I’ll give my first impressions of OpenAI’s newest mannequin GPT-5.6. This mannequin was launched a couple of days in the past, however since its launch I’ve had the chance to check it extensively and evaluate it to different fashions similar to GPT-5.5, Opus 4.8, and Fable 5.
I’ll go my judgment on this mannequin, share my expertise with it and clarify the best way to use it as successfully as doable. I believe this mannequin has its professionals and cons when in comparison with different Anthropic fashions similar to Opus 4.8 and Fable 5. However general it is an excellent mannequin and I extremely suggest giving it a strive.
Why use GPT-5.6?
To begin with, I wish to clarify why it is best to take note of this text: why it is best to take note of GPT-5.6 and the best way to use it successfully. To begin with, the explanation why I selected GPT-5.6 is as a result of the earlier technology of the identical mannequin, GPT-5.5, is an excellent mannequin and one which I personally use.
GPT-5.5 was mainly equal to Opus 4.8 for a lot of duties, and significantly better for my part for some duties similar to code opinions.
So, in concept, the following technology of the identical mannequin, GPT-5.6, needs to be higher, so in fact it is best to care. Moreover, GPT-5.6 has some quirks that it is best to undoubtedly concentrate on. For instance, the mannequin is available in three completely different sizes.
Sol refers back to the solar, Terra refers back to the earth, Luna refers back to the moon, and represents the scale of the mannequin. The Sol is one of the best of one of the best and is a frontier mannequin. We additionally launched completely different ranges of inference for every mannequin, permitting the mannequin to assume longer or shorter earlier than offering a solution. The easy trade-off is that the longer the mannequin infers, the upper high quality the response, however the longer it takes to offer the response.
I am going to share my methods and ideas on the best way to use fashions as successfully as doable.
My ideas on GPT-5.6
First, let me clarify my general ideas on GPT-5.6. Total, it is an enchancment over GPT-5.5 in mainly each means. Barely higher than GPT-5.5 in all the things.
In fact, I exploit this mannequin for code opinions, however I nonetheless discover many issues. I believe it is barely higher at catching issues in your code, each when it comes to precision and recall. This implies catching all of the bugs that exist (recall) and being really correct when reporting bugs (precision).
I’ve additionally tried fashions for precise implementations, and I believe GPT-5.6 permits you to work longer to finish a process, and is mostly extra thorough in its method. I nonetheless imagine that GPT-5.5 was fairly good at finishing the duty, however I believe GPT-5.6 is barely higher.
Nevertheless, it is price noting that I do not assume this can be a very sturdy enchancment. I believe it is an incremental enchancment over the earlier GPT mannequin.
One of many downsides you need to pay attention to is that when you use GPT-5.6 in very superior or tremendous superior considering, you will primarily run out of utilization limits shortly.
It is price noting that OpenAI has really eliminated the 5-hour utilization restrict, not less than quickly. That is undoubtedly very useful right here. So it is solely a weekly restrict. Nevertheless, I believe when you use an ultra-high efficiency or ultra-high efficiency mannequin, you’ll mainly run out of utilization shortly. Additionally, in case you have a subscription, it’s extremely tough to make use of that mannequin for lengthy durations of time, and it is undoubtedly very tough to make use of a number of fashions in parallel.
Moreover, I believe utilizing a mannequin with any of those superior inference modes can be very sluggish. When engaged on easy duties, it is slower than anticipated. So, in observe, we ended up utilizing very high-level considering for planning and medium-level reasoning for precise implementation. This can be defined intimately within the subsequent part. Nevertheless, I believe it is price noting that I needed to modify the inference stage to keep away from hitting the speed restrict too shortly.
That is in fact essential. As a result of whenever you take a look at benchmarks, what’s normally reported is without doubt one of the hyper-inference ranges. And if you cannot use it as a result of it shortly runs out of utilization, then in fact you will be at an actual drawback when making an attempt the mannequin in opposition to a benchmark.
I’d additionally prefer to have a brief part explaining my ideas on the completely different sizes you possibly can select from. I primarily use the most important mannequin, Sol, as I imagine it’s the greatest mannequin. Nevertheless, I’ve learn some benchmarks that say that in some conditions it is higher to make use of Terra, which has a better stage of reasoning, than Sol, which has a decrease stage of reasoning. I additionally examined this a bit and could not discover any apparent distinction. So I continued with the reasoning efforts I discussed earlier and ended up sticking with Sol.
How one can successfully apply GPT-5.6 to resolve issues
Examples of utilizing GPT-5.6
Now let’s talk about the best way to successfully use GPT-5.6 to resolve your issues. The primary factor to start out doing straight away, particularly when you’re utilizing Claude code, is to have GPT carry out code opinions for you.
