Perplexity launches Multiagent for Computer and Orchestrator, an API for models from OpenAI, Anthropic, Google and xAI

AI & Compute

Perplexity Bets on Model Mixing With 16-Agent Teams and One API Bill

Perplexity launched Multiagent, coordinating up to 16 agents on Claude, GPT and Gemini models, plus Orchestrator, a single-bill API spanning 85-plus models from four labs and open weights.

By
Nathan Brooks
Filed
Channel
AI & Compute
Read
3 min read

Perplexity launched two products on Nov. 17 that rest on a single wager: customers would rather mix models from rival labs than commit to one. Multiagent runs a team of up to 16 AI agents on different vendors' models; Orchestrator sells those models through one API and one bill.

The company points to the supply side as evidence. Labs such as Anthropic and OpenAI have released more than a dozen models this year, and more than 100 open-weight models have shipped, according to Perplexity's own announcement.

How does Multiagent work?

Multiagent sits above Ultra, the highest of the four effort settings Perplexity added to Computer in September. A lead agent running Claude Fable 5.2 plans each task and can assign work to as many as 16 agents. By default those agents run Claude Opus 5.5, GPT 6.1 Sol and Gemini 4 Carbon; Fable 5.2 and GPT 6 Astra join as supporting agents only when a user asks or the task demands it.

The architecture differs from Ultra in one structural way. Under Ultra, supporting agents report only to the lead. In Multiagent, agents share findings with one another, question assumptions and edit shared files, Perplexity says.

The target audience is finance and diligence work: parallel forecast scenarios, annual report reviews that flag inconsistent figures, data room diligence and year-end tax planning. Access requires a Perplexity Max subscription on the web. Perplexity placed its always-on Teammates agents behind the same tier earlier in November.

What does Orchestrator change for developers?

Orchestrator opens the model layer behind Perplexity's own products to outside developers. The company says that layer already routes hundreds of millions of requests across more than 85 models and more than a dozen providers.

The API supports OpenAI Chat Completions, OpenAI Responses and Anthropic Messages formats, so most teams can switch by changing the API key and base URL. Open-weight models, including GLM, Kimi and DeepSeek, are served either by Perplexity or by inference partners.

Pricing is the plain part. Perplexity charges no platform fee on top of credits, offers spend reports by model, project and key, and provides organization allowlists and spend caps. Large customers can negotiate volume pricing.

Can failover beat a single provider?

The routing layer tracks each provider's health, capacity and latency, and retries a failed request on another deployment of the same model. Perplexity claims that in one month of production traffic, Claude and GPT requests through Orchestrator posted higher success rates than any single provider it measured. The company did not publish the underlying figures, and the measurement is its own. Take the claim as a vendor benchmark, not an audit.

Two features remain in preview: Routes, plain-language rules that decide which model handles which request, and Auto effort.

Who is the competition?

Orchestrator competes with model routers such as OpenRouter and with the labs' direct APIs. The launch timing was not gentle. OpenAI released GPT-6.2 Astra and Sol on the same day and cut Astra's API price by 20 percent, tightening the margin story for anyone reselling frontier models.

A router with no platform fee lives or dies on reliability and billing simplicity rather than price. Perplexity has now published the product but not the success-rate data behind its central claim. The next test is whether enterprises move production traffic from OpenRouter and the labs themselves, or simply add one more key to the pile.

Source: mlq.ai

Share this article:

More from Nathan Brooks

Nathan Brooks

Show full bio

Senior reporter covering people, launches and day-to-day practice in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

1,056 articles

Related articles

« Previous articleNext article »