In late March, release metadata briefly surfaced details for a model above Opus. 10 trillion parameters. Anthropic hadn't announced it. The name: Mythos 5. A companion mid-tier model called Capabara came with it. The Haiku/Sonnet/Opus triad, which the AI community had treated as a stable architecture, turns out to be a floor, not a ceiling.
The dominant story right now is that AI is commoditizing. Models are getting smaller, cheaper, faster. Inference costs have dropped sharply. Edge deployment is becoming real. That story is correct.
But it only describes one direction of movement.
While the floor drops, the ceiling is also expanding. The Mythos leak suggests Anthropic is building a tier of capability that won't run on your laptop, won't be accessible through a standard API, and won't be priced like a utility. A 10T parameter model reserved for cyber reasoning and academic-grade analysis isn't a product you buy per-token — it's infrastructure you contract for, probably at the enterprise or government level.
Two tiers, not one curve
Most current architectures treat model selection as a point on a single curve: bigger model for harder tasks, cheaper one for simpler tasks. That model breaks when the harder tier is structurally inaccessible to most buyers.
What's emerging looks more like two distinct economic goods. Commodity intelligence: fast, cheap, good enough for most tasks. Haiku-class models, eventually running locally. The floor of this tier keeps dropping, and that's great for routine automation. Then frontier intelligence: ultra-heavy, reserved for problems where the cost of being wrong is measured in contracts or lives. This tier won't commoditize on any planning horizon that matters for the next three to five years.
The gap between them isn't just cost. It's access structure. If Mythos-class capability comes through enterprise or government agreements rather than standard APIs, then the players who get there first aren't necessarily building the best product. They're the ones with the procurement relationships.
What this breaks
The "AGI wrapper on Sonnet" play looks different through this lens. If your product sells frontier intelligence as a service, and your actual access to that intelligence is mediated by a commercial API at standard pricing, you're not building a moat. You're building on Anthropic's distribution decisions. When a new tier surfaces above your tier, your differentiation gets compressed from above.
The useful question isn't "which model do I use for which task." It's "which of my tasks are in the commodity tier, and which ones require frontier-tier reasoning, and do I have reliable access to both?" For most teams, the honest answer: commodity tier yes, frontier tier maybe, frontier tier with reliable SLAs probably not yet.
What I'm doing about this
Our agent architecture already separates by task profile. Routine orchestration — parsing, routing, summarization, code generation at scale — runs on cheap models and local inference where possible. We capture the cost deflation aggressively.
For tasks where the cost of a wrong answer is high, we treat frontier API access as a separate cost line, not an afterthought. The budget isn't computed per-token anymore — it's computed per-decision. That reframe changes when we invoke a frontier model versus using a cheaper one with a human review step.
The Mythos leak adds a third assumption to that picture: there is likely a tier above what we currently access, and I don't know the pricing, access model, or timeline. That uncertainty changes how I think about building durable positions in verticals that actually require that level of reasoning. If that tier is locked behind enterprise agreements, the game in those verticals becomes about access, not just quality.
The commoditization story is real. But it describes one axis. The other axis — frontier capability expanding upward, getting heavier and more expensive — is equally real and less discussed.
Building as if there's one curve means you'll be surprised when the bifurcation makes your position legible: you're either on the commodity tier or you're not.
Figure out which tier your use case actually requires. Then check whether you have reliable access to it.