It's the first time in the history of the world.
OpenAI connects to the big game!
Today, a Slack message inside OpenAI broke that ChatGPT will launch an "epic-level" performance exchange.
In terms of hardcore test data, the optimisation of the ChatGPT front-end performance is exaggerated to almost "unreal":
Application upload speeds rose directly by 94%, stack memory growth decreased by 87.8%, and overall memory occupancy decreased by 41.2%.
The number of network requests dropped by 98.2%, while the overall conversation record load decreased by 99.6%.
Almost at the same time, the Codex launched the GPT-5.6 version of "Multi-Smart Body v2".
Major Agents can now assign different subtasks, automatically assign them to different models, and each sub-Agent can also reason on its own strength.
Greg Brockman, president of OpenAI, said, "We are moving in the direction of saying goodbye to the manual selection model".
There's no denying that OpenAI's wave of "double pipes together".
Directly eliminating the burdens of "waiting" and "choosing" for the user, ChatGPT truly delivers a fast, seamless flowing interaction experience.
ChatGPT is a blood exchange.94% loaded with rabiesWhat many people don't know is that this performance upgrade of ChatGPT was actually a key battle after the departure of "King of Programming" Scott Gray.
Since then, people have been using ChatGPT, and what they fear most is that the longer the conversation is, the more they talk about the card, the more the web page will collapse.
On the official OpenAI forum, someone posted a post asking "Is the new version of the Codex slowing down?" with an appendix below.
This time, Andrew Ambrosino, the head of the Codex, unveiled an internal Slack screenshot, directly to you:
The internal test was a 741 rounds of 231MB "monster-level" conversations.
It sounds like an extreme example, but it's the day-to-day life of the Agent era.
The conversation used to be a conversation record, with dozens of rounds completed; now it's a slightly more complicated task to have it read the code, run the test, change it again, and hundreds of rounds come out.
This is the kind of high-intensity conversation that is now open, with the average time dropping from 27.62 seconds to 1.66 seconds.
The memory of the entire application increased from 1030.7 MiB to 606 MiB;
The number of sessions requests that need to be loaded has dropped from 15529, to 64.
This is a major front-end optimization of OpenAI for ChatGPT ultra-long conversations.
What changes can it make?
Simply put, those months-long chat records open up much faster.
Developers who frequently run heavy codex sessions using hundreds of tools will find that the interface is noticeably lighter and the device's memory occupancy decreases with it.
Cutting back a very old conversation is no longer as hard as "reading archives from a fast-decomposed PS3".
The trick is that ChatGPT no longer has to load and render all historical records because you open the conversation.
It can store most of the content, only loading a small portion of what people really need right now.
The same dialogue spared a lot of unnecessary work.
For those who are accustomed to over-the-top AI conversations, this is one of those infrastructure upgrades that sounds boring, but can dramatically improve the product experience.
Multi-smart body V2 up and runningGPT is not hand-selectedThis week, OpenAI has released an update to ChatGPT.
It wasn't until today that OpenAI engineer Eric Provencher officially announced that Multi-Smart Body v2 (Multi-Agent v2) was fully available.
The main one is clever division of labour.
The main agent can automatically assign subtasks to arbitrary support models, including Luna, with each subagent supporting an independent set of reasoning strengths.
Now, in the hands of ChatGPT/Codex, this card is like this.
- GPT-5.6 Sol: The strongest and most complex Agentic code
- GPT-5.6 Terra: The Power of Daily Programming
- GPT-5.6 Luna: the fastest and cheapest
- Daybreak: Cybersecurity special
- GPT-5.5: Complex programming, research and general tasks
In fact, three weeks ago, in the Model List of the Codex, Sol and Terra were labeled as multi_agent_v2, while the cheapest Luna was labeled as v1.
As a result, the master agent wanted to send a live feed to Luna, and the system responded directly to a "unknown model".
On GitHub, the developers have opened at least two issues for this.
In a post on the official OpenAI forum titled "Give us back Luna", the person who posted the post said that he had previously bypassed the restrictions on using Luna with a hook, which was surprising, and now the road is blocked.
In the past, people had to select models manually, and in the future, the Agent system will automatically break down, execute and aggregate results, converting the model into an internal computing resource.
Only 20% of the steps for complex tasks require the strongest model, and the rest can be handed over to cheaper models.
This will significantly reduce the cost of reasoning and accelerate the transition of ChatGPT from a chat tool to a workflow platform.
ChatGPT evolved locallyAll-Automatic Monsters are ComingThe front end clears the "history box" and the back end opens the "smart distribution".
At the bottom of OpenAI's wave of "combined fists", ChatGPT is not only saying goodbye to the Cartoon era, but also announcing to the entire industry:
The big model is accelerating its evolution from a "chat tool" to a truly "full-automated workflow platform".
It is no longer necessary to worry about which model weighs the cost, and hundreds of rounds of monster-level conversations slowly unfold.
Just throw away the complex tasks, and the rest is automatically disassembled, modeled and rendered at high speed, and handed over to the system.
The era of Agent's extreme slippery slope is really here.
In the face of such a "senseless" and powerful ChatGPT, is your productivity ready to take off?