GLM-5.3 vs Opus 5

I used GLM-5.3 and Opus 5 to generate trading strategies inside a trading harness. Both received the same prompt and both had access to the same trading knowledge hub.

This experiment interested me because Opus 5 is one of the most expensive models on the market, whilst GLM-5.3 sits at the cheaper end of the high-performing ones.

Opus 5 costs roughly $25 per million tokens, whilst GLM-5.3 charges $4.40. That makes Opus 5.6 times as expensive.

Does that mean GLM-5.3 delivers performance that is 5.6 times worse? Let us find out.

The full comparison at a glance — GLM-5.3 against Opus 5.

What Is a Trading Harness?

It is the environment in which AI models such as Opus 5 or GLM-5.3 are executed and fed with feedback on the performance of the trading strategy they generated.

Months ago I started thinking about how to help the models and steer them in the right direction, so that they build strategies that actually hold up. In this article I use the system I built around the models. It does three things:

The harness: the knowledge hub goes in at the front, and two feedback loops run back into the model before the optimiser ever sees the strategy.

The Prompt

GLM-5.3 and Opus 5 both received exactly this prompt:

Prompt

Build me a trend reversal strategy with a maximum of 1 open position at a time. Research the knowledge hub and propose the most suitable entry signal. Ask no questions. Simply implement it so the code is compliant and can run. Use parameters so I can optimize the strategy.

This prompt is far from ideal, because it tells the model not to ask any questions. To get the best out of either model you need to steer the idea as it takes shape. For this experiment, though, I deliberately removed my own intervention so that we can see the raw performance of the models.

The First Draft

There is a great deal of talk on the web about asking an AI to build a profitable strategy. In my experience it does not work straight away. The model knows nothing about your market, your timeframes or what you consider a good strategy. That is why the first draft it produces does not work — at least not at the start.

Opus 5 produced this on the first attempt:

Opus 5 — the first draft, before any optimisation.

GLM-5.3 delivered this:

GLM-5.3 — the first draft, before any optimisation.

Both variants are useless. Let us see whether the harness can improve them through analysis and optimisation.

Optimisation

This is the crucial step. In the prompt I asked the model to use parameters so that the strategy could be optimised, and that matters more than it might appear. The models genuinely have no feel for trading or for the way market conditions need to be handled. So we keep the options open and search for a configuration that works well on gold, XAUUSD.

The first observation is that the optimised metric and the profit factors are twice as high for Opus 5 as for GLM-5.3. In other words, the ratio between the total gain and the maximum drawdown over the two-year backtest is twice as good with Opus. It therefore appears to have found a better way to keep losses small and let the winners grow.

Note that both strategies settled on a low win rate and relatively few trades: 45 for Opus 5 and 35 for GLM-5.3. The charts that follow explore this from several angles.

Performance Comparison

Equity Curve

Both equity curves have a similar shape, though GLM-5.3 ends with the smaller gain.

Statistics

Opus 5 finished +$13,048.93, a gain of 123.78%, against a maximum drawdown of $672.30 (6.38%). It took 45 trades in total, and only 6 of them were winners. That is a win rate of 13%, yet it still produced a recovery factor of 19.41. The longest losing streak ran to 16 trades and the longest position stayed open for 112.4 days.

GLM-5.3 finished +$6,011.55, a gain of 60.21%, against a maximum drawdown of $765.47 (4.87%). It took 35 trades in total, of which 8 were winners. That is a win rate of 23% and a recovery factor of 7.85. The longest losing streak ran to 9 trades and the longest position stayed open for 60.2 days.

The two models arrive at their profit in much the same way. GLM-5.3 wins more often and sits through a shallower drawdown, but it closes its positions early and the winners never grow large. Opus 5 is wrong on almost every trade it takes and endures a losing streak of 16, yet it more than doubles the account, because the few trades it gets right are allowed to run for months. A recovery factor of 19.41 against 7.85 is a striking difference.

The strategy Opus 5 implemented would therefore probably do well if the gold price resumes its upward move.

PnL Histogram

The x axis is closed PnL and the y axis is the number of trades.

Opus 5: the losing side runs from roughly -$165 up to zero and is not evenly spread. 24 of the 39 losses fall into the two bins closest to zero, around -$41 and -$20, and only one trade loses more than about -$150. The winning side occupies four bins: one trade near +$440, one near +$1,320, one near +$2,190, and three in the top bin at roughly +$3,510. The bars add up to 39 losses and 6 wins, which matches the stats panel.

GLM-5.3: the losing side occupies only two bins, about 22 trades around -$114 and about 5 trades around -$201. A spread that narrow is what a fixed stop-loss produces, and the second bin would then hold fills that slipped past it, though the histogram on its own cannot confirm that. The eight wins are spread thinly from about +$200 to about +$2,810, with no bin holding more than two of them. The bars again reconcile with the stats panel: 27 losses and 8 wins.

The difference between the models is one of degree. Opus 5 kept its typical loss smaller and held its winners longer, which is what produces the higher recovery factor of 19.41 against 7.85.

Whether that difference holds on a different instrument or a different period is something anyone can find out on AI-Backbone.com.

Holding Time

Both scatter plots show the same structure. Every loss closes within days, every win is held for weeks, and the band in between is empty.

• Opus 5: losses close inside about 8 days; all 6 wins sit between 90 and 113 days.
• GLM-5.3: losses close inside about 2 days, apart from three between 7 and 28 days; all 8 wins sit between 55 and 61 days.

The split comes from the exit logic rather than from a better entry signal.

Summary

Both models were given the same task: design and build a trend reversal strategy. Opus 5 managed to keep its losses small in relation to its wins. The best strategy from GLM-5.3 ended up with a moderate drawdown and a higher win rate, but a smaller overall profit.

Two things stand out.

On the first attempt, neither model produced anything usable. Both first drafts were worthless. Everything that makes these strategies work came out of the optimisation loop, not out of the model's initial idea. The model picks the design of the strategy, and the harness then finds the parameters.

Nobody told either model to trade rarely, to cut losses within days and to hold winners for weeks. The prompt asked for a trend reversal strategy and nothing else. Both still arrived at the same archetype: a low win rate, fewer than fifty trades in two years, losses capped tightly, and roughly 70% of the gross profit coming from three trades. Two models from different vendors converging on the same structure suggests that the harness and the knowledge hub shape the result at least as much as the model does. It is also the pattern that experienced traders take years to learn, because sitting through sixteen losses in a row to keep one runner open is not something most people manage.

How to Choose

• If you want peak performance straight away, Opus 5 is the right choice.
• If you want to generate several strategies on the same budget, GLM-5.3 will serve you better. It also looks like a good way to explore how the AI-Backbone trading harness works.

If you want to build a strategy like this one, join the waiting list.

If you want to build a strategy like this one, join the waiting list and get ahead of the other traders.