老叶
Feed
Feed
grok 4.6 is here too!
It was released almost simultaneously with DeepSeek-V4-Pro-0813!
Honestly, seeing the parameters of grok 4.6 really excites me! grok 4.5 was already one of my favorite daily models because it’s really fast and doesn’t have many restrictions, making it perfect for everyday use.
However, this time grok 4.6 has parameters close to the top two flagship models and even boasts faster speed than them.
Old Ma’s grok is really on the table this time! I think it can truly replace gemini as one of the new top three contenders!
Official announcement! DeepSeek-V4-Pro-0813 is coming soon!
Just now, the official DeepSeek website documentation has published the model ID for the official version of DeepSeek-V4-Pro. According to the date listed, the scheduled release date should be August 13.
Moreover, the V4-Pro official version now supports the Responses API, allowing direct use on codex without any conversion.
Exploded, OpenAI is being a bit ruthless this time.
The Pro x20 plan just had its quota reset, and I was dumbfounded when I saw it.
The weekly quota seems to be only about 1/3 of what it was before.
It's not a small cut, it's a direct cut to the bone.
In the past two days, OpenAI's model access has been overloaded, frequently lagging, rate-limiting, and queuing.
So now there are only two possibilities:
Either there's a bug in the quota configuration,
Or OpenAI really can't handle it anymore and has started actively tightening the quotas for Pro users.
If it's the latter, then the cost-effectiveness of the Pro x20 plan needs to be recalculated.
Not sure if this is an isolated case.
Brothers with x20, check your own quotas to see if yours have been cut too.
Some netizens asked why not just set directly: model_provider = "OpenAI"
This way, you can generate images and perform searches.
Let me explain why. Generally, proxies avoid model_providers.OpenAI mainly because this remote compact is a synchronous request, and many proxies using CF risk timing out after more than 100 seconds. If you are using your own reverse proxy and not CF, then you can definitely use this method, no need to read further.
After reverse engineering, I found that besides checking model_providers, it also checks the x-openai-actor-authorization header.
Therefore, according to my solution, you can bypass this remote compact and still achieve image generation and search.
Secondly, why reverse engineer? Because I found that the desktop version of codex checks the openai-actor-authorization header, and the desktop version is currently not open source.
Shocking! The new version of the Codex client is secretly "dumbing down" third-party relay stations.
On July 11, I discovered that the new Codex could not generate raw images and immediately published a complete solution.
I initially thought this was just a limitation on raw image generation, but after further reverse engineering, I found the situation was far more complex.
Once the new Codex detects that you are using a third-party Provider, it directly restricts two key capabilities:
1. Image generation
2. Web Search
The restriction on raw image generation is easy for everyone to notice.
But the silent disabling of Web Search is the real impact on the model's capabilities.
Many people may not realize how important Web Search is.
The internal knowledge base of large models has a cutoff date; it doesn't know what just happened today and cannot spontaneously access the latest documents, news, or technical changes.
Therefore, model providers equip the model with a real-time online search capability, which is Web Search.
It acts as the model's "eyes."
Without Web Search, the model can only rely on outdated knowledge to answer questions;
With Web Search, it can query information across the entire web in real time, verify facts, and read the latest documents.
But now, as long as you use a third-party relay station in the new Codex, even if the backend connects to the exact same official model, the client may directly block the Web Search capability.
The model hasn't changed.
The API hasn't changed.
It's just that because the Provider is not official, the client proactively cuts off the capability.
This is the true meaning of "client-side dumbing down."
The good news is, after reverse engineering, I have found a way to bypass the restrictions and restore raw image generation and Web Search.
I have already posted a detailed modification tutorial in the previous tweet:
If you are using a third-party Codex relay station, I suggest you check immediately.
You may think you are using the same model, but in fact, the client may have secretly neutered it long ago.
You can see the comparison before and after the modification in the attached images:
I had an in-depth experience with GPT-5.6-sol for a whole day.
My attitude went through roughly three stages:
Excitement → Doubt → Understanding and Acceptance.
First, let me clarify a point that is easy to confuse:
The highest reasoning level of GPT-5.6-sol itself is actually xhigh. The Max and Ultra shown in the UI mostly refer to the client workflow modes, not simply making the model "smarter."
Today, the biggest complaint about Ultra is that it likes to automatically dispatch a large number of sub-agents.
Even for a task that isn’t very complex, it might instantly launch a bunch of agents, consuming a lot of Tokens. The whole process looks grandiose, but the actual benefit may not match the cost.
At first, I was a bit skeptical:
Is it really working, or just showing off by burning Tokens?
Later, after understanding its mechanism, I lowered the reasoning level to High and reran some real tasks, and the experience actually improved a lot.
The speed is faster than GPT-5.5 xhigh, with noticeably stronger capability and stability, and Token consumption is not as exaggerated as Ultra.
At least based on my experience today:
GPT-5.6-sol + High might be the best combination for daily use currently.
For complex refactoring, full project reviews, or large tasks requiring parallel exploration, Ultra can be considered.
For ordinary development tasks, there’s no need to max it out right away.
This model is not unusable, nor does it require blindly running at the highest setting.
The real problem is that many people haven’t figured out the difference between Max, Ultra, and reasoning levels, yet they’ve already started using Ultra as the default mode.
Used correctly, it is indeed much stronger than 5.5.
Used incorrectly, you can only watch Tokens flow out rapidly.
A piece of advice for everyone! If you are going to use 5.6 sol.
Please temporarily throw Superpowers into the trash!
Otherwise, you will see the model's left and right brain battling continuously for hours! Eventually, the problems will multiply endlessly!
老叶1999.eth
Sorry, I really rushed it, please forgive me!
I spent the whole day deeply using it and felt that 5.6 sol was too slow, and secondly, it was too complicated. A small problem, the more you fix, the bigger the scope!
Sorry, Ultraman, this time I really can't help you brag! After teasing for so long, is this it?
When I found out that 5.6 was actually developed using the 5.4 base, I knew it was doomed!
Sorry, I really rushed it, please forgive me!
I spent the whole day deeply using it and felt that 5.6 sol was too slow, and secondly, it was too complicated. A small problem, the more you fix, the bigger the scope!
Sorry, Ultraman, this time I really can't help you brag! After teasing for so long, is this it?
When I found out that 5.6 was actually developed using the 5.4 base, I knew it was doomed!
