AI agents, Claude and ChatGPT

Best LLM for trading: how Claude, ChatGPT, Grok and Gemini compare

Claude, ChatGPT, Grok and Gemini all read a chart, search the web and write code, and none of them trades for you by itself. Here is how they compare task by task, and what happened when each one traded real money in nof1’s Alpha Arena.

[

November 16, 2026

]
•
[

15

 

min to read

]

The short answer

No single model wins. Claude, ChatGPT, Grok and Gemini all search the web with citations, read a chart image and write code. In nof1’s Alpha Arena, a model family that lost money in the first season won the second. Pick by task, test on your own charts, and keep the trade decision in code, not the model.

Which LLM is best for trading?

None of them, consistently. As of November 16, 2026, Claude, ChatGPT, Grok and Gemini each search the web with sources, take a chart screenshot and write code, and the gaps between them are about plumbing (what connects, what remembers, what runs on a schedule) more than about which one is “smarter” at markets. The table below is the overview; each row gets its own section next.

Claude, ChatGPT, Grok and Gemini on four trading tasks, November 16, 2026
TaskClaude (Anthropic)ChatGPT (OpenAI)Grok (xAI)Gemini (Google)
Research with sourcesWeb search with citations on every plan; Research mode on paid plansWeb search with citations on every plan; deep research, limited on Free and GoX Search plus web search; X Search can narrow to handles or a date rangeGrounding with Google Search, cited; Deep Research in the Gemini app
Reads a chart imageYes; Anthropic says to verify high-stakes readsYes; OpenAI says it struggles with precise spatial detailYes, on Grok 4.7Yes, on the Gemini 4 family; no chart-specific caveat found
Coding or building a botClaude Code: terminal, IDE, desktop, web; custom Skills; MCPCodex: CLI, IDE extension, cloud, inside ChatGPT; MCPGrok Build: terminal UI or headless; runs Grok 4.7 by default; MCPAntigravity: CLI, IDE, app; runs on Gemini 4
Live trading resultsClaude Sonnet 4.5 lost 31.1% in Alpha Arena Season 1GPT-5 lost 62.7% in Season 1, the worst of sixGrok 4 lost 45.4% in Season 1; Grok 4.20 won Season 1.5Gemini 2.5 Pro lost 56.2% in Season 1, second worst

Two comparisons already exist and this page doesn’t repeat them: Claude vs ChatGPT for trading goes task by task between those two alone, and can Grok trade for you?, can Claude trade for you? and can ChatGPT trade for you? cover what each one can and can’t do with a broker connected. This page is the wider picture, across all four, plus the live and published results.

Which model researches a market best?

All four now search live, and all four cite what they read, so the honest answer is that the differences are in reach and habit, not in whether they can do it.

  • Claude turns on web search with citations on every plan, including Free, and reads a page you paste directly. Its slower Research mode, on paid plans, runs several searches before it answers.
  • ChatGPT matches that: web search with citations on every plan, and deep research for bigger questions, limited on Free and Go and full on Plus and Pro.
  • Grok is the outlier, because it also has X Search: keyword or meaning search across X, narrowed to up to 20 accounts or a date range, built for finding who posted what, and when, rather than general web reading.
  • Gemini grounds its answers in Google Search: the model decides whether a search would help, runs it, and returns text with inline citations that link a specific part of the answer to a source URL, according to Google’s own API docs. In the Gemini app, Deep Research goes further, reading up to “hundreds of websites” and, with your permission, your own Gmail, Drive and Chat, to write a multi-page report in minutes.

None of this is a price feed. A search result can quote a stale number, and a social post is a claim, not a fact. Run the same question through two of them and compare the sources they found, not just the summary:

Same research question, two assistants

Why did [ticker] move [up or down] about [X]% on [date]? Search the web. Give the three most likely reasons, each with a link and its publish date. Separate the company’s own statements from reporting and from opinion, and say clearly if you found no company-specific news.

