How a coding wedge became the management layer for how software gets built, and what the IPO filing now asks the public to price
By Jennifer Evans, Principal, Pattern Pulse AI; Co-founder, Tech Reset Canada
On June 1, Anthropic confidentially filed a draft registration statement with the SEC. The number of shares and the price are not set. The filing stays sealed while the regulator reviews it, and a public debut could land as early as this fall. The framing across the financial press was a three-way race to the public markets between Anthropic, OpenAI, and SpaceX, three of the largest listings the market has ever contemplated.
The number that sits underneath the filing tells the real story. Anthropic reported a revenue run rate of roughly $47 billion, up from about $10 billion a year earlier, and closed its most recent private round at a $965 billion valuation, ahead of OpenAI. A company that earned its first dollar less than three years ago is asking public investors to value it near a trillion dollars.
The filing is the visible event. The strategy that produced it came from a deliberate decision to go after the corporate market and to do it through code, then to expand outward from code into the operating system, the agent layer, legacy infrastructure, the law office, and the security stack. Each expansion landed as a product. Several of them moved markets. The throughline is a company that stopped selling a model and started selling the layer where software actually gets made.
Background: The Coding Wedge
Claude Code launched publicly in May 2025. By November it had crossed $1 billion in annualized run-rate revenue. By February 2026 that figure had more than doubled to $2.5 billion. The growth was not driven by a campaign. It came from engineering teams evaluating the tool on real work and adopting it through word of mouth, the most durable adoption there is.
The market-share figure is the cleaner signal. Per Menlo Ventures, Anthropic’s share of the enterprise coding-model market rose from 42% in mid-2025 to an estimated 54% by December, against OpenAI’s 21%. A twelve-point gain in six months, in a category that was already concentrated, is unusual. Concentration normally plateaus once a category matures. Here it widened into the largest single-vendor lead of any LLM workload category.
Code also became the commercial center of gravity inside Anthropic. Claude Code accounts for more than half of all enterprise spending on Anthropic products, and enterprise use makes up over half of Claude Code revenue itself. The wedge is not a feature line. It is the load-bearing wall of the business that is now filing to go public.
The reason code worked as a wedge is the reason it generalizes. Code is where the pilot-to-production gap is most expensive. A demo that writes a function is trivial. A system that ships pull requests, survives review, and runs reliably inside an existing codebase requires the scaffolding most organizations cannot build: sandboxes, state management, error recovery, verification. Anthropic’s bet was that the value was never the raw model. The value was making that gap crossable. Claude Code worked because it needed less governance overhead to be trusted, not because it had novel architecture. That insight is what the company then carried into every adjacent layer.
Hypothesis: The Layer, Not the Model
The argument here is that Anthropic’s 2026 is best read as a single strategic motion: occupy the management layer of software production, one adjacent surface at a time, and let the model recede behind the layer.
A management layer is the place where work gets coordinated, executed, and governed. It sits above the raw capability and below the human decision. Anthropic moved into that position deliberately. It built Claude into the operating system, into the agent-orchestration layer, into legacy modernization, into legal workflows, and into security review. In each case the pattern held: take the infrastructure that teams used to assemble themselves, fold it into the platform, and price it as a service. The strategic question Anthropic answered was not how to build a smarter model. It was which layer of software work to absorb next.
If the hypothesis holds, two things should be observable. First, incumbents whose business sat in the layer being absorbed should react, and react financially. Second, the company should reach an outer edge where the strategy meets a constraint it cannot yet price away. Both are visible in the record.
A Claude for Every Layer
Start with the operating system. Anthropic extended its Cowork desktop agent to Windows and Intel-based macOS with full feature parity, then standardized it as a cross-platform layer that reads, writes, and modifies files inside a user’s working environment. The framing was deliberate. The integration target was not an app. It was the OS, the surface every other piece of software runs on.
Then the enterprise workflow surface. Anthropic shipped interactive integrations across Slack, Figma, and Asana, then a directory of pre-built, plugin-style agents for finance, engineering, legal, HR, and design. The positioning language was platform language: your work tools are now interactive inside Claude. Traditional SaaS applications were repositioned as subordinate services, with Claude as the orchestration layer coordinating across them.
Then legacy infrastructure. Anthropic highlighted that Claude Code could analyze and modernize large COBOL systems, the language still running banking, government, and mission-critical back ends. IBM fell 14% on the news, its worst single day in more than 25 years. The reaction outran the substance. The capability was real, but it was a PDF and a demonstration, not a product that would meaningfully erode IBM’s mainframe business in the near term. The selloff is the evidence that matters: the market now treats an Anthropic product reveal as a direct threat to whichever incumbent sits in the targeted layer.
