Study how frontier AI models reason about stocks. Put your own reasoning on the same bench.

Claude’s Season 2 BUY on ADBE was driven by a cheap-valuation thesis plus solid fundamentals. In the June 29 decision, Claude said Adobe traded at a deeply discounted valuation: P/E 11.6x, forward P/E 8.31x, PEG 0.554, revenue TTM $25.2B, profit margin 28.7%, ROE 62.9%, EPS $17.48, and an analyst target of $282 versus price $202.73. The judge scored the decision 66 and noted the next ADBE decision moved +1.83% the following day. Other ADBE claims in the packet reinforce the same pattern: supported notes on profit margin 28.7%, ROE ~63%, EPS ~$17.48, and a later statement that ADBE was still below the analyst target. Some valuation figures were only partially supported and a momentum claim about the 52-week high was contradicted.

What the data suggests: The packet suggests Claude bought Adobe because it viewed the stock as undervalued relative to earnings, growth, profitability, and analyst targets. The strongest evidence is consistent on low-valuation and quality metrics, while some exact price-ratio details were only partially supported. This is a small sample from one season and one ticker, so it shows the model’s reasoning style more than a broad pattern; judge scores also vary across claims.

On the latest judge evaluations, Anthropic Claude Sonnet 4.6 scored 72 on claim grounding versus OpenAI GPT-5 at 68, so Claude leads by 4 points. Claude had 64 evaluations and GPT-5 had 61. The judge summaries also differ in flavor: Claude was described as a coherent hybrid value+momentum process with strong metric use but weakened by record-keeping errors and stop-execution issues; GPT-5 was described as disciplined and value-driven, but with a few metric inaccuracies and rising concentration risk. Both had the same judge spread of 13 points. Across example claims, Claude has both clearly supported grounding cases and unverifiable ones, while GPT-5 shows some well-grounded decision records but also at least one sector-narrative claim judged not verifiable.

What the data suggests: The data suggests Claude is slightly stronger on evidence grounding in the latest evaluation, but the gap is small and should be treated cautiously. The sample is limited to the benchmark records provided, with a single latest date and mixed claim types. Judge disagreement remains material: both models have a 13-point spread, and GPT-5’s median grounding was higher than its latest score, so the ranking is not perfectly stable over time.

The biggest judge disagreement in the packet was on xAI Grok 4.3. The widest single-record spread was 45 points on 2026-08-01, where OpenAI scored it 60, Anthropic 42, and xAI 87. Other large splits for the same model were 42 points on 2026-08-27 and 40 points on 2026-08-21. Across all 315 panel-scored evaluations, the overall composite range across judges ran from 2 to 45, with a median of 21. By model, xAI Grok 4.3 had the highest median judge spread at 28 over 65 evaluations, above Google Gemini 3.1 Pro at 21 over 62, Anthropic Claude Sonnet 4.6 at 20 over 63, Google Gemini 3.5 Flash at 20 over 62, OpenAI GPT-5 at 15 over 60, and OpenAI GPT-6 Astra at 5 over 3.

What the data suggests: The data suggests judges disagreed most on xAI Grok 4.3, especially in late-summer evaluations where the xAI judge gave much higher scores than OpenAI and Anthropic. The pattern points to disagreement over reasoning quality, risk control, and evidence grounding rather than over final outcomes alone. Caveats: this is a benchmark sample, not a full market study; the spread measures panel disagreement on scored evaluations; and the highest-spread model also appears multiple times, so one model’s repeated runs may dominate the top end.

Strong reasoning sometimes led to poor outcomes, but it was not common. Across 1,494 matched BUY/SELL decisions, strong reasoning had 81 poor follow-up outcomes versus 75 good and 99 flat. The median follow-up change for strong-reasoning decisions was -0.67%, while weak-reasoning decisions had a median of 0%. Examples include OpenAI GPT-5’s ACN SELL on 2026-07-02 (judge score 75), which was later followed by a -38.3% same-symbol move, and Google Gemini 3.5 Flash’s APP BUY on 2026-07-24 (judge score 68), followed by -22.76%. GPT-5’s ACN SELL on 2026-07-06 scored 78 but later mapped to -32.03% and its 2026-07-19 ACN SELL scored 74 with -26.31% later change. Another strong case, APP BUY on 2026-07-25, scored 71 and later saw -21.41%.

What the data suggests: The data suggests that high judge scores do not guarantee favorable subsequent price moves; even coherent, well-grounded reasoning can precede poor outcomes. However, the pattern is modest, not decisive: strong-reasoning poor outcomes are only slightly more common than strong good outcomes, and the benchmark is based on a single matching framework with judge scores that themselves show spread/disagreement. The follow-up change is also tied to the next same-symbol decision, so timing and symbol-specific volatility matter.

Contestants most often ignored concentration risk. The benchmark records 603 ignored-risk mentions across 9 models, far above valuation at 180 mentions and stop/exit discipline at 155. The next most common misses were momentum/technical signals at 106, earnings/guidance at 84, and liquidity/cash at 57. The judge notes repeatedly describe oversized or repeated adds to single names, especially without explicit sizing or beta-aware controls. Valuation errors were also common, including buying expensive names while citing growth. Stop/exit discipline and momentum were frequent blind spots, but less so than concentration.

What the data suggests: The data suggests concentration and process discipline were the most persistent weak spots across the evaluated runs, with valuation a clear secondary issue. This is aggregate evidence across 367 evaluations, not a ranking of every model’s failures, and the examples span different seasons, styles, and judge passes. Some models had better risk awareness scores than others, but judge spread and short-horizon runs mean the exact importance of each risk type can vary by case.