ChatGPT, Claude, Gemini or Perplexity for earnings season? What four published tests show

Every earnings season, traders paste a press release into a chatbot and ask what matters. Which one should get the job? We read four tests published in 2026 that put ChatGPT, Claude, Gemini and Perplexity on real earnings material. None of them was run by us, and none agrees with the others on a winner.

Test 1: a 24-answer quiz with Gemini on top

The newsletter AI Investing Lab published a test on March 3, 2026 using one prompt across six assistants and three companies (NVIDIA, Apple, Rocket Lab), with eight questions each: fiscal period, revenue versus expectations, EPS, margins, CEO, headquarters, current ratio and debt-to-equity.

The sample is small (8 questions x 3 companies per tool), the model versions are not named and the scoring rubric sits in a downloadable spreadsheet rather than the post. The author's most useful observation is qualitative: every tool did worse on Rocket Lab, the smallest and least covered company.

Test 2: 500 single numbers, and the ranking flips

Daloopa, a financial data vendor, ran 500 single-number questions drawn from more than 4,000 companies through each chatbot's web app, with web search on. The models were the versions of that run: GPT-5 Thinking, Claude Opus 4.1, Gemini 2.5 Pro, Grok 4 and Perplexity Finance.

Here Gemini finishes last, the reverse of Test 1 nContext: n=500 single-number questions · test run with GPT-5 Thinking, Claude Opus 4.1, Gemini 2.5 Pro era models (date not shown) · Daloopa benchmark · margin: not disclosed. Daloopa attributes most Gemini misses (352 of its errors) to misreading the question, and most Claude misses to refusals (159). On the 170 companies whose fiscal year differs from the calendar year, ChatGPT matched 42.9% exactly nContext: n=170 companies with non-calendar fiscal years · same run · Daloopa benchmark · margin: not disclosed. Two caveats: Daloopa sells the tool that tops its own table, and it published the full question set and conversation links on Hugging Face so the result can be rechecked.

Test 3: through the API, near-perfect but not perfect

Financial Datasets, another data vendor, published EarningsBench on June 29, 2026: 60 earnings releases across 12 sectors, with each model asked to extract sector KPIs such as load factor, net interest margin or FFO per share, checked by hand. This is not the chat app; it is the API with a detailed instruction set refined over 10 rounds.

Claude Opus 4.6 with extended thinking reached 99.3% accuracy at $0.249 per release; GPT 5.4 on high reasoning reached 98.4% at $0.210; Sonnet 4.6 reached 95.3% nContext: n=60 earnings releases, 12 sectors · published Jun 29, 2026 · Financial Datasets EarningsBench · margin: not disclosed. Four sectors were solved by every model. Insurance was not: Sonnet 4.6 scored 74% there without extended thinking nContext: n: insurance releases within the 60, count not stated · Jun 29, 2026 · Financial Datasets EarningsBench · margin: not disclosed. The vendor's conclusion is the line worth keeping: 99.3% nContext: n=60 releases · Jun 29, 2026 · Financial Datasets EarningsBench · margin: not disclosed is still not good enough to ship without validation.

Test 4: one earnings day, two failures worth knowing

Ben Dixon's AI Reliability Tracker ran ChatGPT, Claude and Perplexity on Meta's Q1 2026 release on May 15, 2026. Perplexity Pro was cleanest on live retrieval and flagged on its own that some outlets quoted GAAP EPS of $10.44 and others adjusted EPS near $7.31. It also ran 10 web searches when told to work only from a pasted document. The author later published a correction: his pasted summary contained three wrong figures, so the document-reading results are not a clean ranking.

In a separate dated test, Perplexity reported a small company's annual revenue as $6K on May 14 because it skipped the filing's 'in thousands' label; the real figure was $6.095 million. The same query returned the right number on June 18.

Why the rankings disagree

• Different jobs. Test 1 mixes facts and ratios; Test 2 asks for one exact reported number; Test 3 extracts KPIs from a supplied document; Test 4 is a live session.
• Different products. A chat app with web search and an API call with a tuned prompt are not the same tool, even when the model name matches.
• Model churn. Test 2 used models that have since been replaced. A ranking from spring can be stale by fall.
• Small samples. Only Test 2 crosses a few hundred questions. None reports confidence intervals.

What it costs (prices checked October 14, 2026)

Claude Pro lists at $20 a month billed monthly, or $17 a month on the annual plan ($200 up front); Max starts at $100 a month (claude.com/pricing). ChatGPT lists Free, Go, Plus and Pro tiers and Google lists AI Plus, Pro and Ultra, but the dollar amounts did not appear in the page versions we opened, so we do not print them. Perplexity's Pro page did not load for us. For how current models score on public finance benchmarks, see our AI Model Index.