Model reports

Quant-Arena

Can an agent develop a stock-scoring strategy on a research period, freeze it, and retain performance on a sealed year? Quant-Arena tests the complete research loop.

Official project Original paper Official project Catalog source

AI Benchmark Radar / Equality Machine · Reviewed 2026-10-27

Paper v1 and the author leaderboard: 76 model/effort configurations plus one proprietary reference. Iterative and one-shot tracks are kept separate.

Research, commit, then assess

A run writes Python features and daily stock scores for a selected universe of 3,042 A-share stocks, using daily and raw Level-2 data. Iterative research permits at most five candidate evaluations on 2022–2023. One program is then frozen and evaluated once on the sealed year 2024. One-shot generation writes one strategy without that research feedback.

Quant-Arena paper v1 · protocol, Table 2, Appendix B and reference strategy

Total score is not return

Total S = 0.3 S_in + 0.7 S_out on a 0–100 scale. The in-sample score combines return-dominated quality with feature diversity; the sealed-period combination rule is fixed per version but unpublished. The cash-account evaluator targets 75 stocks and applies lot constraints, costs and simulated market impact. Its equal-weight market reference is gross of costs.

Quant-Arena paper v1 · protocol, Table 2, Appendix B and reference strategy

Read every effort level

There are 38 iterative and 38 one-shot configurations of nine models. A model’s best total-score configuration need not maximize its annualized return. In Table 2, GPT-6 Sol at medium effort returns 18.4% against the 17.1% gross market, while its highest total score belongs to xhigh. The statement that no model beats the market concerns the selected best-score configurations, not every run.

Quant-Arena paper v1 · protocol, Table 2, Appendix B and reference strategy · Quant-Arena · author leaderboard, all efforts and both settings

Iterative research · complete configurations

All 38 model/effort configurations are retained. Kepler-1 is the publisher’s proprietary reference, displayed separately in the source and not another LLM effort setting.

Iterative research · author leaderboard read 2026-10-27
Model	Effort	Total	S_in	S_out	2024 ARR	Excess (pp)	RankIC	Sharpe	MDD
Kepler-1 Einsia AI	—	77.56	77.22	77.70	94.0%	+76.9	0.085	2.98	−15.9%
GPT-6 Astra	Low	47.22	58.10	42.56	2.3%	−14.8	0.056	0.22	−25.7%
