Finance Agent (v2)

Name: Finance Agent (v2)
Creator: Vals AI

Whether a model can answer the questions entry-level financial analysts face about public-company filings: finding the right source, applying the right industry convention and carrying precise numbers through several steps.

Financial analysisVals AI73 models run by Vals AIUpdated 29 September 2026Reported by Google DeepMind

How it is run

Finance Agent v2 carries 927 expert-reviewed questions over public, private-validation and held-out test splits, and Vals reports two scoring rules side by side: partial credit, and a stricter all-pass rate where every required element must be present. Nine question categories, including earnings analysis, financial modelling, precedents and comparables.

What it cannot tell you

Partial-credit and all-pass scores can order models differently. All-pass requires every required element to be correct, so a high partial-credit result can still include incomplete answers. One of Opus 5's 1,350 run-task results came from a refusal fallback but had already scored zero, so the published number is unchanged.

Leader: Gemini 4 Argon65.40% · Vals AI's own list
Top-ten spread: 7.30%Leader minus tenth place, or last place in smaller fields.
Median: 50.39%of the 73 rows published
No published result: 9of 46 models on this site

Independent leaderboard · Vals AI

Top 73 of 73 runs by Vals AI, one harness.

#	Model	Score
1	Gemini 4 Argon	65.40%
2	Gemini 3.8 Flash	61.44%
3	Muse Spark 1.2	60.60%
4	Muse Spark 1.3 Max	59.96%
5	Gemini 3.7 Flash	59.04%
6	Muse Spark 1 3	58.90%
7	Claude Fable 5.1	58.88%
8	Claude Opus 5	58.63%
9	Claude Opus 5.5	58.59%
10	Claude Sonnet 5.5	58.10%
11	Gemini 3.5 Flash	57.86%
12	GLM 5.3 Flash	57.85%
13	MiMo V2.6 Pro	57.34%
14	Muse Spark 1 1	57.21%
15	Claude Fable 5	56.31%
16	Gemini 3.6 Flash	56.30%
17	MiMo-V2.6-Flash	56.28%
18	GLM 5.3	55.84%
19	Hunyuan Hy4 Preview	55.06%
20	GPT-5.6 Luna	55.04%
21	Ling 3.0 Flash Af Rc3	54.93%
22	GPT-5.6 Terra	54.44%
23	Kimi K3	54.36%
24	Claude Opus 4.8	53.92%
25	Claude Sonnet 5	53.91%
26	GPT-5.6 Sol	53.76%
27	Grok 4.6	53.68%
28	GPT-6 Astra	53.54%
29	DeepSeek V4.1 Flash	53.48%
30	Grok 4.7	52.25%
31	GPT-6.1 Sol	52.03%
32	Ember 1	51.85%
33	GPT-5.5	51.76%
34	Claude Opus 4.7	51.51%
35	Claude Sonnet 4 6	51.03%
36	Qwen3.8 Max	50.59%
37	DeepSeek V4 Pro	50.39%
38	GPT-6 Luna	49.87%
39	Glm 5.2	49.70%
40	Deepseek V4 Flash 0731	49.52%
41	GPT-6 Sol	49.05%
42	Qwen3.8 27b	48.55%
43	Grok 4.5	48.35%
44	MiniMax M3	48.27%
45	Qwen3.7 Max	47.78%
46	Gemini 3.5 Flash Lite	47.44%
47	Inkling	46.60%
48	Gpt 5.4 Mini 2026 03 17	45.36%
49	Kimi K2.6	44.90%
50	Glm 5.1	44.79%
51	Deepseek V4 Pro	44.08%
52	Gemini 3.1 Pro	42.98%
53	Gemini 3 Flash Preview	42.55%
54	Mimo V2.5 Pro	41.50%
55	Inkling Small	41.26%
56	Qwen3.6 Plus	40.85%
57	Qwen3.7 Plus	38.22%
57	Gpt 5.4 Nano 2026 03 17	38.22%
59	Grok 4.3	37.73%
60	Nemotron 3 Ultra 550b A55b	37.67%
61	Mimo V2.5	36.73%
62	Kimi K2.5 Thinking	35.79%
63	Mistral Medium 3.5	32.10%
64	Claude Haiku 4 5 20251001 Thinking	31.01%
65	Ling 3.0 Flash 2607	30.31%
66	Gemini 3.1 Flash Lite Preview	29.99%
67	Grok 4.20 0309 Reasoning	28.49%
68	MiniMax M2.7	27.89%
69	Laguna M.1	25.03%
70	Mercury 2.5	18.89%
71	Nemotron Lightning 3p5 30b A3b	18.50%
72	Laguna Xs.2	15.60%
73	Command A Plus 05 2026	9.04%

