Perplexity’s expanded Computer system routes tasks across 20+ frontier models and pitches itself as the anti-lock-in layer — a direct structural bet against every lab trying to own the stack.
Perplexity announced a significant expansion of its Computer system this week, broadening the roster of models that can act as orchestrators in agentic workflows. According to Perplexity, the system now supports Anthropic’s Opus and Sonnet, xAI’s Grok, Zhipu’s GLM, and OpenAI’s GPT-5.5 as top-level orchestrators — with smaller and multimodal models fanned out as subagents underneath. The design is explicitly framed by Perplexity as insulation against single-vendor lock-in and the performance volatility that comes from betting on one lab’s roadmap.
The headline feature is a “Model Council” — Perplexity’s term for a mode in which multiple models run in parallel on a query and surface distinct answers, letting users (or the system itself) arbitrate between them. Corroborating coverage from MarkTechPost confirms the 20+-model routing claim and the multi-orchestrator architecture. Perplexity is also teasing on-device local runtimes as a near-term addition, which would let the system execute tasks without a cloud round-trip — a move with real implications for enterprise privacy and latency.
It’s worth being precise about what this is and isn’t: the capability and model-list claims come from Perplexity’s own announcement. Whether its orchestration is objectively superior to alternatives is unverified. And “model agnosticism” is a strategic positioning choice, not a neutral stance — Perplexity still captures the user relationship and, critically, the routing data that accumulates as it picks models across millions of tasks.
The key insight: Perplexity is not building a better model. It is building the layer that decides which model wins each task — and accumulating the routing signal to make that decision smarter over time. That is a fundamentally different business than any of the labs it routes through.
The Orchestration Race — Key Moments
2024 — Perplexity launches first agentic search features
Perplexity begins positioning beyond search, adding tool use and early agent workflows to its answer engine.
Early 2025 — Labs race to close their APIs
Meta, OpenAI, and xAI each move to tighten distribution control — own the model, own the surface, force adoption through proprietary APIs and closed ecosystems.
Mid-2025 — Frontier model performance begins converging
Benchmark leapfrogging accelerates. No single lab holds top position across all task types for more than weeks at a time. The routing thesis gains structural credibility.
July 2026 — Perplexity Computer expands to 20+ models
Opus, Sonnet, Grok, GLM, GPT-5.5 as orchestrators. Model Council. On-device runtime teased. The anti-lock-in pitch goes explicit.
The Structural Read
There are two coherent theories of where durable value accretes in the AI stack right now. The first — call it the verticalization thesis — says the winning move is to own the model, force it through proprietary distribution, and make switching costs as high as possible. Meta’s closed pivot on Muse Spark is a clean example: own the creative model, own the API, own the surface. OpenAI and xAI are running the same playbook in different verticals. The logic is straightforward — if you own the model that everyone has to use, you capture the economics of AI adoption directly.
Perplexity Computer is the mirror image. Its thesis is that frontier models will keep leapfrogging each other — which is already empirically true — making any single model bet a volatile proposition for enterprise buyers. If no lab holds pole position for more than a quarter, the durable asset is not the model. It’s the router that picks the right model per task, accumulates signal about which model is best for which workload, and sits between the user and every lab simultaneously. The enterprise buyer never has to marry one lab; Perplexity captures the relationship regardless of which model wins next month’s benchmark.
This maps directly onto the orchestration layer as a control point in the Map of AI framework — the argument that whoever owns the workflow coordination layer above the models can extract value from the entire stack beneath it. Perplexity is explicitly racing to occupy that position. But the defensibility risk is real and worth naming clearly.
The Routing Paradigm — Defensibility Risk
Routers are powerful but capturable from above
A model maker that bundles its own orchestration layer can bypass the independent router entirely. OpenAI’s operator framework, Anthropic’s tool use, and Google’s Gemini ecosystem are all partial attempts to own orchestration from within the model layer. Perplexity’s answer to this threat is exactly what it announced: on-device local runtimes (reducing cloud dependency), proprietary workflow data (making routing smarter than any single-lab alternative), and the Model Council interface (making multi-model comparison a user-facing habit). Without all three, agnostic routing is a feature, not a moat. With all three, it starts to look like a platform.
The on-device runtime announcement is the most strategically underappreciated element here. Local execution means Perplexity can run subagent tasks without a cloud round-trip — which matters for enterprise latency, data residency requirements, and, critically, reducing its own dependency on any lab’s API availability. It also means the routing layer becomes partly hardware-resident, which is a meaningful barrier to replication. This is not yet shipped; Perplexity is teasing it. But the direction is clear: own enough of the execution surface that no single lab can bundle you away.
Three Implications
IMPLICATION 1 — Enterprise Procurement Shifts
If Perplexity’s framing lands with enterprise buyers, AI procurement stops being a model-selection decision and becomes an orchestration-platform decision. The question is no longer “which model do we standardize on?” but “which router do we trust to pick the right model?” That is a fundamentally different sales motion — and a larger contract size — than per-token API consumption. Labs that assumed enterprise stickiness would come from model quality may find the stickiness accretes one layer up.
IMPLICATION 2 — The Routing Data Flywheel
Every task routed through Perplexity Computer generates a data point: which model was selected, for what task type, with what outcome. At scale, that dataset is more valuable than any single benchmark. A router that has processed billions of real-world task-model pairings can make better routing decisions than any lab’s internal test suite — and that advantage compounds. This is the hidden asset in Perplexity’s agnostic positioning: it is not neutral; it is accumulating the signal that makes neutrality into a proprietary capability.
IMPLICATION 3 — Labs Face a Two-Front War
Labs pursuing verticalization now have to compete not just on model quality but on whether their bundled orchestration is better than an independent router with access to all their competitors’ models simultaneously. Perplexity’s Model Council makes that comparison explicit to the user. A lab that wins on model quality but loses on orchestration UX may still lose the enterpriserelationship. The pressure to ship credible orchestration layers accelerates — which paradoxically may further commoditize the model layer itself, strengthening the router’s position.
Business Engineer Framework
The Map of AI Redrawn — The Orchestration Layer as a Control Point
The Map of AI framework maps 200+ companies across nine layers of the AI stack. Perplexity Computer is a direct play for Layer 6 — workflow orchestration — the layer that sits above every model and below every end-user application. The framework explains why control at that layer can be more durable than model ownership, and what it takes to defend it. Read the full structural analysis to understand which companies are winning at each layer — and which are about to be routed around.
Perplexity is making the opposite bet to every major lab this week: don’t own the model, own the decision about which model to use — and make that decision smarter, faster, and more local with every task routed. It’s a coherent thesis with real structural logic behind it, and the on-device runtime move suggests Perplexity understands exactly where the defensibility gap is. The risk is that every lab it routes through is also building an orchestration layer, and the race to bundle from above is already on. Whether agnostic routing becomes a durable platform or gets absorbed
91,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.
Gennaro is the creator of FourWeekMBA, which reached about four million business people, comprising C-level executives, investors, analysts, product managers, and aspiring digital entrepreneurs in 2022 alone | He is also Director of Sales for a high-tech scaleup in the AI Industry | In 2012, Gennaro earned an International MBA with emphasis on Corporate Finance and Business Strategy.
Scroll to Top
Discover more from FourWeekMBA
Subscribe now to keep reading and get access to the full archive.