GPT-6 Astra	Medium	47.17	48.09	46.78	9.8%	−7.2	0.044	0.53	−17.9%
GPT-6 Astra	High	43.70	52.34	39.99	−0.9%	−17.9	0.053	0.11	−26.3%
GPT-6 Astra	XHigh	51.29	53.94	50.16	10.7%	−6.3	0.101	0.86	−6.8%
GPT-6 Astra	Max	50.63	52.94	49.64	11.2%	−5.9	0.082	0.81	−7.0%
GPT-6 Sol	Low	38.51	48.85	34.09	1.4%	−15.7	−0.020	0.19	−30.1%
GPT-6 Sol	Medium	49.80	44.14	52.23	18.4%	+1.4	0.018	0.65	−35.9%
GPT-6 Sol	High	42.89	52.16	38.92	−2.5%	−19.5	0.042	0.10	−37.4%
GPT-6 Sol	XHigh	51.14	55.50	49.26	8.6%	−8.4	0.053	1.89	−2.8%
GPT-6 Sol	Max	45.31	46.49	44.81	8.1%	−9.0	0.018	0.41	−30.0%
GPT-5.6 Terra	None	33.09	26.67	35.85	2.4%	−14.7	−0.042	0.22	−24.9%
GPT-5.6 Terra	Low	41.19	50.58	37.17	−4.6%	−21.7	0.022	0.02	−31.4%
GPT-5.6 Terra	Medium	39.50	43.17	37.93	−4.3%	−21.3	0.036	0.00	−29.0%
GPT-5.6 Terra	High	41.54	44.72	40.17	0.0%	−17.0	0.035	0.14	−29.9%
GPT-5.6 Terra	XHigh	44.77	50.60	42.28	9.9%	−7.1	−0.012	0.44	−32.6%
GPT-5.6 Terra	Max	47.58	50.37	46.39	8.4%	−8.7	0.026	0.42	−28.8%
Claude Opus 5.5	Low	30.15	37.76	26.89	−29.4%	−46.4	0.035	−1.21	−44.1%
Claude Opus 5.5	Medium	44.28	56.78	38.92	4.4%	−12.7	−0.002	0.29	−32.7%
Claude Opus 5.5	High	39.89	52.87	34.32	−11.2%	−28.3	0.039	−0.25	−32.7%
Claude Opus 5.5	XHigh	41.92	54.48	36.53	−7.0%	−24.0	0.043	−0.08	−33.2%
Claude Opus 5.5	Max	41.39	50.41	37.52	−4.0%	−21.1	0.030	−0.03	−25.1%
Gemini 3.8 Flash	Low	41.46	53.44	36.32	−6.2%	−23.3	0.037	−0.04	−35.8%
Gemini 3.8 Flash	Medium	43.04	53.58	38.52	−2.9%	−19.9	0.045	0.06	−33.8%
Gemini 3.8 Flash	High	35.39	48.90	29.61	−19.1%	−36.2	0.022	−0.49	−38.5%
Claude Sonnet 5.5	Low	34.97	47.91	29.43	−9.5%	−26.6	−0.009	−0.13	−35.5%
Claude Sonnet 5.5	Medium	39.42	43.39	37.73	−3.5%	−20.6	0.041	−0.02	−24.3%
Claude Sonnet 5.5	High	40.50	54.11	34.67	−9.0%	−26.1	0.044	−0.19	−34.7%
Claude Sonnet 5.5	XHigh	40.10	44.68	38.13	−0.5%	−17.6	0.015	0.15	−35.0%
Claude Sonnet 5.5	Max	40.93	57.33	33.90	−12.2%	−29.2	0.045	−0.23	−36.1%
GLM-5.3	High	38.67	49.31	34.11	−8.6%	−25.7	0.018	−0.14	−31.6%
GLM-5.3	Max	38.91	47.96	35.03	−7.8%	−24.9	0.046	−0.18	−27.8%
Kimi K3	Low	38.31	45.87	35.07	−8.5%	−25.6	0.027	−0.10	−36.4%
Kimi K3	High	31.43	36.85	29.10	−20.0%	−37.0	0.036	−0.84	−36.9%
Kimi K3	Max	31.86	39.87	28.43	−25.6%	−42.7	0.041	−1.14	−39.9%
DeepSeek V4.1 Flash	None	22.00	21.21	22.33	−24.1%	−41.2	−0.022	−1.03	−29.5%
DeepSeek V4.1 Flash	Low	36.45	48.46	31.30	−5.7%	−22.7	−0.009	−0.06	−30.3%
DeepSeek V4.1 Flash	High	35.97	46.02	31.67	−11.3%	−28.4	0.015	−0.25	−34.0%
DeepSeek V4.1 Flash	Max	35.90	42.49	33.08	−11.0%	−28.1	0.037	−0.28	−34.2%

