MiniMax Switches to Perplexity and Cuts Agent Costs by 27%

MiniMax Agent has quietly made a decision that reveals something important about how agentic AI systems actually work in production: search quality is not a nice-to-have, it is the core cost driver. The team benchmarked three AI-native search providers across more than 700 agent tasks, and the results were decisive enough to switch their default search provider from Serper to Perplexity.

The Numbers That Made the Call

The benchmark compared MiniMax Agent's previous default, Serper, against Perplexity Search across real agent workflows. Serper is a classic SERP API -- it returns raw Google search results cheaply and quickly. Serper is one of the cheapest options for raw Google SERP data at $1 per 1K requests on the Starter plan, or $0.30 per 1K on Ultimate. But cheap per-query pricing can be deceptive when agents loop.

The MiniMax team's results across 700+ tasks told a clear story:

• Tool calls per task: 17.8 (Perplexity) vs. 32.6 (Serper) -- a 45% reduction
• Token usage: 94.6M vs. 162.3M -- a 42% reduction
• Pass rate: +2% improvement with Perplexity
• Total cost: 27% decrease end-to-end

The key insight: fewer, better searches meant the agent needed less context to arrive at a correct answer. One good search replaced nearly two dozen bad ones.

Why Search Is Different Inside an Agent

The distinction between a search API built for humans and one built for agents is fundamental. The fundamental question is: when your agent calls this API, what does it get? A synthesized answer with citations, or raw links it has to process itself? Full page content, or just snippets? The difference determines your entire agent architecture.

Perplexity changes the fundamental contract. Other search APIs return results -- documents, URLs, snippets. Perplexity returns a synthesized answer with citations. For an agent running in a loop, this matters enormously. Every irrelevant token in a search result is context the model has to wade through before it can decide whether to search again.

Perplexity has been investing heavily in exactly this problem. The most significant investment has been in snippet quality, optimized along two dimensions: relevance and size. Returning the right content in the right amount directly determines downstream answer accuracy and token efficiency. In practice, these improvements let them generate smaller snippets that are more relevant to the query. Their self-improving content understanding pipeline now handles a broader range of structured data formats, including tables, nested lists, and dynamically rendered content.

The Architecture Behind Perplexity's Search API

Perplexity's Search API uses low-latency hybrid search, combining semantic methods, LLM ranking, and human feedback. It offers fine-grained content understanding, segmenting documents into sub-units for targeted retrieval. This approach reduces preprocessing needs and accelerates integration into AI pipelines. The result is that each API call returns denser, more relevant evidence -- which is exactly what an agent needs to avoid spinning up another search call.

The Agent API orchestrates multi-step workflows where the model can search the web, reason about results, call additional tools, and verify information across multiple turns. This makes it suitable for complex queries that require synthesis from multiple sources. Developers can access OpenAI, Anthropic, Google, xAI, and more through one unified API, without needing to manage multiple API keys.

The Trade-offs Are Real

Perplexity is not a universal upgrade. The same synthesis that makes it powerful for agent grounding creates specific failure modes worth knowing about. Synthesis can hallucinate -- the answer looks authoritative but may contain errors the agent cannot detect. Rate limits are lower than Serper or Exa at equivalent price points. It is not suitable for agents that need raw web data to make their own inferences, since the synthesis happens inside Perplexity, not inside your agent.

Perplexity's 50 RPM is the most restrictive rate limit in the category, suitable only for sequential agent workflows. If your architecture relies on parallelizing many search calls simultaneously, this becomes a hard constraint. MiniMax Agent's workflow is sequential enough that the quality gains outweighed the throughput ceiling.

What This Means for the Agentic Search Market

This benchmark is a data point in a broader shift happening across the industry. Within Perplexity's own systems, single tasks have been seen to invoke hundreds or even thousands of retrieval operations within a few minutes -- a workflow that is impossible for humans but absolutely natural for agents. Search providers are now competing not on queries-per-second for human users, but on evidence density per token for machine consumers.

MiniMax itself is moving fast on the model side. MiniMax M2.7 adds enhancements to the popular MiniMax M2.5 model, built for agentic harnesses and complex use cases in fields such as reasoning, ML research workflows, software engineering, and office work. On May 27, 2026, MiniMax announced an overall upgrade of MiniMax Agent and renamed the upgraded agent Mavis. The Perplexity integration is shipping as part of this broader platform push.

For teams building their own agent stacks, the MiniMax benchmark offers a useful framework: do not optimize search cost per query in isolation. Measure tool calls per task, tokens consumed per task, and pass rate together. A search provider that costs more per call but cuts your total call volume by 45% is almost always the better deal -- and the MiniMax numbers show that gap can be substantial enough to matter at scale.

The Perplexity API Platform is available now with pay-as-you-go pricing, and the Agent API supports models from OpenAI, Anthropic, Google, and xAI through a single endpoint at POST https://api.perplexity.ai/v1/agent. The Search API improvements including the new span-labeling pipeline for snippet quality are already live.