Perplexity Computer: Multi-Model AI Breaks Single-Vendor Lock-In

Perplexity launched Computer, a multi-model orchestration system that deploys tasks to 19 different AI models (Claude Opus, GPT-5.4, Gemini, and others) in a single agentic workflow. The CEO quoted Claude's biggest weakness: "It only coworks with Claude."

That is a deliberate hit at Anthropic's fortress strategy, and it signals something important: the single-vendor AI era is ending.

What Perplexity Computer does

Computer is a sandbox environment that runs autonomous tasks, routing each subtask to the best model for the job. If the task requires creative writing, it might use Claude. For mathematical reasoning, GPT-5.4. For structured information retrieval, Gemini. A single high-level request becomes a multi-model workflow orchestrated by Perplexity's logic.

The system runs for extended periods (days to months in sandboxed environments) with minimal human intervention. You pay consumption-based pricing and can hand-pick which models execute each task layer.

This is not a new idea. Model routing and multi-model pipelines have been theoretically possible for years. What is new is that Perplexity productized it and made it accessible at scale.

Why this breaks vendor lock-in

For the last two years, the dominant AI strategy has been single-vendor dependence: you commit to Claude or GPT, build your workflows around that model, and accept the vendor's roadmap, pricing, and strategic direction.

This made sense when model capabilities were highly differentiated. Claude was better at reasoning, GPT was better at coding, Gemini was better at multimodal tasks. Lock-in was a rational tradeoff for competitive advantage. As we covered in our analysis of OpenAI's latest benchmarks, the capability differentiation has collapsed.

GPT-5.4 hits 83% human parity across most job categories. Claude is equivalent on most tasks. Gemini is competitive. The frontier moved past "which model is best?" to "which model is cheapest for this specific subtask?"

Multi-model orchestration exploits that commodification. If GPT-5.4 costs 30% more per token than Claude, and they are equivalent on market analysis, route market analysis to Claude and save the margin. If Gemini's multimodal handling is 15% cheaper than Claude's, use Gemini for document processing.

Vendor lock-in assumes you are betting on a model's superior capability. When models converge on capability, lock-in becomes a liability instead of a feature.

The implications for procurement and partnerships

This shifts power back to customers.

Under single-vendor lock-in, you negotiated with one company. You accepted their terms, timeline, and pricing because switching costs were prohibitive. You built your entire stack around their model.

Under multi-model orchestration, you negotiate with multiple vendors for pieces of your workflow. You can threaten to route your most profitable tasks to a cheaper alternative. You can switch models without rebuilding your infrastructure.

For sales and enterprise teams, this is blunt: your lock-in advantage just declined. Customers can now commoditize your model within a broader workflow. Price pressure will increase. Margins will compress.

For customers, this is liberation. You can optimize each task for cost, speed, and reliability without being constrained by a single vendor's capabilities or pricing strategy.

What it means for Anthropic (and everyone else)

The CEO's quote was pointed: "Claude only coworks with Claude." This is half true. Claude works well with other models in a multi-model system. What does not work well with other models is Anthropic's broader strategy, which has been built on the idea that customers would commit deeply to a single vendor.

Perplexity is explicitly positioning itself as the anti-vendor lock-in platform. That is a credible differentiator if multi-model systems actually work, and early signals suggest they do.

For Anthropic, this requires a strategy shift: either compete on capability so overwhelming that single-vendor dependence is rational again, or pivot to being one model in a larger ecosystem, accepting lower margins on API consumption and compensating with volume.

For OpenAI, it is less of a threat. GPT-5.4 is flagship and expensive; it will be the preferred model for high-stakes tasks. But even OpenAI will lose margin on routine tasks that cheaper models can handle adequately.

For smaller models and open-source alternatives, this is an opportunity. If models are being routed based on cost and capability for specific subtasks, open-source alternatives become viable on price grounds alone.

The bigger picture

Perplexity Computer is the logical consequence of model commodification. When capability converges, users optimize for cost. When that happens at scale, multi-model orchestration is inevitable.

This does not mean vendors lose power. It means power concentrates differently. The vendor who controls the orchestration layer (Perplexity in this case) gains power. Individual model vendors lose leverage.

This is where the geopolitical dimension emerges. As Anthropic demonstrated with its recent distillation accusations, IP disputes will become more common as model value concentrates. Multi-model systems offer a legal hedge: if you do not depend on a single vendor's proprietary weights, you reduce your exposure to licensing disputes and capability theft.

For your business, the question is: what is your role in this ecosystem? Are you building the orchestration layer (you win as models commodify)? Are you selling a differentiated model (you lose if models converge)? Or are you using AI as a tool to build something else (you win either way)?

The single-vendor era bought you time. The multi-model era demands you choose what you are actually building.

---

Frequently Asked Questions

Q: If I use multi-model orchestration, do I need multiple API keys and vendor relationships?

A: Yes, initially. Perplexity handles the routing, but you still negotiate pricing with each model vendor. As the market matures, expect aggregators to emerge that simplify this process.

Q: Does multi-model orchestration reduce quality compared to single-vendor workflows?

A: Not necessarily. If you route each task to the best model for the job, you can exceed single-vendor quality while reducing cost. The risk is coordination complexity and latency if tasks are distributed across multiple vendors' APIs.

Q: Will Anthropic and OpenAI build their own multi-model systems?

A: Anthropic could, but it would undermine their single-vendor strategy. OpenAI has less incentive because GPT-5.4 is their flagship. Expect vendors to resist multi-model orchestration until margin pressure forces them to adapt.