I Tested Every Major LLM for Algorithmic Trading. There is One Clear Winner

1. I created a sample of trading strategies that the NexusTrade platform can generate. This includes strategies such as “Create a strategy that rebalances the Magnificent 7”, “Create a strategy that buys and holds this list of stocks”, or “Create a simple moving average crossover strategy”
2. I took a dozen of the best AI models and had them generate the trading strategies
3. I took Gemini 2.5 Pro and GPT-5 and evaluated the trading strategies using the above system prompt
4. I generated summary statistics and sorted the models by their medium score

After running the script, I ranked the tested models by their judges’ scores for matching requested rules. Illustrative semantic check: if the request requires both a cooldown and a low RSI, an OR configuration can buy when only one is true; an AND configuration requires both. Both can be valid JSON, but they express different strategies. This example explains the grading question; it is not an identified scored fixture from this run.

Some of the things I saw shocked me.

Opus 4.1 Comes Out as King of Algorithmic Trading

Press enter or click to view image in full size

According to this experiment, Claude Opus 4.1 is the best at understanding how to create algorithmic trading strategies. It achieved the highest median score (1/1), the highest average score (0.95/1), and an extremely high amount of perfect scores (72%). Even Claude Opus 4, which was released 2.5 months ago, outperforms the rest of the models on this list. Unlike other models, Opus seems to truly understand the nuances of creating algorithmic trading strategies.

Not only is Opus 4.1 the best, but it’s also the fastest.

But it comes with a cost.

The Opus models are between 5 to 10 times more expensive than even the second most expensive model on the list (GPT-5 and Gemini 2.5 Pro). While you are getting the best results, it doesn’t come cheap.

After the Opus series, we have GPT-5 and Gemini 2.5 Pro. Unsurprisingly, these models are also extremely good at creating algorithmic trading strategies. GPT-5 was significantly slower, but they both scored around the same for median and average score, with Gemini 2.5 Pro being slightly better.

Next comes GPT-5-mini, which actually surprised me. GPT-5 mini is one of the cheapest models on the list, costing less than Gemini 2.5 Flash and GPT 4.1, but performing much better. It even outperforms models like Grok 4, Claude Sonnet 4, and OpenAI o3, which are significantly more expensive. This is the outcome that shocked me the most.

Knowing that Opus 4.1 is the best model for algorithmic trading, I knew I had to do something with these insights.

I had to make it available for everybody.

Updating NexusTrade With The BEST AI Model

Claude Opus 4.1 led this August 2025 configuration-generation test. I added that version to the chat model list described below so users could try translating their own rules.

I had to make it accessible.

To do this, I updated NexusTrade and added a new model to the AI Chat.

When you click the Settings icon in the top right corner, a new model appears in the dropdown box ready to use.

Press enter or click to view image in full size

If you’re serious about learning how to be an algorithmic trader and you want the very best tools at your disposal, now is the perfect chance.

You now have access to a free platform to create, optimize, and deploy your own algorithmic trading strategies. You don’t have to be a cracked out software expert or a Wharton Finance graduate.

You just have to explain your ideas to the world’s most powerful AI model. How much easier could it be?

Concluding Thoughts

This exercise taught me a few valuable lessons.

For one, it reinforced the importance of benchmarking. While I’ve tested models for SQL Query Generation in the past, (and found that Opus 4 was severely disappointing for this use-case), I didn’t think about how vastly different these tasks are. I now know better.

GPT-5-mini performed better than several more expensive models in this configuration-generation test. Its separate SQL-query results concern another task; neither comparison measures all daily tasks.

Third, I learned that even expensive models can be lightning fast. At a whopping $15/M input tokens and $75/M output tokens, Opus was somehow able to outspeed even the smallest models on this list, while delivering on exceptional performance.

That’s insane.

Finally, this August 2025 test gave me a configuration-generation winner: Opus 4.1. The reported median was 1.0, mean 0.95 and perfect-score share 72%. Those are semantic judge scores, rather than returns from trading the generated population.

Translate a precise trading idea in Aurora. Inspect the returned entry, exit and purchase size, then correct any mismatch before a historical test and separate paper observation.