Values are copied as displayed; S is the normalized score, while ARR and MDD are percentages. The two settings are separate protocols. Kepler-1’s construction is undisclosed.

Quant-Arena paper v1 · protocol, Table 2, Appendix B and reference strategy · Quant-Arena · author leaderboard, all efforts and both settings

One-shot generation · complete configurations

All 38 model/effort configurations are retained. Kepler-1 is the publisher’s proprietary reference, displayed separately in the source and not another LLM effort setting.

One-shot generation · author leaderboard read 2026-10-27
Model	Effort	Total	S_in	S_out	2024 ARR	Excess (pp)	RankIC	Sharpe	MDD
Kepler-1 Einsia AI	—	77.56	77.22	77.70	94.0%	+76.9	0.085	2.98	−15.9%
Claude Opus 5.5	Low	35.04	43.21	31.53	−4.5%	−21.6	−0.011	0.00	−33.0%
Claude Opus 5.5	Medium	42.61	47.19	40.65	5.6%	−11.5	−0.001	0.34	−21.0%
Claude Opus 5.5	High	36.07	39.68	34.52	−7.5%	−24.5	0.020	−0.14	−29.7%
Claude Opus 5.5	XHigh	39.13	49.05	34.88	−9.9%	−26.9	0.040	−0.20	−35.0%
Claude Opus 5.5	Max	39.70	42.07	38.69	−3.7%	−20.8	0.046	0.03	−35.9%
GPT-5.6 Terra	None	42.14	46.56	40.25	9.5%	−7.6	−0.010	0.44	−27.8%
GPT-5.6 Terra	Low	27.82	26.59	28.35	−10.3%	−27.4	−0.007	−0.24	−25.7%
GPT-5.6 Terra	Medium	21.01	20.44	21.25	−32.4%	−49.4	−0.010	−1.51	−34.1%
GPT-5.6 Terra	High	40.61	42.88	39.64	6.2%	−10.9	−0.001	0.35	−27.5%
GPT-5.6 Terra	XHigh	34.21	42.94	30.46	−9.8%	−26.8	0.000	−0.17	−33.6%
GPT-5.6 Terra	Max	20.98	20.98	20.97	−32.3%	−49.3	−0.009	−1.60	−32.6%
Claude Sonnet 5.5	Low	37.75	42.41	35.76	1.1%	−16.0	−0.007	0.19	−30.3%
Claude Sonnet 5.5	Medium	33.82	41.81	30.39	−11.5%	−28.6	0.002	−0.21	−35.5%
Claude Sonnet 5.5	High	40.24	44.28	38.50	1.4%	−15.6	0.010	0.20	−32.1%
Claude Sonnet 5.5	XHigh	37.05	47.90	32.40	−14.3%	−31.3	0.029	−0.33	−36.0%
Claude Sonnet 5.5	Max	33.89	41.19	30.77	−14.8%	−31.9	0.018	−0.33	−36.9%
GLM-5.3	High	38.02	52.54	31.79	−4.2%	−21.3	−0.023	0.05	−37.3%
GLM-5.3	Max	31.46	34.70	30.07	−15.4%	−32.4	0.022	−0.51	−32.1%
Gemini 3.8 Flash	Low	27.75	21.26	30.53	−12.1%	−29.1	0.005	−0.41	−23.5%
Gemini 3.8 Flash	Medium	35.91	35.62	36.04	1.6%	−15.4	−0.006	0.19	−23.8%
Gemini 3.8 Flash	High	30.26	25.93	32.11	−5.4%	−22.5	−0.003	−0.06	−24.6%
GPT-6 Astra	Low	30.08	35.05	27.95	−11.2%	−28.3	−0.010	−0.28	−30.3%
GPT-6 Astra	Medium	31.02	36.73	28.57	−9.7%	−26.7	−0.012	−0.20	−30.3%
GPT-6 Astra	High	35.22	38.27	33.92	−4.8%	−21.9	0.007	−0.03	−30.9%
GPT-6 Astra	XHigh	30.30	38.46	26.80	−16.7%	−33.8	−0.002	−0.48	−34.7%
GPT-6 Astra	Max	31.47	40.13	27.75	−18.3%	−35.3	0.007	−0.54	−34.8%
GPT-6 Sol	Low	32.15	38.29	29.52	−7.6%	−24.7	−0.014	−0.08	−32.2%
GPT-6 Sol	Medium	22.60	22.62	22.59	−31.4%	−48.4	−0.006	−1.31	−34.1%
GPT-6 Sol	High	23.15	21.07	24.04	−17.4%	−34.4	−0.020	−0.51	−29.6%
GPT-6 Sol	XHigh	23.29	23.84	23.05	−19.0%	−36.1	−0.022	−0.60	−29.6%
GPT-6 Sol	Max	22.31	24.08	21.55	−23.7%	−40.7	−0.018	−0.98	−32.5%
Kimi K3	Low	31.58	38.58	28.58	−13.4%	−30.5	0.000	−0.34	−33.8%
Kimi K3	High	32.01	37.89	29.48	−20.5%	−37.6	0.038	−0.84	−38.6%
Kimi K3	Max	31.65	40.87	27.70	−20.3%	−37.3	0.016	−0.86	−35.1%
DeepSeek V4.1 Flash	None	20.48	22.01	19.82	−34.1%	−51.1	−0.015	−1.79	−37.5%
DeepSeek V4.1 Flash	Low	29.36	34.21	27.28	−17.4%	−34.4	0.001	−0.49	−33.3%
DeepSeek V4.1 Flash	High	29.63	37.51	26.25	−25.4%	−42.4	0.010	−0.92	−39.5%
DeepSeek V4.1 Flash	Max	29.92	35.95	27.34	−11.9%	−29.0	−0.008	−0.42	−26.7%

Values are copied as displayed; S is the normalized score, while ARR and MDD are percentages. The two settings are separate protocols. Kepler-1’s construction is undisclosed.

Quant-Arena paper v1 · protocol, Table 2, Appendix B and reference strategy · Quant-Arena · author leaderboard, all efforts and both settings

What this evidence establishes

This is an author-run historical assessment on one stock universe and one sealed year, with a controlled data interface and benchmark-specific trading assumptions. It measures research and strategy construction under that setup. It does not establish forward investment performance, an independently reproduced result, or a general finance knowledge ranking.

Quant-Arena paper v1 · protocol, Table 2, Appendix B and reference strategy

Original publication

Paper v1 and the author leaderboard: 76 model/effort configurations plus one proprietary reference. Iterative and one-shot tracks are kept separate.