For the deeper Claude vs ChatGPT comparison on this exact point, including how each handles a linked brokerage account, see Claude vs ChatGPT for trading.

Which model codes a trading bot or a backtest best?

Here the four companies have converged on the same shape: a terminal agent that reads your files, runs commands and talks to outside tools through MCP, the Model Context Protocol. Picking one is less about which model writes better Python (all four write usable Python) and more about which surface fits how you work.

The four companies’ agentic coding tools, as their own docs describe them, November 16, 2026
ToolMakerRuns inWhere it stands, November 2026
Claude CodeAnthropicTerminal, VS Code, JetBrains, a desktop app and the webGeneral purpose; connects MCP servers with one command and loads custom Skills from a folder
CodexOpenAIThe ChatGPT apps, a CLI, an IDE extension and Codex CloudBuilt-in agent approvals and an auto-review step; connects to MCP servers and to GitHub, Slack and Linear
Grok BuildxAIAn interactive terminal UI, headless in a script, or other apps through the Agent Client ProtocolRuns Grok 4.7 by default; xAI’s separate Grok Build 0.1, an “intelligent coding model for agentic software, engineering, and workflow tasks”, is in early access
AntigravityGoogleA CLI, an IDE and a desktop app, on macOS, Windows and LinuxRuns on the Gemini 4 family; Gemini 4 Hydrogen is pitched at “everyday coding and agentic workflows”

None of these four is a trading platform. Each one can scaffold a backtest script, call a broker’s API and run it on a schedule, but the broker connection, the paper-trading default and the approval settings differ tool by tool and broker by broker. Building a backtest or a bot with Claude Code specifically, including the broker side, is its own guide; see how to build an AI trading agent for onchain markets for the architecture that holds regardless of which coding tool writes it: the model proposes, and plain code, not a prompt, enforces the limits.

Which model reads a trading chart best?

All four take an image. None of them should be trusted to read an axis the way a trader does, and two of the four say so in their own documentation.

What each company’s own documentation says about reading images, November 16, 2026
ModelWhat its maker says about reading images
ClaudeAnthropic’s vision documentation says Claude “might hallucinate or make mistakes when interpreting low-quality, rotated, or very small images” and asks developers to verify its readings “especially for high-stakes use cases”
ChatGPTOpenAI’s API guide to vision says the model “struggles with tasks requiring precise spatial localization” and “may struggle to understand graphs” whose line colors or styles vary
GrokGrok 4.7 reads images as part of its standard multimodal input; xAI’s docs make no specific claim about reading charts or graphs
GeminiThe Gemini 4 family takes text, image, audio, video and PDF input; Google’s own docs make no specific claim about chart-reading accuracy either way

In practice, a chart screenshot is pixels, not numbers. Every model estimates a price from where a line sits against the axis labels, and none of them can see anything the frame cut off. The fix is the same for all four: crop to one chart, keep the axis and the ticker visible, and ask for every number with how sure the model is.

A chart-reading test you can run in any of the four

This is a [timeframe] chart of [ticker]. Using only what you can see in the image: 1) describe the trend in two sentences; 2) list the support and resistance levels you can see, with the price you read off the axis for each; 3) name any visible pattern or indicator signal and how sure you are; 4) say which parts of the image you can’t read clearly. Don’t use outside information, and don’t tell me whether to buy or sell.

Run it on a chart you already know well, in two of the four, and keep the one that misreads fewer levels for you. Can Claude trade for you? and can Grok trade for you? go further into each model’s own limits once a broker is connected.

How have these models done trading real money?

Nof1, an AI research lab, ran public contests in 2025 in which it gave several models real money and no human stepping in. The results are the closest thing to a head-to-head trading test these four families have, and they are a lesson in how fast a leaderboard moves.

Season 1: six models, crypto, two to three weeks

From October 17 to November 3, 2025, nof1 gave six models $10,000 each to trade Hyperliquid perpetual futures on six coins, with the same prompt and the same market data for all of them.

