Anthropic Is About to IPO. The Code Demo Is the Best Argument It Has.
Anthropic has filed an S-1 registration statement with the SEC, putting the company on a path toward a public debut that will be the first for a frontier AI lab that has built its reputation on safety as a competitive differentiator. The filing is happening at a moment when the company's coding tool has become the clearest real-world evidence of a broader pattern: AI is eliminating the step between a person deciding what they want and a machine building it.
At a Code with Claude event in San Francisco on Tuesday, the company demonstrated a version of its AI coding assistant that can take a feature request in plain English and produce a working pull request — a chunk of code submitted for review — without the developer reading a single line of it. Nearly half the people in the room raised their hands when asked if they had shipped a pull request written entirely by Claude, according to MIT Technology Review. Most hands stayed up when the follow-up asked whether they had not read the code at all.
The same elimination of the intermediate step is showing up in other knowledge work domains. Legal teams at several large firms have begun piloting AI tools that take a case brief and produce a draft motion; financial analysts are using models that turn a data query into a finished memo. The pattern is consistent across cases: the human moves from writer to approver. What Anthropic demonstrated in San Francisco is the coding version of that shift, played out in front of an audience of developers who are living it.
The shift changes where the bottleneck in knowledge work sits. If AI can build across domains — code, contracts, analysis, design — the scarce resource becomes the human who knows what to ask for and can evaluate what comes back. Anthropic's product lead Angela Jiang said at the event that the absolute end state they are trying to get to is Claude being able to build itself, MIT Technology Review reported. If that happens, the remaining human task is specification: deciding what good looks like, and catching errors in what AI produces.
Anthropic's engineering lead Katelyn Lesse said at the same event that Claude is probably as good as a midlevel engineer at writing code, MIT Technology Review reported. The productivity claim is self-reported and the audience was self-selected, so treat it as directional signal rather than verified benchmark. The more immediate open question is what happens to code quality when developers stop reading what ships. Security researchers have flagged that unreviewed AI-written code introduces vulnerabilities that existing review processes are not designed to catch. That concern has not stopped adoption; it has slowed it in regulated industries while proceeding faster in startups and individual workflows where individual engineers bear the consequences directly.
Context windows have stayed at roughly 1 million tokens for over a year across AI providers, Chris Ebert noted in his writeup of the event. Anthropic's answer to that ceiling is a feature called dreaming, where the agent writes notes to itself between sessions, consolidating patterns so it does not have to relearn context each time it picks up a task. The feature was announced two weeks prior and is a structural workaround rather than a raw capability jump — the model is not suddenly smarter, but it is better at maintaining continuity across longer work sessions.
The dreaming feature matters in practice because it changes the economics of long projects. A developer who previously had to re-explain a large codebase to an AI each session now leaves the agent to maintain context across days or weeks of work. That changes what a team can attempt: a project that once required two weeks of setup time might now be underway in hours.
Demand for Claude's coding tools has grown a startling 80x so far this year, CEO Dario Amodei said at the event, without specifying the base period or exact time frame. That growth runs against a backdrop of heavy spending to train and run frontier models while showing enough momentum to make the public markets case.
Anthropic signed an agreement with SpaceX for all of the compute capacity at the Colossus 1 data center: more than 300 megawatts and over 220,000 NVIDIA GPUs, according to an Anthropic blog post. The deal is one of the largest single compute infrastructure commitments in the industry's history and reflects the capital intensity of the bet Anthropic is asking public investors to share.
What to watch next is whether the S-1 disclosures confirm the revenue trajectory Amodei has cited privately, and whether the IPO allows Anthropic to access public equity markets at a valuation that reflects its infrastructure commitments. The intent-to-execution story and the IPO story are converging: Anthropic is about to ask public investors to fund a future where AI handles more of the building, and the filing will be the first public test of whether that future looks as close as the demo suggests.
The implications extend beyond individual developer workflows. If AI systems can take a specification and produce working code, legal documents, or financial analysis without the human reading the output, the entire quality-assurance layer that exists inside organizations today — code review, legal review, peer analysis — becomes optional rather than standard. That is the real shift Anthropic is pointing toward, and it is the one that will be hardest to walk back once it becomes normal practice.