched by HTTP GET only

Read all five source gaps and the method

Source gaps, verbatim

LMArena (Chatbot Arena) Elo — only export found was a pickle file; unpickling an untrusted file can run arbitrary code, so it was not loaded. Left out.
ARC-AGI leaderboard — no downloadable JSON/CSV endpoint found without a browser. Left out.
HumanEval / GPQA / MATH as standalone numbers for closed frontier models (GPT, Claude, Gemini, etc.) — each lives in a separate model card or paper with no bulk feed; not pulled in this pass. GPQA/MATH/MMLU-PRO/BBH/MUSR/IFEval are present here for the 11 open-weight models whose base weights matched Epoch AI's notable-models list exactly, sourced from the Hugging Face Open LLM Leaderboard v2.
SWE-bench Lite and SWE-bench Multilingual 'introduced' dates are the earliest submission date in the published leaderboard data, not a separate primary announcement — documented as a data-floor proxy in raw/SOURCES.md.
1057 models is Epoch AI's notable-models list as of the 2026-09-07 CSV pull, not literally every model ever trained — it is the closest thing to a maintained, sourced, comprehensive spine that exists.

How this page counts

63 of 214 scores carry no date in the source export’s date field, and that field is a model date, not the day a result was recorded. The frontier chart places a score only by its leaderboard run date; 66 scores have none and are listed beside the chart, never plotted.
82 of the 109 scored names are not in the Epoch catalogue: they are searchable score-only entries with no lab, release date or size on record. Only 27 of the 1,057 catalogue models carry any score. A benchmark’s own counts are stated with its result; these are counts for the whole dataset.

The 11 benchmarks

Every benchmark, what it measures and its source
Benchmark	What it measures	Introduced	Scores	Source
SWE-bench	resolving real GitHub issues end-to-end in a live repo, as a coding agent	2023-10-10	13	Source
SWE-bench Lite	a 300-instance curated subset of SWE-bench for cheaper agent evaluation	2023-10-10	25	Source
SWE-bench Verified	a 500-instance subset of SWE-bench human-confirmed as solvable, OpenAI/Princeton collaboration	2024-08-13	87	Source
SWE-bench Multimodal	resolving GitHub issues in visual/frontend software domains (JS/TS UI repos), needs image understanding	2024-10-04	10	Source
SWE-bench Multilingual	resolving GitHub issues across non-Python languages; date is the earliest submission in the published leaderboard data, not a separate announcement post	2026-02-13	13	Source
IFEval	following verifiable natural-language instructions (formatting, length, keywords)	2023-11-14	11	Source
BBH	BIG-Bench Hard: 23 challenging multi-step reasoning tasks	2022-10-17	11	Source
MATH Lvl 5	hardest tier (level 5) of the MATH competition-mathematics benchmark	2021-03-05	11	Source
GPQA	graduate-level Google-proof multiple-choice science questions	2023-11-20	11	Source
MUSR	multistep soft reasoning over long narrative text (murder mysteries, object placement, team allocation)	2023-10-24	11	Source
MMLU-PRO	harder, 10-option-per-question revision of MMLU across 57 subjects	2024-06-03	11	Source

vyke score (proposed)

Unpublished, not a benchmark.

For each (model, benchmark) score, the model’s percentile rank among every model scored on that same benchmark within ±12 months of its own release date. A model’s vyke score is the weighted average of those percentiles, weighted by how many models sit in each cohort. 64 of the 109 scored names receive one (27 of them catalogue models). It is shown only in a model’s profile and takes no part in the highest recorded score or the frontier.