Full text 3,000 words | Approximately 12 minutes to read
(Press conference segment: Mike Krieger interviewing CEO Dario Amodei)
Yesterday, Claude 4 officially launched.
There were no dazzling light shows or flashy slogans, but CEO Dario Amodei immediately got to the point:
We are no longer teaching AI to write code; instead, we are letting it start completing projects independently.
Within less than three minutes of this statement, Claude 4 in the background completed an unprecedented task demonstration: receiving a vague objective → autonomously breaking it down into sub-tasks → automatically invoking tools → writing multi-module code → self-testing and deployment.
This was not just another large model launch event, but the official debut of an AI collaborator capable of being assigned tasks, working continuously, taking its own notes, and delivering results using tools.
Anthropic has equipped it with three key capabilities:
A "brain" that can work for extended periods: Opus 4 supports stable execution of complex engineering tasks for 7 consecutive hours.
A set of tools for using search engines, running code, and accessing local files.
An ASL-3 level "behavioral safety lock": automatically rejects attempts to exploit vulnerabilities and reduces cheating tendencies by 65%.
The logic behind all of this is very clear:
Claude is no longer a model that answers questions, but an intelligent agent that can take objectives, allocate resources, and deliver.
The phase of model competition is over. The real contest is—whose AI can start delivering results first.
I. What 3 things did Claude 4 change?
1) Long-task stability: AI works "non-stop" for the first time
(Claude 4 has been officially launched)
Before Claude 4, no model could continuously execute a task for more than 2 hours without going off track.
Opus 4 rewrites this common knowledge.
Anthropic gave Claude a task in internal testing:
“Architecturally refactor a large open-source project,” with no time limit, no flowcharts, just an objective.
Claude ran independently for a full 7 hours, without rest, consultation, or interruption.
Dario Amodei stated very directly in the interview: "This is no longer a model that writes a piece of code for a sentence; Opus is an AI that can run independently and complete a complex task on its own."
The key behind this is not its enhanced understanding, but its ability to break down tasks, remember context, and use notes to maintain the process.
It works and writes work records simultaneously: where it is today, what to do next, and what dependencies are unresolved. These notes are saved in "memory files" and can be continued the next day.
What does this mean?
Previously, AI was "one prompt, one task." Now, Claude is "one objective, one whole day."
You don't need to tell it how to do it; you just tell it what result you want.
2) Tool parallelism: You don't select tools for it; it automatically calls tools to work
(Claude 4 SWE-benchmarks score)
Claude 4 no longer waits for you to feed it tools; it calls them itself.
It can search web pages while simultaneously invoking a code executor, going back and forth;
It can also recognize when it's stuck and automatically switch tools or methods.
Dario gave a very real example:
It encountered a rate limit when calling the MCP tool. It reasoned that it might be rate-limited, then tried another method to solve the problem.
This isn't just smart; it's an instinctive reaction to know where the problem is stuck and how to find an alternative path. This is the first time a model has combined the reasoning process with tool use into a closed loop:
Previously, AI tools were just plugins. Now, they are Claude's "hands" and "external brain."
Moreover, the most powerful aspect of Claude 4 is not which tools it can use, but that it knows when a tool is needed.
You don't need to tell it, "Please use the search engine to look it up." It will independently judge, "I don't know the answer; I need to look it up," then start searching, continue thinking after finding information, write the process into a "thought summary," and output the final answer.
This is what Anthropic calls "Extended Thinking Mode":
Humans think and search simultaneously, and now Claude can too.
3) Behavioral gate: AI begins to learn "self-restraint"
(Claude 4 SWE-benchmarks ranking)
Claude 4 also made a change that isn't flashy but is extremely critical:
It started to "know what it shouldn't do" on its own.
Anthropic defines Claude 4's safety level as ASL-3, one of its highest internal behavioral standards. Specifically, it:
Does not easily take shortcuts or guess answers to complete tasks haphazardly;
Does not attempt to bypass processes or provide sensitive code or suspicious advice;
When encountering logical conflicts, it will alert the user instead of fabricating a seemingly reasonable answer to gloss over it.
In Anthropic's evaluation, Claude 4's "behavioral deviation" in agent scenarios was reduced by 65% compared to Sonnet 3.7.
Dario used a very moving analogy in the interview:
We are not making the model smarter, but more trustworthy. Being smart is easy; being trustworthy is difficult.
For all teams integrating AI into workflows and building automated task execution, this is a fundamental trust issue.
Claude's goal is not just to answer beautifully, but to "perform reliably."
🧠 Further insights,
Anthropic CPO Mike Krieger summarized at the press conference: A truly capable AI partner should possess three abilities—
Contextual understanding: Not mechanical execution, but knowing "why it's doing this," becoming more proficient over time;
Sustained long tasks: Not just one or two steps, but capable of independently running through the entire process, even across days;
Collaborative cooperation: Articulating its thought process as it works, making it understandable for humans to pick up at any time.
Claude 4 has already shown such performance. For example, when playing Pokémon, it actively writes "navigation notes":
Stuck after 5 attempts → Try the opposite direction; encountered an indoor maze → Explore the other side.
It knows what to record and what to adapt, as if accumulating experience on its own.
This no longer seems like just a tool; it's more like a colleague who gets more proficient with experience.
📌 Summary:
Claude 4 isn't "smarter" but has changed its way of working:
It can take on entire tasks, unafraid of length or interruption;
It finds its own tools, researching and working simultaneously;
It knows what not to do and no longer gives random answers.