Alpha Arena Season 1, account values at the close on November 3, 2025, from nof1’s own data
ModelAccount at the closeChange from $10,000
Qwen3-Max$12,202+22.0%
DeepSeek V3.1$10,453+4.5%
Claude Sonnet 4.5$6,890−31.1%
Grok 4$5,456−45.4%
Gemini 2.5 Pro$4,376−56.2%
GPT-5$3,734−62.7%

All four companies in this comparison lost money, in this order from least to worst damage: Claude, Grok, Gemini and GPT-5, and the last three filled the bottom half of the table. The only two models to make money were Qwen3-Max and DeepSeek V3.1, which this page doesn’t cover further: it compares the four assistants above, not every model that has ever traded in public.

Season 1.5: US stocks, and a changed leaderboard

Nof1’s second season traded only US stocks, in four separate competitions, each with its own rules. Alpha Arena’s own site states the result plainly: “The official competition has ended as of December 3rd, 2025 at 5:00 PM EST. Models are no longer running. Mystery Model was the winner with a 12.11% aggregate return in 2 weeks.” Nof1’s leaderboard names that Mystery Model as Grok 4.20, released on xAI’s API on March 10, 2026, a different model from the Grok 4 that lost in Season 1. The older Grok 4 stayed in the contest and, in the Situational Awareness competition, fell to $385.02, down 96.15%, the lowest result on the whole 32-entry board.

Read next to each other, the two seasons say the same thing nof1 itself said before Season 1 even ended: results this short can come down to luck. A model family from the losing side of one season’s board topped the next one, under a newer version number. That is not a reason to ignore live results, but it is a reason not to crown a “best” model from one contest. Can Grok trade for you? has the full detail on both seasons, including how each model traded.

What do other benchmarks say about AI models trading?

Outside nof1’s contests, three academic benchmarks have tested LLMs as traders, and none of them hands the crown to one company either.

Published benchmarks of LLMs trading, read from their own papers on arXiv, November 16, 2026
BenchmarkDatedWhat it testedWhat it found
StockBenchPosted October 2, 2025; revised March 2, 2026Agents built on several models trading 20 large Dow stocks daily, March to June 2025Its own abstract: “most models struggle to outperform the simple buy-and-hold baseline, while some models demonstrate the potential to achieve higher returns and stronger risk management”. In its results table, 8 of the 13 models beat buying and holding all 20 (0.4%), none by much more than two points: Anthropic’s Claude Sonnet 4 returned 2.2%, OpenAI’s o3 1.9% and GPT-5 0.3%
AI-TraderPosted December 1, 2025Six mainstream LLMs trading across three real marketsIts own abstract: “general intelligence does not automatically translate to effective trading capability, with most agents exhibiting poor returns and weak risk management”, and that risk control, not raw intelligence, decided how well an agent held up across markets
FINSABERPosted May 11, 2025; last revised June 26, 2026LLM-based strategies backtested over two decades and more than 100 stocksIts own abstract: reported LLM advantages “deteriorate significantly” over a longer run and a wider set of stocks, with strategies “overly conservative in bull markets… and overly aggressive in bear markets, incurring heavy losses”

The pattern across all three, and across nof1’s own contests, is the same: being a strong general model doesn’t automatically make a good trader, and in FINSABER’s test an edge measured over a short window or a few stocks shrank once the test ran longer and wider. Treat any single benchmark, this page’s table included, as a snapshot, not a verdict.

How do you pick the right model for your trading tasks?

  1. Match the tool to the job, not the brand

    Research, coding and chart reading are different skills inside the same company’s lineup. The model that writes your best Pine Script may not be the one you’d trust to summarize a filing.

  2. Run your own test before you trust a leaderboard

    Use the chart-reading and research prompts on this page, in two models, on data you already understand, and keep the one that gets your specifics right more often.

  3. Don’t chase one season or one benchmark

    Alpha Arena’s own Season 1 writeup warned that “early successes may be the result of luck”. A model that wins once can lose the next time out, and FINSABER, above, found the same pattern over longer runs.