Reviewed 2026-10-27

Evidence available for this benchmark

This source record contains 77 numeric observations. The source distinguishes 10 model identities or configurations. A model may appear more than once; check the version, tools and run budget before comparing. Read results and row-level evidence →

Reported metric: Total score S = 0.3 S_in + 0.7 S_out (0–100)
Recorded unit: points
Metric direction: higher is better

Start with these resources

Catalog source · lab.einsia.ai

Cited documents and dates (1)

Quant-Arena paper v1 · configuration results and proprietary reference · 2026-10-21

Evaluation context and evidence

Source: Model reports
Publisher: Not recorded
Benchmark release: 2026-10-21
Collected: Not recorded
Modality: Not recorded

These are source-reported observations, not independent reruns by this site. Check the task version, evaluation protocol, unit and reporting date before comparing models. Results from different setups are not a universal ranking.

Reported results 77

Shown in source observation order, not ranked. Dates retain the precision provided by the source.

Model	Value	Unit	Protocol	Reported date and basis	Source
Kepler-1 (proprietary strategy)	77.56	points	Kepler-1 proprietary reference strategy; construction undisclosed; cash-account evaluator; not a supported-effort LLM run	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Opus 5.5	39.89	points	iterative research; at most five research-window candidate evaluations; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Sonnet 5.5	40.5	points	iterative research; at most five research-window candidate evaluations; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
DeepSeek V4.1 Flash	35.97	points	iterative research; at most five research-window candidate evaluations; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GLM-5.3	38.67	points	iterative research; at most five research-window candidate evaluations; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-5.6 Terra	41.54	points	iterative research; at most five research-window candidate evaluations; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Astra	43.7	points	iterative research; at most five research-window candidate evaluations; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Sol	42.89	points	iterative research; at most five research-window candidate evaluations; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Gemini 3.8 Flash	35.39	points	iterative research; at most five research-window candidate evaluations; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Kimi K3	31.43	points	iterative research; at most five research-window candidate evaluations; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Opus 5.5	30.15	points	iterative research; at most five research-window candidate evaluations; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Sonnet 5.5	34.97	points	iterative research; at most five research-window candidate evaluations; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
DeepSeek V4.1 Flash	36.45	points	iterative research; at most five research-window candidate evaluations; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-5.6 Terra	41.19	points	iterative research; at most five research-window candidate evaluations; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Astra	47.22	points	iterative research; at most five research-window candidate evaluations; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Sol	38.51	points	iterative research; at most five research-window candidate evaluations; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Gemini 3.8 Flash	41.46	points	iterative research; at most five research-window candidate evaluations; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Kimi K3	38.31	points	iterative research; at most five research-window candidate evaluations; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Opus 5.5	41.39	points	iterative research; at most five research-window candidate evaluations; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Sonnet 5.5	40.93	points	iterative research; at most five research-window candidate evaluations; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
DeepSeek V4.1 Flash	35.9	points	iterative research; at most five research-window candidate evaluations; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GLM-5.3	38.91	points	iterative research; at most five research-window candidate evaluations; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-5.6 Terra	47.58	points	iterative research; at most five research-window candidate evaluations; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Astra	50.63	points	iterative research; at most five research-window candidate evaluations; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Sol	45.31	points	iterative research; at most five research-window candidate evaluations; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Kimi K3	31.86	points	iterative research; at most five research-window candidate evaluations; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Opus 5.5	44.28	points	iterative research; at most five research-window candidate evaluations; reasoning effort medium; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Sonnet 5.5	39.42	points	iterative research; at most five research-window candidate evaluations; reasoning effort medium; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-5.6 Terra	39.5	points	iterative research; at most five research-window candidate evaluations; reasoning effort medium; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Astra	47.17	points	iterative research; at most five research-window candidate evaluations; reasoning effort medium; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Sol	49.8	points	iterative research; at most five research-window candidate evaluations; reasoning effort medium; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Gemini 3.8 Flash	43.04	points	iterative research; at most five research-window candidate evaluations; reasoning effort medium; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
