A Meta AI Model That Runs on One Graphics Card Edged Out Claude Fable at Picking US Stocks

WILMINGTON, Del., November 12, 2026. FinzData, the market-data company built for AI, quants, data scientists and financial analysts, today announced the results of its AI Trading Championship, a head-to-head test of 11 leading AI models as stock-market investors.

Meta's Muse Glimmer, a 30-billion-parameter open model that Meta released in August to run on a single consumer graphics card, finished second. Its strategy returned 15.5% a year from January 2019 to September 2026 after trading costs, with a Sharpe ratio of 0.75. It finished just ahead of Anthropic's flagship Claude Fable (Sharpe 0.74) and ahead of Claude Opus and Claude Sonnet. Only NVIDIA's Nemotron-3 Ultra did better, with a Sharpe ratio of 0.81 and 16.1% a year.

"Everyone is asking which AI is best at trading. We wanted an answer that holds up: the same data, the same rules, and years of market history no model had seen," said a FinzData spokesperson. "The answer is not simply the biggest model. A model you can download for free and run at home placed second, ahead of flagship models from the largest AI labs."

How the championship worked

Each model designed one strategy for US stocks with a single rule: use FinzData's point-in-time data, covering more than 16,000 US stocks since 2000, delisted companies included. Point-in-time means each model saw only what an investor could have known on the day, and companies that later failed stay in the record. Seven models were served by NVIDIA NIM and four came from Anthropic. Every strategy ran on the same engine with the same trading costs. The models designed on 2011 to 2018 data and were scored once on January 2019 to September 2026, a period none of them had seen.

Results (Sharpe ratio, return a year)

• 1st: Nemotron-3 Ultra, NVIDIA: 0.81, 16.1%
• 2nd: Muse Glimmer 30B, Meta: 0.75, 15.5%
• 3rd: Claude Fable, Anthropic: 0.74, 14.0%
• 4th: GPT-OSS-20B, OpenAI: 0.69, 14.0%
• 5th: Gemma 4 31B, Google: 0.60, 11.1%
• 9th: Claude Opus, Anthropic: 0.57, 8.9%

What set Muse Glimmer apart

Muse Glimmer built the broadest strategy in the field. It bought cheap, profitable companies with little debt, recent price strength and low volatility, five ingredients where most models used two or three. It was the only model to use company debt at all. Its record on the design years was modest (Sharpe 0.19), and it did far better on the years it had not seen, the opposite of a strategy tuned to its backtest.

What investors should take from it

• Bigger is not better. Claude Opus, Anthropic's largest model, finished 9th of 11.
• AI models think alike. Ten of the 11 models leaned on the same input, earnings yield.
• Beating the market is hard, even for AI. Seven of the 11 strategies trailed a simple equal-weight portfolio of the 500 largest US stocks (Sharpe 0.63).
• AI judges graded the scoreboard. Three AI judges scored every rival strategy, and their scores matched the results they were shown almost exactly (rank correlation 0.82 to 0.99).

"These are simulations on historical data, not live money, and we publish every number, including where luck could explain a result," the spokesperson added. "As brokers open the door to AI agents that trade for people, investors deserve evidence about which models make sound decisions. That is what FinzData is built for."

The timing matters. Anthropic released Claude Fable 5.2 on November 9, after the championship was scored. The Claude Fable entry here is the earlier version, and Round 2 is open to new entrants, Fable 5.2 included. Brokers such as Robinhood now let customers' AI agents trade stocks, and FinzData's data reaches those agents directly through its MCP server.

Why AI models, quants and traders build on FinzData

The championship ran on the same data FinzData customers use every day. Five things set it apart:

• Results you can trust. Every record is point-in-time and survivorship-free: more than 16,000 US stocks since 2000, with more than 9,000 delisted names kept and fundamentals as first filed. A backtest sees only what was knowable that day, so it reflects the truth and is never flattered by hindsight.
• Built for AI. AI agents connect directly through FinzData's MCP server in Claude and ChatGPT, with integrations for Grok, Perplexity, Cursor and Windsurf. Developers get the same data in one line, pip install finzdata, plus a Rust crate and a REST API.
• The whole market in one place. Stocks, 5,100+ ETFs, 139 futures markets, options chains, point-in-time dealer gamma (GEX) for more than 11,200 stocks, ETFs and indexes, 123 indexes, FX, crypto, more than 22,700 macro series with every vintage, COT positioning and 13F holdings, in bars from 1-minute to daily.
• The real economy, day by day. FinzData's Global Macro Pulse shows activity before the official statistics arrive: ship transits through 28 chokepoints and port calls at more than 2,000 ports, about 200,000 flight legs a day across 7,400+ airports, and toll and congestion-zone traffic from New York to Hong Kong. Electricity demand covers 87 countries and every US grid operator.
• Yours to keep, at the industry's best value. A free API key needs no card. History starts at $19 for one ticker and $195 for five years of the whole US market, and what customers buy stays theirs to download and take to any backtester or cloud.

Every strategy, result and test is public:

• Interactive results
• Full research paper
• Re-run any entry on the FinzData Hub
• Free API key, no card

About FinzData

FinzData provides point-in-time, survivorship-free market data built for AI, quants, data scientists and financial analysts who backtest. FinzData offers the industry's best breadth and depth of data at the best value. Clean Data. Real Insights. Real Alpha. finzdata.com

Media contact

The FinzData Team · [email protected] · Logos and company description

Results are historical simulations on FinzData data, after trading costs. Past performance does not guarantee future results. This release is for information only and is not investment advice.

← All news