  4. Pin the exact model version

    Every company ships new versions under the same family name. If a model you’re calling by API changes under you without warning, your results change with it; pin the version you tested.

  5. Keep the trade out of the model

    None of the four discussed here holds a brokerage account, watches a position while you sleep, or enforces a stop by itself. That logic belongs in code or in a product built for it, never only in a prompt.

Where does an AI trading agent fit next to these four models?

Claude, ChatGPT, Grok and Gemini are general assistants that happen to be very good at reading and writing. An AI trading agent is built around the trade itself, with the market data, the account and the limits already in place. Ethy is one, the AI trading agent you ask before you trade, and it is not a fifth row in the table above: it doesn’t compete on which base model is smartest, because the chart read, the backtest and the order are part of the product, not something you have to wire up.

  • A structured chart read, not a screenshot guess. The daily, 4-hour, 1-hour and 15-minute charts together, with the trend on each, support and resistance, RSI and MACD divergences, Fibonacci levels, VWAP, volume nodes, patterns and an overall bias.
  • Research with sources, next to the market. A web search with citations, shown alongside the price and how the market’s peers moved.
  • Rules tested on real history. Describe an idea in plain words and Ethy backtests it on up to 365 days of data, with every entry and exit marked.
  • Your plan, read first. Your own rules for risk, markets, process and goals, checked at the start of every chat, with a flag when an idea goes against them.

It trades too: from the chat, Ethy places orders on Hyperliquid perpetual futures, which cover crypto plus the stocks, indices, commodities and currencies listed on trade.xyz, and it runs spot swaps. New users see a confirmation card before an order goes out, and that can be switched to place orders at once instead. Automations trade on their own once activated, live or in paper mode with no money at stake, with stop loss and take profit sitting on Hyperliquid itself so they fire even when Ethy is closed. Your balance lives in a wallet that belongs to you, and you can export its key at any time. Ethy is software for research and analysis, and it doesn’t give investment advice.

Summary

  • No model wins across the board: Claude, ChatGPT, Grok and Gemini all search the web with citations, read a chart image and write code, each through its own coding tool.
  • Two of the four publish a caution about reading images precisely; the other two make no specific claim either way, so check every number a model reads off a chart.
  • In nof1’s Alpha Arena Season 1, Claude, Grok, Gemini and GPT-5 all lost money trading crypto; in Season 1.5, on stocks, Grok 4.20 won while the older Grok 4 posted the lowest of 32 results.
  • Three academic benchmarks (StockBench, AI-Trader, FINSABER) found that general intelligence doesn’t reliably translate into trading skill, and FINSABER found that edges measured in short tests shrink over longer runs.
  • Pick by task, test the exact prompts you’ll use on your own charts, pin the model version you tested, and keep the trade decision in code rather than only in a prompt.
  • An AI trading agent such as Ethy isn’t a fifth model to rank: the chart read, the backtest and the order are already part of the product.

Questions people ask

Which LLM won nof1’s Alpha Arena?

It depends on the season. Season 1, on crypto, was won by Qwen3-Max, with DeepSeek V3.1 second; Claude, Grok, Gemini and GPT-5 all lost money. Season 1.5, on US stocks, was won by a “Mystery Model” that nof1’s leaderboard names as Grok 4.20, a newer model than the Grok 4 that lost in Season 1 and finished last in one of Season 1.5’s four competitions.

Is there one reliable benchmark for the best trading LLM?

No. Alpha Arena ran for a few weeks on a handful of models; academic benchmarks such as StockBench, AI-Trader and FINSABER run longer or wider tests, and FINSABER found that an edge measured over a short run tends to shrink over a longer one. Treat any single benchmark, including the ones on this page, as a snapshot.

Is Claude or ChatGPT better for trading?

For most trading tasks they’re close: both read chart screenshots, search the web with citations and run code. Claude adds custom MCP connectors on every plan; ChatGPT adds investment data from accounts you link. The full task-by-task comparison is in Claude vs ChatGPT for trading.

