Token Monsters: The New Economic Playbook for AI Vendors
When Anthropic announced Mythos, the company framed it as a model so capable that withholding release became part of the message. The implication was not subtle: this was a system with unusually powerful offensive capabilities, demonstrated in part by its ability to discover obscure vulnerabilities in old FreeBSD kernels. But subsequent analysis suggested that this feat may not have been as singular as the announcement implied. If other frontier models can achieve similar results under comparable conditions, then the real question is not whether Mythos has superpowers. It is why the story was told that way in the first place.
A plausible answer is that the announcement was less about technical discontinuity than economic positioning. AI model vendors are no longer competing only on benchmark scores or raw intelligence. They are increasingly competing on the ability to embed inference inside persistent, high-frequency workflows that generate recurring demand. In other words, they are looking for token monsters.
A token monster is a workflow that runs continuously, ingests ongoing machine-generated events, and requires sustained inference over time. These are not one-off prompts or occasional copilots. They are systems that stay awake. They monitor, classify, triage, decide, escalate, and act. Their value lies not in a single impressive answer but in their ability to consume an endless stream of events and convert that stream into billable model usage.
This helps explain why code generation and cybersecurity have become such attractive categories for model vendors. Both are event-dense, machine-mediated, and structurally open-ended. Codebases change constantly. Build pipelines fail. Dependencies break. Security alerts fire. Logs accumulate. Configurations drift. New vulnerabilities emerge. These are environments where intelligence does not appear once and disappear. It remains in the loop.
That persistence matters because the economics of AI are not especially attractive when products are used only intermittently. A chatbot consulted a few times a day is useful, but it does not necessarily create deep or durable revenue. A system that is continuously scanning repositories, reviewing pull requests, tracing failures, investigating alerts, or proposing remediations is different. It generates a much thicker layer of ongoing inference demand. The winning workflow is not merely helpful. It is metabolically expensive.
This is why product moves around the major model labs increasingly point toward service layers rather than standalone APIs. The interesting play is not just to sell a model call. It is to wrap intelligence in tooling, orchestration, monitoring, and enterprise workflow. Once a vendor sits inside the operating loop of a valuable system, the revenue profile changes. The model is no longer consumed ad hoc. It becomes part of a recurring process.
Taken together, the benchmark narrative around Mythos and the accompanying product posture suggest that the economic value may lie less in the model alone than in the operational layer around it: tooling, deployment, orchestration, remediation, and service delivery. This does not mean models no longer matter. They do. But in commercial terms, the model may increasingly function as one component in a larger stack whose real purpose is to lock inference into persistent workflows.
If that is the logic, then the next question is not whether AI can automate work in general. It is which categories of work naturally produce nonstop inference demand.
The first category is software. Modern software development is already mediated through machines. Repositories, CI pipelines, linters, tests, runtime telemetry, issue trackers, deployment logs, and infrastructure state changes all produce structured signals. This creates fertile ground for systems that can continuously observe and intervene. A coding assistant that merely helps complete functions is useful. A system that continuously watches the software lifecycle, spots regressions, proposes patches, explains failures, and helps guide remediation is economically much more interesting.
The second category is cybersecurity, which may be the purest token-monster domain of all. Security operations are built around endless event streams: endpoint alerts, identity anomalies, traffic spikes, suspicious binaries, misconfigurations, and changing vulnerability data. Every enterprise already has too many signals and not enough attention. This makes cybersecurity a natural home for persistent inference systems that can triage, correlate, investigate, and recommend action. Unlike a casual consumer application, a security system does not wait politely to be asked a question. It lives in the flow of events.
The same logic applies beyond code and security. Financial markets generate continuous signals: price movements, filings, news, analyst revisions, transaction anomalies, and macro data releases. Industrial systems produce telemetry from sensors, machines, and process-control environments. Healthcare generates streams from imaging, diagnostics, wearables, monitoring systems, and administrative events. In each case, the commercial opportunity is not simply “AI for X.” It is AI attached to a stream that does not stop.
That distinction matters because it clarifies where defensibility may actually form. Much of the popular debate in AI still assumes that the central moat is model quality alone. But if model capabilities converge at the top end, then the stronger position may belong to whoever owns the harness around the model: the workflow, the interface, the integration layer, the remediation loop, the trust boundary, and the event stream itself.
This is also why application companies should not assume they are doomed to be crushed by the labs. If the key asset is not only frontier intelligence but persistent workflow control, then there is plenty of room for value to accrue above the model layer. A company that owns the operational context, the user relationship, the deployment surface, and the event pipeline may be in a stronger position than a model vendor with excellent raw capability but weak embeddedness.
Seen this way, the model lab strategy begins to look less like a race to build the smartest oracle and more like a race to attach inference to the thickest, stickiest sources of machine-generated activity. The economic objective is not simply to produce occasional brilliance. It is to become part of an always-on process that repeatedly consumes tokens because the underlying environment keeps generating new states to interpret.
This is the deeper significance of categories like coding and cybersecurity. They are not just fashionable or technically impressive. They are commercially attractive because they produce recurring inference demand with relatively clear willingness to pay. They offer a path from intelligence as a feature to intelligence as an operating layer.
That is what the Mythos story may really point toward. The future battle in AI may not be decided solely by who has the most impressive model in isolation. It may be decided by who can bind model inference to the largest recurring streams of machine-generated activity, and then wrap that inference in a product surface that enterprises trust enough to leave running.
The next great AI businesses, in other words, may not be built on one miraculous answer. They may be built on systems that never stop having to answer.