Just now, GPT-6 has increased speed by 50% across the board

Author: ASI Revelation, New Intelligence Yuan

Editor: Peach, New Intelligence Yuan

In the next 28 days, ChatGPT is set to make a full comeback!

Just now, "Foster Father" Tibo suddenly announced the Day 1 plan, igniting the entire internet------

The default inference speed of GPT-6 Astra and GPT-6.1 Sol has increased by 50%

TPS has skyrocketed from the previously sluggish 30 to 50, covering all products logged into ChatGPT, as well as OpenCode, Pi, Amp, Devin, and others.

Moreover, the entire GPT-6 series has switched to the "optimized tokenizer," requiring fewer tokens to complete the same tasks.

The AI community hadn't even had time to shout "Ultraman's conscience" when a heavy blow struck Silicon Valley.

SemiAnalysis released extreme testing results, comparing the subscription value of the two giants across the internet, and found that:

In heavily loaded scenarios like Agent programming, Anthropic's Opus 5.5 provides actual API value that is over 5 times that of GPT-6.1!

Now, the GPT-6 series runs faster, but is the money still worth it?

GPT-6 Series Speed Increase of 50% Day 1 Delivered

Just yesterday, Tibo set a bold flag for himself on X.

In the next 28 days, there will either be a visible improvement every day or a reset for everyone.

The countdown of 28 days starts today.

The first day's update directly addresses the biggest pain point for many developers------inference speed.

Starting today, the default generation speed of GPT-6 Astra and GPT-6.1 Sol has increased by a full 50%, jumping from the previous 30 TPS (tokens per second) to 50 TPS!

What's even more impressive is that this speed increase not only covers all official ChatGPT endpoints, but the entire ecosystem has been seamlessly integrated:

• The well-known AI programming tool Devin
• The popular personal terminal and workflow tools OpenCode, Pi, Amp

No matter where it's called, without changing a single line of code, as long as it's through the subscription channel, the speed instantly increases!

The official statement claims that this is not just an adjustment in computational concurrency, but more importantly, a replacement of the extremely aggressive underlying tokenizer.

This means that when faced with the same long code generation and complex mathematical analysis, the total number of tokens consumed by the model at the underlying level has been compressed to a very low level, achieving a qualitative leap in effective throughput.

However, many developers have tested it and felt that it is not very fast, and you need to go through the API channel to reach 50.

Claude Subscription Value Blows GPT Away by 5 Times in Authoritative Testing

On the same day, the well-known semiconductor and AI research institution SemiAnalysis released a report that can be described as "slapping the face."

They directly purchased the high-priced subscription services of OpenAI and Anthropic to conduct long-term programming and heavy-load tasks like Agent testing until reaching the weekly usage limit, then calculated the theoretical costs based on the public API prices.

The results were shocking.

At the $200 monthly fee level, the equivalent API value of ChatGPT Pro dropped from about $14,000 to about $2,084 after the limit was halved.

In contrast, the Claude Max 20x plan reached about $11,726, with a difference of over 5 times. More intuitively:

$200 tier: Claude's equivalent value is about 5.6 times that of ChatGPT

$100 tier: Claude is about 5.4 times that of ChatGPT

$20 tier: Claude is about 5.6 times that of ChatGPT

The well-known AI expert Chetaslua created a visual comparison chart of the bills for ChatGPT Pro and Claude Max.

Even more bizarre operations are yet to come.

Many people remember that before June of this year, OpenAI was the wealthy "price butcher," with the highest tier package capable of generating nearly $14,000 in computational power.

However, just last week, Ultraman made a swift move, directly halving the actual usage limit of the $200 ChatGPT Pro top-tier package.

The real reason behind this is hidden in the subscription data analyzed by SemiAnalysis.

Taking Claude Pro and Claude Max 5x as examples, when the actual usage reaches about 20%, Anthropic can achieve a break-even point.

In contrast, OpenAI's situation is more severe: for ChatGPT Plus and Pro 5x subscriptions, once the utilization exceeds about 11.4%, the company starts to incur losses.

In other words, these subscription plans are essentially betting that most users will not use up their limits.

However, the popularity of Agent programming is breaking this balance, with multi-step, self-invoking tool workflows consuming tokens at levels up to a thousand times that of traditional single-turn conversations.

Super Application Arrives! Four Major Entrances Completely Integrated

Beyond this, OpenAI has a grander strategy unfolding.

The latest revelations indicate that OpenAI plans to merge the four major product lines: Chat, Work, Codex, and Dots into a unified interface, transitioning to a single super application.

This is not only an integration at the UI level but also signifies a shift in the entire product philosophy.

Specifically, the changes include:

The model selector is hidden, and the inference process is no longer visible to users.

In the future, no one will be able to choose which model to use, nor will they know how many steps the AI "thinks"------you just need to tell it what to do and wait for the result.

All text generated by ChatGPT and Codex will be forcibly embedded with invisible watermarks to identify AI content.

Just this morning, OpenAI announced this.

Additionally, OpenAI's commercialization strategy is also being adjusted; users waiting for the image generator to process images will see inserted advertisements.

It is expected that the Chat and Work modes of ChatGPT will merge to reduce complexity, and all capabilities of Dots will ultimately be integrated into ChatGPT, covering 1.2 billion users.

Now, Day 1, 50 TPS has been submitted.

The remaining 27 days will be OpenAI's real test!