What do Claude, ChatGPT and Gemini cost for a trader?

As of November 16, 2026, Claude Pro is $20 a month ($17 billed yearly) and ChatGPT Plus is $20 a month, both adding the fuller model and more usage; both companies’ pricing pages list higher tiers above that. Google’s Antigravity coding tool has a free plan on macOS, Windows and Linux, and paid Google AI plans raise its limits. Grounding with Google Search gives 5,000 free search requests a month, then bills $14 per 1,000. Check each company’s own pricing page before you commit, since tiers change often.

Can any of these models watch a position while I’m away?

Only through a scheduled job you build and connect, and even then with limits: Claude Code’s cloud scheduled tasks and ChatGPT’s scheduled tasks exist, but a prompt that runs later is not a resting stop order. A real stop belongs at the broker or exchange, not in a model’s memory.

Does a model that wins one season keep winning?

Not reliably. Nof1 itself warned before Season 1 even ended that “early successes may be the result of luck”, and a newer Grok model won Season 1.5 after an older Grok model lost Season 1. FINSABER’s backtests over two decades found the same pattern: reported LLM advantages shrink once the test runs longer and wider.

Sources

  1. Anthropic docs: Models overviewplatform.claude.com
  2. Anthropic docs: Vision (reading images, high-stakes caveat)platform.claude.com
  3. Anthropic docs: Agent Skills overviewplatform.claude.com
  4. Anthropic: Introducing Claude Fable 5.2 and Claude Mythos 5.2 (November 9, 2026)anthropic.com
  5. Claude: Plans and pricingclaude.com
  6. OpenAI API docs: Images and vision (spatial localization limits)developers.openai.com
  7. ChatGPT docs: Codex surfaces, MCP and approvals (read November 16, 2026)learn.chatgpt.com
  8. ChatGPT: Plans and pricingchatgpt.com
  9. xAI docs: Modelsdocs.x.ai
  10. xAI docs: Grok Build 0.1docs.x.ai
  11. xAI docs: Grok Build, the coding agentdocs.x.ai
  12. xAI docs: Grok 4.7 (image input; the default model in Grok Build)docs.x.ai
  13. xAI docs: Release notes (Grok 4.20, Grok 4.7)docs.x.ai
  14. Google AI for Developers: Gemini modelsaistudio.google.com
  15. Google AI for Developers: Grounding with Google Searchaistudio.google.com
  16. Gemini: Deep Research overviewgemini.google
  17. Google Antigravity: overviewantigravity.google
  18. Google Antigravity: plans and pricingantigravity.google
  19. Nof1: home page (read November 15, 2026)nof1.ai
  20. Nof1: Alpha Arena home page (Season 1.5 result, read November 15, 2026)alphaarena.ai
  21. Nof1: Alpha Arena leaderboard (Season 1.5 winner and account values, read November 15, 2026)alphaarena.ai
  22. Nof1: Exploring the Limits of Large Language Models as Quant Traders (October 27, 2025; Wayback Machine copy of November 12, 2025)web.archive.org
  23. Nof1: Alpha Arena Season 1 account totals at the close (Wayback Machine copy of November 4, 2025)web.archive.org
  24. Chen et al., StockBench: Can LLM Agents Trade Stocks Profitably In Real-world Markets? (arXiv, posted October 2, 2025, revised March 2, 2026)arxiv.org
  25. Fan et al., AI-Trader: Benchmarking Autonomous Agents in Real-Time Financial Markets (arXiv, posted December 1, 2025)arxiv.org
  26. Li et al., Can LLM-based Financial Investing Strategies Outperform the Market in Long Run? (FINSABER, arXiv, posted May 11, 2025, revised June 26, 2026)arxiv.org

Trading involves risk, including the loss of your money. Ethy AI is a research and trading tool and does not give investment advice or recommendations. Backtests are simulations on historical data and do not predict future results.

Ethy blog

Read more

The $ETHY ecosystem

Stop watching charts.
Start living.

Deploy your AI trading agent in one click.