, I’ll give my first impressions of the most recent OpenAI mannequin, GPT-5.6. The mannequin was launched just a few days in the past, and I’ve gotten the possibility to check it extensively for the reason that launch and evaluate it towards different fashions comparable to GPT-5.5, Opus 4.8, and Fable 5.
I’ll give my judgment on the mannequin, share my experiences with it, and clarify work as successfully as attainable with it. I consider it’s a mannequin with some professionals and a few cons in comparison with the choice Anthropic fashions comparable to Opus 4.8 and Fable 5. However general, it’s an excellent mannequin, and I positively suggest making an attempt it out.
Why use GPT-5.6?
To start with, I’d prefer to cowl why you need to care about this text, and on this case, it’s why you need to care about GPT-5.6 and use it successfully. To start with, the rationale you need to care about GPT-5.6 is that the earlier technology of the identical mannequin, GPT-5.5, was a terrific mannequin and a mannequin I personally used extensively.
I consider GPT-5.5 was mainly on par with Opus 4.8 in lots of duties, and in some duties, comparable to code evaluation, it was far superior in my view.
Thus, you need to in fact care in regards to the subsequent technology of the identical mannequin, GPT-5.6, because it ought to, on paper, be higher. Moreover, GPT-5.6 has some quirks which you need to positively learn about. For instance, the mannequin is available in three completely different sizes:
The place Sol refers back to the Solar, Terra refers back to the Earth, and Luna refers back to the Moon, representing the sizes of the fashions. Sol is the most effective one and the frontier mannequin. In addition they, with every mannequin, launched completely different reasoning ranges, which might make the mannequin assume for longer or shorter earlier than offering solutions. The straightforward trade-off is that should you make the mannequin motive for longer, it is going to give higher-quality responses, however it is going to take longer to offer responses.
I’ll share my strategies and ideas on use the fashions as successfully as attainable.
My ideas on GPT-5.6
First, let me cowl my general ideas on GPT-5.6. General, I’d say it’s an enchancment over GPT-5.5 in mainly each facet. It’s only a bit higher than GPT-5.5 for every part.
I in fact use the mannequin for code opinions, wherein it nonetheless discovers lots of points. I consider it’s barely higher at catching points within the code, each with regard to precision and recall, i.e., catching all of the bugs that exist, which is recall, and truly being right when reporting a bug, which is precision.
I’ve additionally tried the mannequin for precise implementations, and I consider GPT-5.6 is ready to work for longer to finish duties and is mainly extra thorough in the best way it does its work. I nonetheless consider GPT-5.5 was fairly good at getting duties carried out, however I believe GPT-5.6 is barely higher.
Nevertheless, it’s price noting that I don’t assume this can be a very robust enchancment; I consider it’s an incremental enchancment over the earlier GPT mannequin.
One draw back I wish to observe is that should you use GPT-5.6 with further excessive and even extremely considering, it mainly drains your utilization limits straight away.
Now it’s price noting that OpenAI truly eliminated the five-hour utilization restrict, not less than for a short lived period of time, which positively helps lots right here. So that you solely have the weekly restrict; nevertheless, I consider that should you use the mannequin with further excessive or ultra-thinking, it mainly drains your utilization straight away, and should you’re on a subscription, it’s very exhausting to make use of the mannequin for an prolonged time period, and positively very exhausting to make use of a number of fashions in parallel.
Moreover, I believe that should you use the mannequin with considered one of these excessive reasoning modes, it’s additionally very, very gradual. Slower than I’d count on it to be, positively when engaged on easier duties. So, I’ve truly ended up utilizing further excessive considering for planning and medium reasoning for precise implementations, which I’ll cowl extra within the subsequent part. However I believe it’s price noting that I needed to tune down the reasoning ranges to keep away from hitting price limits straight away.
That is, in fact, essential as a result of if you take a look at the benchmarks, what’s reported is normally one of many tremendous reasoning ranges, and should you’re not ready to make use of that as a result of it drains your utilization straight away, then that in fact makes the mannequin successfully worse for you when making an attempt it in comparison with the benchmarks.
I’d additionally prefer to have a brief part masking my ideas on the completely different sizes that you could select. I principally use Sol, the biggest mannequin, as I consider it’s the most effective mannequin. Nevertheless, I’ve learn some benchmarks saying that in some conditions you’re higher off utilizing Terra with a better reasoning degree than Sol with the decrease reasoning ranges. I examined this a bit myself and wasn’t actually capable of discover any stark variations, so I ended up staying with Sol with the reasoning efforts that I discussed earlier.
How you can successfully apply GPT-5.6 to unravel issues
Use instances for GPT-5.6
Now I’ll get into successfully use GPT-5.6 to unravel issues. The primary factor you need to begin doing straight away, particularly should you’re utilizing Claude Code, is to have GPT carry out your code opinions.
