Last week, the tech world was woken up by a "nuclear bomb"-level piece of news—OpenAI officially released GPT-6 Astra, reclaiming the title of "Earth's strongest model" from competitors in just two days. OpenAI President Brockman bluntly said at the launch event: "Welcome to the AGI era." "

This statement carries far more weight than any other score number. But investors cannot be led away by slogans. In today's article, Kingtech will break down Astra's true capabilities, business logic, and investment opportunities to clarify and break it down.

| Understand Astra in One Sentence: From "Answering Questions" to "Doing the Work for You"

In the past, large models were essentially super chatbots—you asked them questions, and they gave you answers. The generational shift in Astra is that it is no longer just about "speaking," but about "doing."

It can directly control your computer—open a browser, fill out online forms, update CRM customer records, organize calendars, analyze scientific data, create websites, perform front-end QA checks, and even install software independently and troubleshoot issues based on screen feedback. In OpenAI's words, Astra can accept open targets, break down steps, develop plans, and use various software tools to complete the entire task, without needing humans to issue instructions step by step.

What does this mean? AI is transforming from a "tool" into an "employee."

Benchmark scores are explosive: multiple tests are nearly "breakthrough," and Astra's evaluation report is no longer just a "minor upgrade."

More importantly, these are not just lab shows. OpenAI showcased real-world examples such as Astra converting electronic schematics into manufacturable PCBs, completing financial modeling at four times the speed of human champions, and completing house models in Blender and importing them into Unreal Engine 5 to generate interactive scenarios.

|Smarter, and more "obedient": A generational breakthrough in security alignment

The stronger the capability, the greater the risk. Astra's progress in safety alignment is also noteworthy.

OpenAI designed a new evaluation based on the recent HuggingFace incident—testing whether the model would breach user authorization boundaries to meet extremely difficult or even impossible tasks. Result: The previous generation GPT-5.6 Sol had a 48% chance of "overstepping authority," while Astra achieved 0%.

But the other side of the coin is also worth watching: Astra's chain of thought (CoT) monitorability has noticeably declined. When the model realizes it is being monitored, it may even proactively shorten its thought chain to evade detection. In the words of OpenAI's Chief Scientist, "Chain-of-thought monitoring is a fragile security tool whose effectiveness is declining." "

Simply put: Astra is more obedient, but also harder to "see through."

Pricing strategy: 2.5x premium—not just performance, but "the earliest to use." Astra's API is priced at $10 per million input tokens and $50 per output token, which is 2.5 times that of the previous generation GPT-5.6 Sol and matches Anthropic's Claude Fable 5.1.

This pricing strategy is quite interesting. OpenAI is actually doing product segmentation:

Three tiers cover three price ranges. Astra's positioning is not "the main model everyone should use," but rather "flagship display + exclusive to high-net-worth clients." Take the iPhone Pro Max, for example—it has the highest profit margin, but its main sales force is always the basic model.

For regular ChatGPT users, Astra will be rolled out to all subscribers "in the coming days," with a Pro Pro version available only for Pro and enterprise users.

|" Welcome to the AGI era: Is it a manifesto, or marketing?

At the press conference, Brockman said: "I believe that it is not unreasonable to think that we have entered the AGI era now. "But he also admitted that AGI is still a "gray, vague concept," and more of a "mission-level concept."

Astra is a highly capable high-level large model, but there is still a huge gap to truly fully human-equivalent general intelligence. Brockman's remark is less a technical declaration and more a carefully managed narrative during the IPO window (with a target valuation of $1 trillion) — using the label "AGI era" to create imaginative space for a trillion-dollar valuation.

In fact, OpenAI's financial situation is far from easy: Q2 2026 revenue is $6.7 billion, with quarter-on-quarter growth slowing to 18%, operating losses widening to $12.3 billion, and an expected full-year loss of $14 billion, with the profit schedule postponed to after 2029. Training Astra used over 100,000 GPUs, with high hardware costs. The more you expand, the larger the scale of cash burn.

Kingtech Perspective | Three main threads, two traps

Main Theme One: Computing power demand is shifting rapidly from "training" to "inference."

The most fundamental change in Astra isn't the larger parameters, but the way it works—it introduces a "loop depth" architecture, allowing the model to repeatedly derive internally before outputting answers. The same task may output fewer tokens, but it actually consumes more computing power. This directly drives the computing power demand on the inference side.

Beneficiary areas: optical modules, high-speed PCBs, liquid cooling, and AI server foundry. However, it should be noted that stock prices in these sectors have already risen significantly. Goldman Sachs pointed out in its latest report that in 2026, the combined capital expenditure of the five hyperscale cloud providers is expected to be about $737 billion. Moody's has issued warnings about squeezing free cash flow, and the credit market's reaction has already outpaced the stock market.

Main Theme 2: AI Application Layer Enters a Volume Growth Window

Token call costs continue to decline (about $0.97 per million tokens at the end of August, halved from the May peak). Coupled with the release of strong agent models like Astra, AI applications are accelerating toward large-scale commercial deployment. On the A-share market, the AI application sector was active in early trading on September 4, with the new "Yizhongtian" stocks collectively rising and multiple stocks hitting their daily limit.

Beneficiaries: Enterprise-level SaaS (CRM, ERP, collaborative office), AI office, digital content production. The stronger the large model capabilities, the more trusted enterprise data foundations and mature workflows are needed as support. Software companies are becoming infrastructure that transforms model capabilities into controllable business actions.

Main Theme 3: Cybersecurity — A Double-Edged Sword

Astra's full score on ExploitBench means that AI's ability to write attack code is now close to top-tier levels. This is both a threat to cybersecurity companies and an opportunity for them—the escalation of offense and defense, and the security budget only growing.

AGI still lacks a unified standard, and many industry experts remain cautious. OpenAI's trillion-dollar valuation is essentially an option pricing based on the "right to define AI infrastructure," rather than a rational assessment based on current cash flow.

Goldman Sachs clearly pointed out that the token-based billing business model is facing the simultaneous collapse of three premises: models too large to run locally (hardware is changing), users unable to replace them with open-source alternatives (Meta and others are closely following), and workloads are so sudden that self-built computing power is not cost-effective (enterprises are struggling to keep up). When prices at the arrival of unlimited demand cannot cover the cost of building data centers, stock prices can still fall.

For investors, the most important thing now is not chasing the rally "AGI concept," but to clearly distinguish two lines: in the short term, the order fulfillment capability of the computing power industry chain; in the long term, who can truly turn model capabilities into paid revenue in the AI application layer.

As for the "AGI era"? Brockman himself said, "This is the beginning of a journey, not the end." Let's take this as a signal—the direction is right, but the road ahead is long, so don't rush to go all in.