Source: www.vals.ai/benchmarks/fabv2

No published result tracked

9 of the 46 models on this site have no published number on this benchmark, so their cells in the home table stay blank rather than reading as zero.

Step 5 Preview Qwen3.8 2.4T A95B Qwen3.8-Flash-Next Qwen3.8 Max (0902)GPT-5.4 DeepSeek V4.1 Flash 1017 GLM-Legal Claude Haiku 5.5 Beam

Read this with

Nothing else on this site shares this benchmark's family or category, so these are other evaluations the same publisher reports.

As reported by each publisher

The same benchmark, measured by the labs themselves. Compare down a column, never across two publishers.

Google DeepMind

Gemini 4 Argon: model evaluation · October 2026

Rank	Model	Score
1	Gemini 4 Argon	65.4%
4	GPT-6 Astra	53.5%
2	Claude Fable 5.1	58.9%
3	Claude Opus 5.5	58.6%

Ranks compare reported scores in the same published row; ties share a rank.

Vals Finance Agent v2. Quoted by Google from Vals AI; retain the report’s scoring label.

Page 5 of Google DeepMind’s model evaluation, read 1 October 2026. Gemini runs use the API at the highest thinking setting, generally pass@1 without majority voting or parallel test-time compute; smaller benchmarks average multiple trials. Other models use their highest available published reasoning results. Google quotes Vals AI for knowledge-work and Vibe Code results, Zapier’s private-set AutomationBench leaderboard, Proximal for FrontierSWE v2, Surge AI for RiemannBench and no-tools Chartography, and the CWE-bench v1 leaderboard. Argon’s DeepSWE uses mini-swe; Terminal-Bench 4.0 is self-computed for Argon and quoted from the author board for other models. PostTrainBench v1.1 uses OpenCode with one H100 and a 10-hour budget for all models. Terminal-Bench Science uses a 6x verifier timeout for Argon. LABBench 2 uses terminal, bioinformatics tools, Python, R and internet access for all models. GraphWalks uses BFS F1 on two separate subsets: 650 items up to 128k tokens and 200 items from 256k to 1M. Agent’s Last Exam uses ALE-Claw, a 5-hour window and binary pass rate, with safety filters enabled for Argon. OSWorld 2.0 uses offline partial reward, the best of three single-attempt runs for Argon, safety filters, 1080p, 500 steps, Gemini CUA with parallel batch tools and compaction, screenshot-only observation and the 08.08 patch; Claude’s combined online/offline results are omitted. LVBench runs without tools, using 1 FPS for Gemini, 800 frames for Astra, 300 for Fable and 600 for Opus. Missing cells remain blank. These are the report’s published figures, preserved separately from this site’s independent source snapshots.

storage.googleapis.com/deepmind-media/gemini/gemini_4_argon_model_evaluation.pdf

Sources

Every page used for this benchmark, read on 17 October 2026.

Benchmark site or the page that publishes itwww.vals.ai/benchmarks/fabv2
Independent resultswww.vals.ai/benchmarks/fabv2
Google DeepMind tablestorage.googleapis.com/deepmind-media/gemini/gemini_4_argon_model_evaluation.pdf