Capability Compression

What It Is

Capability compression names a pattern now visible on a quarterly cycle: a model capability that required the most expensive, largest AI system in one season becomes standard in a cheaper, faster alternative the next. In the first half of 2026, Anthropic’s Sonnet line closed most of the agentic gap with Opus in a single model generation. OpenAI’s midtier absorbed the performance profile that once required Pro-tier access. Midtier models are no longer slower versions of their flagship siblings. They are frontier models from six to twelve months ago, priced for volume.

The compression is not gradual. It arrives in discrete jumps: a new Sonnet ships, and the prior definition of “advanced agent” becomes the new floor for free users. Each jump resets operator expectations and, more importantly, resets pricing. What cost $15 per million output tokens in late 2025 costs $10 in early 2026 and $3 by summer. The economics of building on AI are not merely improving. They are compressing on a curve most cost models have not accounted for, and the curve does not flatten in the near term.

Understanding capability compression is not about tracking benchmarks. It is about knowing when to revise your routing assumptions, your cost forecasts, and your product differentiation strategy, because those assumptions now expire quarterly rather than annually.

How It Actually Works

The mechanism behind capability compression is a combination of inference optimization, model distillation, and hardware improvement. Labs train frontier models that establish a new performance ceiling, then distill the most useful behaviors into smaller, faster architectures. Distillation is the technical process of training a smaller model to reproduce the outputs of a larger one. It now happens fast enough that the gap between flagship and midtier closes in months rather than years.

Simultaneously, inference hardware keeps improving. Custom silicon like Anthropic’s Trainium integration and competing labs’ in-house chips means the cost per token drops independently of model size. A smaller distilled model running on purpose-built inference hardware can deliver comparable output quality at a fraction of the compute cost of the flagship on last year’s hardware. The two curves, model compression and hardware cost reduction, compound each other. Every quarter, the same quality gets cheaper and the same price buys more capability.

Why It Matters Right Now

For anyone building on AI today, capability compression changes the planning horizon for both cost forecasts and product decisions. A use case that required Opus-level capability to be commercially viable in early 2026 may be economically feasible at Sonnet pricing by Q3. Builders who passed on certain agentic workflows because of agent-loop-cost constraints should revisit those decisions on a quarterly cadence, not an annual one.

The competitive pressure matters equally. If every product’s baseline model improves every quarter, holding a competitive advantage through model access alone becomes harder each cycle. The moat shifts from “we use the best model” to “we use the best model better,” meaning workflow design, context engineering, and domain-specific integration matter more as raw capability becomes commodity. The labs are compressing the capability gap between tiers. Operators who build on model access rather than workflow depth are exposed.

The Cost / Tradeoff

Capability compression does not mean all models are equal. Flagship models retain meaningful advantages in the hardest reasoning tasks, the longest context windows, sustained coherence over very large inputs, and the highest-stakes safety scenarios. When Sonnet 5 closes the gap with Opus 4.8 on most agentic tasks, “most” is the operative word. For legal document analysis at full context length, scientific reasoning requiring deep chain-of-thought, or tasks with very low tolerance for any error, the flagship still outperforms.

The tradeoff is one of selectivity. When midtier capability handles 80% of tasks reliably, the remaining 20% that require the flagship becomes a meaningful routing decision rather than a default. Builders who send everything to the top model pay a tax on capability they do not need. Builders who send nothing to the top model leave performance on the table where it counts. The practical skill is knowing which tasks are in which 20%.

How TWO Uses It

At TWO, capability compression is a routing signal, not a budget story. When a new Sonnet ships, the first step is reassessing which workflows were running on Opus by habit rather than requirement. The question is concrete: does this task require the flagship’s reasoning depth, or was it routed there because the midtier could not handle it last quarter?

Scott’s Take: The quarterly capability recheck is the new cost review. Any model-routing decision you set and forgot is probably wrong by now.

The deeper operator implication is that product differentiation built on model access erodes faster than founders expect. When Sonnet outperforms last year’s Opus, the advantage disappears unless the product compounds something the model alone cannot provide: proprietary data, workflow depth, institutional knowledge, or trust with a specific user base. Capability compression accelerates the pressure to build on something more durable than which model you chose this quarter.

A Concrete Operator Scenario

You run a customer-service pipeline on Claude Opus 4.8, paying $15 per million output tokens. Claude Sonnet 5 ships. The right move is not to switch immediately. It is to run your eval suite on both models against a sample of real tickets: routine inquiries, complex multi-turn complaints, escalations requiring nuanced judgment. On routine tickets, Sonnet 5 scores at parity. On complex escalations, Opus still outperforms by a meaningful margin.

Now the routing decision is specific rather than categorical: route routine volume to Sonnet 5, flag complex escalations to Opus. Effective cost per resolved ticket drops sharply on the routed volume. The evaluation took one afternoon. The savings recur every month. That is what capability compression actually looks like as an operator decision: not a wholesale switch, but a routing refinement triggered by a quarterly recheck you build into your calendar.