In my view, there’s no use for human code evaluation anymore. In fact, you would possibly want Codex in case you have essential items of infrastructure or actually desire a human to take a look at it, however most often I believe Codex is nice sufficient as a code reviewer to keep away from pushing bugs into manufacturing.
In fact, you may as well use GPT-5.6 in your precise implementation. Nevertheless, for my part, the next configuration works higher for implementation: Step 1 is to plan your implementation utilizing Claude Fable. In step 2, we’ll as an alternative change to Claude Opus 4.8 and begin operating the implementation. In my view, utilizing a better inference stage for planning and a decrease inference stage for implementation nonetheless yields higher outcomes than merely utilizing GPT-5.6.
One other level to make use of GPT-5.6 is utilizing your pc or browser. In my view, GPT is superb for pc use and browser use, and may be very quick when working with browsers, particularly when utilizing medium reasoning ranges. I believe GPT-5.6 navigates browsers very nicely. In fact, that is essential if you wish to validate your code end-to-end, carry out actions within the browser, and so forth. So I believe this can be a superb use case.
Strategies for successfully using GPT-5.6
We’ll now transfer on to some particular methods to implement and apply when leveraging GPT-5.6. The primary matter I need to tackle is the inference stage. As talked about earlier than, when you use the Tremendous Excessive or Tremendous Reasoning ranges, this mannequin is simply too sluggish and consumes your spending restrict too shortly, so it does not actually work when you’re on a subscription, even when you’re on a $200 subscription.
So a method I’ve began utilizing is additional high-level considering when planning in GPT-5.6. So everytime you begin a process, you begin in a planning mode with very superior reasoning. As soon as the planning is full, change to the medium stage of reasoning and truly implement the code. We frequently assume that implementing the code is less complicated than planning the code. It is because code planning requires you to undergo your entire repository and think about the best way to implement one thing, whereas in precise implementation you’re merely implementing a plan that has already been created.
One other crucial method when utilizing GPT-5.6 is to verify all the things is accessible. One of many issues that I discovered very inconvenient was that up till now I had primarily been working with Claude Code, giving Claude entry to all the things I exploit via MCP, together with Gmail, Google Calendar, Slack, and Playwright MCP. Then, after I began utilizing GPT-5.6, I did not bear in mind to offer entry to all these instruments, which triggered poor efficiency. So my different important tip right here is to really present GPT-5.6 entry to all of the instruments you want.
OpenAI has primarily all the identical connectors as Claude Code, and when you’ve already granted entry to Claude Code, there is no purpose to not grant entry to Codex.
And eventually, the final tip I need to carry up is keep in mind that OpenAI resets regularly. This is without doubt one of the main variations between Claude Code and Codex. Subsequently, Claude Code could reset utilization limits for everybody. That is in fact good in case you have exhausted a few of your limits. Nevertheless, Codex additionally does this, however normally really supplies a financial institution reset. That is primarily a reset that may be triggered at any time and resets your utilization limits.
That is superb when you count on excessive utilization in a brief time period, or in case you have used up all of your tokens and have to spend extra. Nevertheless, needless to say resetting your utilization restrict not solely units your utilization to 0%, but in addition resets the date on which your utilization will subsequent be reset. So the following 5 hour restrict can be in 5 hours and the following week’s restrict can be in 1 week. Subsequently, the advantages of banked resets are lowered as they’re additionally reset, however in fact nonetheless very priceless.
Traditionally, OpenAI has supplied financial institution resets now and again, so in case you have a subscription it is possible for you to to get them over time.
conclusion
On this article, I coated my opinions and ideas about OpenAI’s newest mannequin, GPT-5.6. I defined why we must always use this mannequin and emphasised that the earlier technology OpenAI mannequin, GPT-5.5, was already superb and that is an enchancment on that mannequin. We additionally shared some particular methods on the best way to use the mannequin successfully. This is essential as a result of GPT-5.6 comes with a number of completely different utilization configurations. For instance, there are three completely different sizes of fashions and completely different ranges of inference to select from. I believe it is best to undoubtedly use GPT-5.6 for code opinions. With regards to code implementation, pc utilization, and so forth., the variations between GPT-5.6 and Opus 4.8 are usually not that huge. Subsequently, your present coding settings will proceed to stay. We use Claude Fable for planning, Opus 4.8 for execution, and GPT-5.6 for code opinions. In fact, it is best to at all times keep updated with the most recent fashions, strive them your self, see in the event that they fit your use case, and resolve if they’re best for you.
👋 contact
👉 My free e-books and webinars:
🚀 10x your engineering with an LLM (free 3-day email course)
📚 Get your free Vision Language Models eBook
💻 My webinar on visual language models
👉 Discover me on social:
💌 substack