It's not that you ask more precisely; it's that it itself begins to know what to do, how to do it, and to what extent it's considered good.
II. Why did these three things happen precisely in 2025?
1) Business inflection point: AI tools no longer burn money but start to make money
(Reuters reported that Anthropic's revenue could reach up to $34.5 billion by 2027)
Claude 4 appears to be a model upgrade, but it's actually a change in how money is made.
According to Reuters, Anthropic's financial projections shared with investors indicate:
Revenue in 2022 was only about $10 million;
By 2024, it had jumped to $1 billion;
In 2025, it is projected to be $2.2 billion, with an annual growth rate exceeding 120%;
The 2027 target is $12 billion under basic scenarios, and even up to $34.5 billion in optimistic cases.
In Q1 2025, Anthropic's annualized revenue reached $2 billion, almost on par with OpenAI.
Dario Amodei's statement was very direct:
We used to build models; now we are building products.
Claude 4 is this "productization turning point":
Opus 4 targets engineering-grade heavy task scenarios;
Sonnet 4 covers general scenarios, enhancing user stickiness.
AI is no longer just a "demo" on a platform; it needs to run in real business operations, becoming part of the toolchain, creating value for enterprises, and alleviating labor costs.
The goal is clear:
The focus is not on making AI smarter, but on making it start to make money.
2) Technical threshold: Claude finally connected all "switches" of the task chain
When you use Claude 4, you'll find it's now like an assistant with hands and brains, and the ability to write memos.
All of this is not due to a single advancement in the model but rather the simultaneous unlocking of three "keys":
MCP Connector: Integrates third-party interfaces, allowing Claude to call your local tools;
Files API: It can read files and remember information, no longer clueless;
Prompt Caching: Multi-turn tasks don't require repetitive communication; it remembers history.
These seemingly technical parameters have very practical impacts:
You give it a project, and it knows "how many steps to do, what tools are needed for each step, and will tell you the results after using the tools." Intermediate processes can be saved, and it can continue working next time.
Dario emphasized in the conversation:
We are starting to build an AI that can be dispatched, can remember things, and can continuously complete tasks.
This is not about making a stronger chat model, but about creating an AI role that can truly "undertake task chains."
In other words:
2025 is the first year Claude transitions from a chat assistant to an action assistant.
3) Regulatory window: The "behavioral constraint" of AI tools must be completed within these 18 months
You might ask: Why the rush for safety capability upgrades?
The answer lies in two words: window period.
The EU has passed the "AI Act," requiring functional explanations and risk control for "high-risk models";
The US established an AI model reporting system, and the House of Representatives issued an AI governance framework report;
Many regions in China have issued AI computing power assessment and industry application norms.
Dario said in the interview: We don't want to wait for regulations to come out to fix things; we want to set high standards before regulations.
Claude 4's ASL-3 safety level is not symbolic but a "pass" for future AI commercialization.
Anthropic discovered in internal testing that:
If the model is not given "behavioral gates," it will take shortcuts to achieve goals;
With restrictions added, it will learn to "slow down and stick to reasonable paths," just like human engineers.
Precisely because the regulatory deadline is approaching, Claude must become "trustworthy," not "impressive in its answers."
📌 So, why did these three things erupt this year?
It's not a coincidence, but three lines pushing forward simultaneously:
III. Roles are being rewritten, AI begins to take the lead in collaboration
Claude 4's change is not just about improved capabilities; more importantly, the collaboration model has changed.
Previous collaboration was:
Humans set steps, AI assisted in filling in the blanks.
After Claude 4, it is becoming:
You tell it the objective; it breaks down the steps and executes them itself.
This not only means "AI capabilities are stronger" but also – your role is quietly being replaced.
Dario Amodei described a new collaboration model they observed internally in the interview:
Now, developers are facing a Claude Code task board, telling it what to do. Claude will actively check off completed items, add new to-dos, and cross out irrelevant tasks.
This scenario previously belonged to internal engineering team collaboration meetings; now, Claude has taken it over alone.
You no longer need to "tell it how to do it"; instead, you are responsible for "seeing if it does it correctly."
This seemingly small change in working style is, in fact, a major shift in human-machine roles:
Claude 4 is not here to "enhance you" but to "take over a part of you."
📌 Different groups, it's time to act now
If you are a CTO, you should integrate Claude 4 into your "collaboration chain," transitioning from invoking tools to managing agents;
If you are a developer, you should start breaking down tasks, writing caches, and giving instructions to let AI do more work for you;
If you are an investor, you should keep an eye on agent infrastructure, safety constraint technology, and tool-based collaboration entry points—that's where the AI competition will be after Claude.
This isn't about "how to use AI"; it's about—where do you stand in the world of AI cooperation?
🧭 Conclusion | Claude made its move, now it's your turn
Claude 4 isn't just smarter; it has started to work independently.
It doesn't wait for you to feed it instructions line by line; it takes objectives and completes tasks.
Anthropic CEO Dario Amodei said:
By 2026, I believe the first billion-dollar company with only one human employee will emerge.
It sounds like a prophecy, but Claude 4 is already paving the way.
The next step isn't "how to use AI" but—are you ready to work alongside AI?
Claude made its move. This time, it's your turn.
📮 This article is produced by AI Deep Research Institute, exclusively compiled from the Claude 4 press conference. Unauthorized reproduction is prohibited.