Frontier AI models beat analyst consensus on earnings surprises, Samaya study finds

What happened

Samaya, a company building AI tooling for financial research, says the newest frontier language models have for the first time beaten professional analysts at forecasting company earnings. According to the company's research write-up republished on dev.to, the three most recent models it tested — GPT-6 Astra, Claude Fable 5.1 and Claude Opus 5.5 — produced smaller forecast errors than analyst consensus across 456 second-quarter earnings releases, and also came out ahead of a deliberately strengthened version of that consensus.

The claim is narrower than beating the market: the models were scored on predicting the figures companies actually report — revenue, gross margin, operating income and adjusted EPS — rather than on trading returns. But the authors present the result as a first for the field and a sign that AI capability has crossed into one of the most demanding areas of professional knowledge work.

How the evaluation worked

Each prediction was made one week, or five trading sessions, before a company reported, and the models could draw only on information published up to that moment — the same information a human analyst would have had. The tested universe covered 456 companies from every sector that reported results from July 14 onward, a date chosen because it falls after the knowledge cutoff of all seven models in the study. Companies needed a market capitalisation above $5 billion and estimates from at least eight brokers. Alongside the three newest models, the study included GPT-5.6 Sol, Claude Sonnet 5, Gemini 3.8 Flash and Kimi K3.

The hardest design problem was stopping hindsight from leaking in. Samaya enforced a hard temporal cutoff at the harness layer using an authentication token, which the company says makes it impossible for a model to retrieve anything published after its prediction date. General web access was removed because it cannot be reliably gated this way, with essential news sources served behind the cutoff as a substitute.

Two baselines, and the results

Predictions were scored three ways: error against the reported actuals, correlation between the predicted deviation from consensus and the actual deviation, and a hit rate for landing on the same side of consensus as the real result. Errors were normalised by each company's historical surprise volatility so that erratic businesses did not dominate the averages.

The first baseline was raw analyst consensus, the averaged forecasts the market already prices in. Samaya notes that analysts systematically shade their estimates down before earnings, so reported figures beat consensus more often than they miss — meaning clearing that bar alone proves relatively little. The team therefore built a tougher baseline by adding each company's historical median surprise back onto consensus.

Even older models such as Claude Sonnet 5 and Kimi K3 outperformed the raw consensus. Only the newest three also beat the corrected baseline, and the write-up reports a clear trend of error shrinking as model capability increases.

Why it matters

Consensus estimates anchor how investors value listed companies, and the write-up puts global stock markets at more than $150 trillion. Because even the best analyst averages miss, the gap between consensus and reported numbers — the surprise — is what moves prices, often sharply. A system that reliably anticipates surprises has direct value for investors hunting an edge, for analyst teams automating part of the modelling pipeline, and for anyone deciding where human research effort is still worth spending.

The finding also matters beyond finance. Earnings prediction is an open-ended research task: it requires searching scattered, time-stamped sources, weighing competitors, supply chains and macro conditions, and committing to explicit assumptions. Frontier models clearing the expert bar here suggests their strengths are moving from self-contained benchmarks toward the judgement-heavy work that defines expensive professional roles.

Caveats are worth holding onto. The evidence covers a single quarter of releases; the study was designed, run and reported by Samaya, whose own product line is this kind of finance harness, so independent replication will be the real test; and out-forecasting averaged consensus is not the same as earning market-beating returns, since consensus expectations are already embedded in prices.