Models, Research & Prompt ·
GPT 5.6 Is Here, but the Hype Needs a Reality Check
Today, the most explosive thing for the technology circle is the three arrows of GPT 5.6, that is, Sol, Terra, and Luna are online at the same time.
Date: 2026-07-10
Today, the most explosive thing for the technology circle is the three arrows of GPT-5.6, that is, Sol, Terra, and Luna are online at the same time.
After the release is completed, all subscriptions will have their quota reset. This is the most important version in the second half of this year, and it is also the last product before the GPT 6.x era. I estimate that 6.x will be online soon.
Regarding the capability description of GPT5.6, I posted it the day before yesterday. You can read my previous article.
You can just update the Codex version directly.
Sol is a lot stronger and beats Fable 5 directly.
But GPT-5.6-Sol is only available with Pro 5x and 20x, and Plus versions are only available with Terra and Luna.
But I don’t know why. It was mentioned before that Plus can use Sol. I don’t know why it has not been grayed out this time. It is different from what was said when it was released. If you have anything you can use, please tell me.
In addition, Sol has also added a new Max and Ultra thinking quota. The reasoning capabilities of Max and Ultra are both very good. Max allows the model to spend more time thinking deeply, while ultra will call multiple subagents to process tasks in parallel. Moreover, Ultra consumes a lot of tokens and needs to be used with caution.
The official also released a review. It seems that this GPT-5.6 Sol not only does the job well, but also saves tokens and is Fable 5. . . Forget it, too lazy to say it.
And the official also specifically emphasized thatLuna is a version that Sol uses for post-training.
Although GPT 5.6 has been officially released, Codex has come to say goodbye.
OpenAI officially launchedChatGPT Codex and ChatGPT Work.
The current Codex is calledChatGPT Codex.
These two modes can be seamlessly switched directly in Codex.
The launch of ChatGPT Work, I feel, may be an entry point for OpenAI to enter the office scene.
Unlike Codex For developers, which shows the thinking process, ChatGPT Work hides the technical details and is mainly aimed at non-developers.
ChatGPT Work is the most noteworthy new feature this time. It no longer just answers questions, it can perform tasks across applications: connect your Google Drive, Slack, email and other tools, break down complex projects yourself, and complete them step by step independently, asking you only when needed.
For example, if you give it a sentence to "organize customer research into a marketing plan," it can read the information, write plans, and generate versions of materials for different markets on its own, with context throughout the process.
This is also what I said before, ChatGPT and Codex may be integrated together.
Work will also be available on the web and mobile phones, which means you can initiate and track tasks without being in front of a computer.
Many domestic products such as Qoder, Trae, and WorkBuddy actually have two sets, one for programmers and one for professionals.
It's just that this time OpenAI integrates them together.
This time the browser also received a major update.
Using Computer use within Codex is smoother and supports certified websites, multiple tabs, and file downloads. In addition, your labels can also be retained, making it smoother.
This Computer use function allows your ChatGPT to use any app more and more like a human being. The most important thing is that it is faster to use and saves tokens.
I used the previous Computer use function very frequently on a daily basis, and I did feel that the efficiency was a bit slow.
In addition, GPT 5.6 also has complete improvements in front-end, visualization and other aspects. OpenAI also released a front-end rendering of different types of generation this time.
There are so many small improvements.
This update still brings a lot of new things to everyone. I think these small details and features can keep me excited for a while.
By the way, after the official release of GPT 5.6, Company A chose to reset the limits for all users.
GPT 5.6 has three arrows fired at the same time. Of course, everyone will fall into a comfortable atmosphere and may not be able to hear some objections.
But I may have to tell you something different, and I may have to pour some cold water on it.
Let me start with a picture of my favorite evaluation agency.
What this picture actually says is that the outcome of Fable 5 and GPT-5.6 should be viewed in terms of comprehensive capabilities and coding agents.