For my part, you principally don’t want human code opinions anymore. In fact, you would possibly want it if there’s a essential piece of infrastructure otherwise you actually desire a human to take a look at it, however for probably the most half, I consider that Codex is nice sufficient as a code reviewer to keep away from bugs being pushed to manufacturing.
Now you may also, in fact, use GPT-5.6 for precise implementations. Nevertheless, in my view, I’ve higher success utilizing the next setup for implementations. The first step is that I take advantage of Claude Fable to plan the implementation, and step two is that I change and begin executing the implementation with Claude Opus 4.8 as a substitute. For my part, this offers me higher outcomes than simply utilizing GPT-5.6, even when I take advantage of a better reasoning degree for planning after which a decrease reasoning degree for implementations.
And one other level you would possibly use GPT-5.6 for is laptop use or browser use. For my part, GPT is excellent at laptop use and browser use, and it’s fairly quick when interacting with the browser, particularly if I take advantage of a medium reasoning degree. I believe GPT-5.6 navigates my browser extremely properly, which is, in fact, essential when verifying code end-to-end, for instance, or performing actions within the browser for you. So I consider this can be a excellent use case.
Methods to make use of GPT-5.6 successfully
Now, I’ll transfer over to some particular strategies that I implement and apply when using GPT-5.6. The primary matter I need to cowl is the reasoning ranges. As I discussed earlier, should you go along with the additional excessive or ultra-reasoning ranges, I consider the mannequin is each too gradual and spends utilization limits means too quick, which doesn’t actually work should you’re on the subscription, even should you’re on the $200 subscription.
Thus, the approach that I began utilizing is that I’ve further excessive considering after I’m planning with GPT-5.6. So, all the time after I’m beginning a activity, I begin with plan mode, which has further excessive reasoning. When the plan is completed, I change to a medium reasoning degree to really implement the code, contemplating code implementation is usually simpler than code planning. It is because code planning requires you to look into the complete repository and take into account how one thing needs to be applied, whereas the precise implementation is simply implementing the plan that was already made for you.
One other crucial approach when utilizing GPT-5.6 is giving it entry to every part. One factor that I observed that was fairly inconvenient was that I’ve labored principally with Claude Code earlier than and had given Claude entry to every part I’m utilizing via MCP, comparable to Gmail, Google Calendar, Slack, Playwright MCP, and so forth. After which after I began utilizing GPT-5.6, it carried out worse as a result of I didn’t bear in mind to offer it entry to all of those instruments. So my different major tip right here is to really present GPT-5.6 entry to all of the instruments that it wants.
OpenAI has mainly all the identical connectors that Claude Code has, and there’s no motive you shouldn’t give entry to Codex should you already give entry to Claude Code.
And lastly, the final tip I need to cowl is to do not forget that OpenAI usually offers out resets, which is among the massive variations between Claude Code and Codex. So, with Claude Code, they generally do resets of utilization limits for everybody, which is, in fact, nice should you’ve spent a few of your limits; nevertheless, Codex additionally does that, however they normally truly present you a banked reset, which is mainly a reset that you could set off at any cut-off date and can reset your utilization limits.
This may be nice in case you have excessive anticipated utilization for a shorter time, or should you simply spent all of your tokens and must spend extra tokens. Nevertheless, it’s price noting that resetting the utilization limits not solely, in fact, units your utilization to 0%, nevertheless it additionally resets the date of when the subsequent utilization reset happens. So your subsequent five-hour restrict can be 5 hours later, and your subsequent week restrict can be one week later. So in addition they reset, which reduces the good thing about the banked resets, although they’re nonetheless very beneficial, in fact.
Traditionally, OpenAI has offered banked resets from time to time, so in case you have a subscription, you’ll get them over time.
Conclusion
On this article, I lined my opinion on the most recent OpenAI mannequin, GPT-5.6, and my ideas on it. I mentioned why you need to use the mannequin, highlighting that the earlier technology of the OpenAI mannequin, GPT-5.5, was already excellent, and that that is an enchancment over the mannequin. I additionally shared some particular strategies on use the mannequin successfully, which is essential as a result of GPT-5.6 comes with some completely different utilization settings. For instance, it is available in three completely different sizes of fashions, and so they even have completely different reasoning ranges you may select between. I consider you ought to be utilizing GPT-5.6 for code opinions, positively. And the distinction between GPT-5.6 and Opus 4.8 shouldn’t be that massive in the case of code implementations, laptop use, and so forth. So my present coding setup will nonetheless stay: Claude Fable for planning, Opus 4.8 for execution, and GPT-5.6 for reviewing my code. In fact, you need to all the time keep on high of the most recent fashions, attempt them out your self, see in the event that they work higher on your use instances, and thus determine if it’s related to you.
👉 My free eBook and Webinar:
🚀 10x Your Engineering with LLMs (Free 3-Day E-mail Course)
📚 Get my free Imaginative and prescient Language Fashions e book
💻 My webinar on Imaginative and prescient Language Fashions
👉 Discover me on socials:
💌 Substack