Then security. Claude Code Security shipped as Anthropic’s first dedicated cybersecurity product, reviewing entire codebases contextually rather than scanning for known patterns. An internal proof of concept using Opus 4.6 surfaced more than 500 vulnerabilities in production open-source code that had survived years of human review. CrowdStrike and Zscaler each dropped more than 7% on the news. The pattern repeated: enter a vertical, demonstrate capability, watch incumbent stocks move.
Then the review layer itself. Code Review dispatches multiple specialized agents at each pull request, some hunting bugs, others verifying findings to suppress false positives, a final pass ranking by severity. It does not approve pull requests; that authority stays with humans. After internal deployment, the share of pull requests receiving substantive review comments jumped from 16% to 54%. The tool addresses the bottleneck the coding wedge created: writing code became cheap and fast, and review capacity did not scale to match.
Then the harness. On April 8, Anthropic launched Claude Managed Agents, a fully hosted service inside the Claude Platform. A developer defines an agent through natural language, system prompts, tools, permissions, and connections, and Anthropic runs everything else: sandboxed execution, durable session state, error recovery, scaling, security isolation. The reaction was a wave of posts declaring that Anthropic had killed thousands of agent startups overnight. The claim is overstated, and the overstatement is instructive. An entire startup category existed to sell the orchestration and sandbox infrastructure that crossing the pilot-to-production gap required. Managed Agents folds that infrastructure into the platform and prices it by the hour. The startups most exposed are the middle-layer infrastructure providers. Application-layer companies with domain depth and high switching costs are not displaced by a stronger platform; they are made more valuable by it. The category mismatch is the part the viral framing misses.
The Edge of the Strategy
If the strategy were frictionless, there would be no outer edge. There is one, and it is named Mythos.
Claude Mythos Preview, announced April 7, sits a full capability tier above Opus 4.7 and is not generally available. Access runs through Project Glasswing, an invitation-only program for twelve founding organizations and roughly forty vetted critical-infrastructure operators, with $100 million in usage credits seeded to participants. Its defining capability is autonomous vulnerability research. It identified zero-day flaws across every major operating system and browser, including a 27-year-old bug in OpenBSD and a 17-year-old remote-code-execution flaw in FreeBSD’s NFS server. It found weaknesses in the cryptographic libraries and protocols the software world runs on, including TLS, AES-GCM, and SSH, the primitives that secure HTTPS connections, encrypt data, and authenticate remote access. Anthropic published cryptographic commitments to prove the exploits existed and is holding technical detail until the underlying flaws are fixed.
Mythos extends the layer thesis to its limit. Every other product moves up the stack of building software. Mythos demonstrates the company can break the security layer underneath all of it. And here the strategy hits two constraints it has not priced away.
The first is access. Anthropic deliberately did not ship Mythos into the software layer the way it shipped Claude Code. The capability was judged to exceed what general availability could safely support, so it was gated. The pattern that defined every other product, push the capability out and let adoption compound, reverses at the frontier.
The second is cost, and it is the sharper one. Palo Alto Networks, a Glasswing member, reported that turning frontier AI models on its own product portfolio surfaced roughly 75 vulnerabilities, more than seven times what it typically finds in a month. Reporting this week added the price: running Mythos at Palo Alto found flaws the company would typically surface using existing tools, but it burned through more than $1 million in tokens to do it. The whole wedge was built on collapsing governance overhead, on this works Tuesday instead of six months of infrastructure. At the frontier, the most capable model in the stack matched what existing tools already do, at a cost that breaks the deployability story. The strategy that won by being cheap to deploy reaches a capability that is neither cheap nor open.
What the Filing Asks the Market to Price
The IPO filing is the strategy submitted for public scrutiny. A confidential S-1, a $47 billion run rate, a $965 billion last round, and a possible debut by fall. Public investors will now price what the private market priced on momentum.
What they are actually pricing is the layer position. Anthropic’s value does not rest on having the smartest model, a lead that compresses with every competitor release. It rests on having become the embedded management layer for how software gets built, across the operating system, the agent harness, legacy code, the law office, and the security stack, with the switching costs and ecosystem dependence that position creates. The coding wedge proved the thesis. Every product since extended it. The stock reactions of IBM, CrowdStrike, and Zscaler are the market already pricing the layer one incumbent at a time.
Mythos marks where the position stops being a clean compounding story. A capability the company can build and prove but chooses not to ship, that costs a million dollars in tokens to match commodity tooling in its one visible deployment, is the boundary of the strategy rather than its next step. The same layer that is cheap to deploy in production is, at the frontier, gated and expensive. A public market will have to decide how to value a company whose core position is real and durable and whose outer edge is neither settled nor priced.
The filing turns that question from a private bet into a public one. The answer will come when the SEC review closes and the shares carry a number.
Jennifer Evans is Principal of Pattern Pulse AI, and co-founder of Tech Reset Canada. Her research covers AI reliability, agentic systems, and LLM failure modes, alongside the Canadian AI Sovereignty Series.