DeepSeek V4.1 Flash	22.0	points	iterative research; at most five research-window candidate evaluations; reasoning effort none; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-5.6 Terra	33.09	points	iterative research; at most five research-window candidate evaluations; reasoning effort none; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Opus 5.5	41.92	points	iterative research; at most five research-window candidate evaluations; reasoning effort xhigh; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Sonnet 5.5	40.1	points	iterative research; at most five research-window candidate evaluations; reasoning effort xhigh; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-5.6 Terra	44.77	points	iterative research; at most five research-window candidate evaluations; reasoning effort xhigh; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Astra	51.29	points	iterative research; at most five research-window candidate evaluations; reasoning effort xhigh; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Sol	51.14	points	iterative research; at most five research-window candidate evaluations; reasoning effort xhigh; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Opus 5.5	36.07	points	one-shot generation; no research-window candidate feedback; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Sonnet 5.5	40.24	points	one-shot generation; no research-window candidate feedback; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
DeepSeek V4.1 Flash	29.63	points	one-shot generation; no research-window candidate feedback; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GLM-5.3	38.02	points	one-shot generation; no research-window candidate feedback; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-5.6 Terra	40.61	points	one-shot generation; no research-window candidate feedback; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Astra	35.22	points	one-shot generation; no research-window candidate feedback; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Sol	23.15	points	one-shot generation; no research-window candidate feedback; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Gemini 3.8 Flash	30.26	points	one-shot generation; no research-window candidate feedback; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Kimi K3	32.01	points	one-shot generation; no research-window candidate feedback; reasoning effort high; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Opus 5.5	35.04	points	one-shot generation; no research-window candidate feedback; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Sonnet 5.5	37.75	points	one-shot generation; no research-window candidate feedback; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
DeepSeek V4.1 Flash	29.36	points	one-shot generation; no research-window candidate feedback; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-5.6 Terra	27.82	points	one-shot generation; no research-window candidate feedback; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Astra	30.08	points	one-shot generation; no research-window candidate feedback; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Sol	32.15	points	one-shot generation; no research-window candidate feedback; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Gemini 3.8 Flash	27.75	points	one-shot generation; no research-window candidate feedback; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Kimi K3	31.58	points	one-shot generation; no research-window candidate feedback; reasoning effort low; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Opus 5.5	39.7	points	one-shot generation; no research-window candidate feedback; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Sonnet 5.5	33.89	points	one-shot generation; no research-window candidate feedback; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
DeepSeek V4.1 Flash	29.92	points	one-shot generation; no research-window candidate feedback; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GLM-5.3	31.46	points	one-shot generation; no research-window candidate feedback; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-5.6 Terra	20.98	points	one-shot generation; no research-window candidate feedback; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Astra	31.47	points	one-shot generation; no research-window candidate feedback; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Sol	22.31	points	one-shot generation; no research-window candidate feedback; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Kimi K3	31.65	points	one-shot generation; no research-window candidate feedback; reasoning effort max; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Opus 5.5	42.61	points	one-shot generation; no research-window candidate feedback; reasoning effort medium; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Sonnet 5.5	33.82	points	one-shot generation; no research-window candidate feedback; reasoning effort medium; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-5.6 Terra	21.01	points	one-shot generation; no research-window candidate feedback; reasoning effort medium; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Astra	31.02	points	one-shot generation; no research-window candidate feedback; reasoning effort medium; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Sol	22.6	points	one-shot generation; no research-window candidate feedback; reasoning effort medium; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Gemini 3.8 Flash	35.91	points	one-shot generation; no research-window candidate feedback; reasoning effort medium; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
DeepSeek V4.1 Flash	20.48	points	one-shot generation; no research-window candidate feedback; reasoning effort none; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-5.6 Terra	42.14	points	one-shot generation; no research-window candidate feedback; reasoning effort none; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Opus 5.5	39.13	points	one-shot generation; no research-window candidate feedback; reasoning effort xhigh; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
Claude Sonnet 5.5	37.05	points	one-shot generation; no research-window candidate feedback; reasoning effort xhigh; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-5.6 Terra	34.21	points	one-shot generation; no research-window candidate feedback; reasoning effort xhigh; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Astra	30.3	points	one-shot generation; no research-window candidate feedback; reasoning effort xhigh; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference
GPT-6 Sol	23.29	points	one-shot generation; no research-window candidate feedback; reasoning effort xhigh; frozen strategy; cash-account evaluator	2026-10-21 · document_publication	Quant-Arena paper v1 · configuration results and proprietary reference

Research, commit, then assess

Total score is not return

Read every effort level

Iterative research · complete configurations

One-shot generation · complete configurations

What this evidence establishes

Original publication

Evidence available for this benchmark

Start with these resources

Evaluation context and evidence

Reported results 77

Explore related benchmarks

Put this benchmark in context

Six Navers Lab / Einsia benchmarks and the work they test

AI benchmarks and leaderboards are different things

How to compare AI benchmark scores without mixing protocols

Choose benchmarks for coding, cybersecurity and computer use agents

Four new agent benchmarks: adaptation, safety and video

Continue exploring