- The top half of the graph is the comprehensive Intelligence Index (mix of 9 different tests): Fable 5 60, GPT‑5.6 Sol 59, Terra 55. In other words, in terms of overall score, Fable is slightly higher than Sol, and Terra is one level lower.
- The lower half of the picture is the Coding Agent Index (the sum of DeepSWE, Terminal-Bench, SWE-Atlas-QnA): Sol 80，Terra 77，Fable 5 77。 In other words, Sol comes first when it comes to code agents; Terra and Fable are basically tied.
Therefore, it cannot be concluded that GPT‑5.6 fully defeats Fable, but it shows:
Sol is the number one code agent; Fable is still slightly ahead in overall capabilities; Terra is positioned very much like a coding agent, a mid-range model with scores close to Fable, but with lower overall capabilities.
In addition, max is the highest inference level, and with fallback means that Fable's safe fallback configuration is also running; so this is not a pure model naked duel. Different tool chains, reasoning levels and evaluation question sets will change the ranking.
Simon Willison mentioned on his blog, GPT 5.6. . . Maybe the effect is not as strong as Fable 5.
Simon Willison was fortunate enough to be one of the people who could experience it in advance, so he had enough testing time to explain clearly the causes and consequences.
He said that they designed a “ultimate agent test” to make every token play a greater role.
In this test, which covers 55 long-running professional workflow evaluations, GPT-5.6 Sol achieved a new high score of 53.6 points, 13.1 points higher than Claude Fable 5 (with adaptive inference mode).
Even in medium mode, it also leads Fable 5 by 11.4 points at about a quarter of the estimated cost.
This is critical to enabling more pervasive, affordable intelligence: GPT-5.6 Terra and GPT-5.6 Luna both outperform Fable 5 at about one-sixteenth the cost.
However, in the field of SWE-Bench that everyone is paying attention to, Fable 5 significantly outperforms GPT-5.6. Fable 5 achieved a score of 80%, while GPT-5.6 Sol was 64.6%.
Of course, OpenAI published a special article the next day, saying that about 30% of the questions in SWE-Bench Pro have problems themselves. There may be incomplete descriptions of questions, excessive testing, and non-unique answers. This report card cannot be regarded as the final judge.
He said that in his current use, the A company model does not perform better than Fable in terms of complex coding tasks.
He demonstrated 18 different "pelican" configurations for different versions of GPT 5.6 - covering the performance of three different models at different inference strengths. Their Token consumption and calculation costs are also listed: the lowest cost is gpt-5.6-luna (inference strength is None), and the cost is 0.71 cents; the highest cost is gpt-5.6-sol (inference strength is Max), and the cost is 48.55 cents.
*Image credit: Simon Willison, three columns Luna, Terra, Sol, six rows none to max. *
I tested a few rounds myself. I didn’t use Sol. I tested the Terra ultra mode. It gave me the feeling that it was not as good as GPT 5.5 and the output speed was very slow. Moreover, I had two rounds of problems when I just started using it.
In the first round, an empty final message is sent at the end of the task, so the interface appears blank.
Then in the second round, English was mixed with Chinese and output.
So who is stronger? I don’t feel the need to jump to conclusions so quickly.
Maybe I didn't use the Sol mode, so I might not be able to measure anything with Terra.
GPT-5.6 Sol may have more advantages in token efficiency, multi-agent collaboration and tool invocation; Fable 5 is still in an irreplaceable position, at least in the complex engineering tasks Simon actually does.
I tested Grok 4.5 yesterday and it feels good. If the effect of GPT 5.6 - Sol is not good after I use it for testing, I may use Grok 4.5 directly.
So what exactly? I’m just waiting for the actual test results from all of you who are smart here.
Finally, let me digress. Many reviewers probably stayed up all night waiting for the release to finish writing the article and then went to bed, which is quite good.
But I got up at 4 o'clock to watch the game between France and Morocco, and then wrote the article.
Should I go to bed then? No, I have to go to work :)