Microsoft's Web IQ turns Bing into a sub-165ms engine for AI agents
Microsoft just launched Web IQ, a specialized suite of grounding APIs that connects AI agents to real-time web data. Instead of dumping full documents, it serves passage-level structured evidence to slash token consumption, delivering sub-165ms p95 latency that is roughly 2.5 times faster than competitors. It gives multi-step reasoning agents a high-speed factual backbone without blowing their context budgets.
Microsoft just launched Web IQ, a new suite of AI-native grounding APIs designed specifically to act as a search engine for autonomous AI agents.
While traditional search APIs are built for humans browsing full pages, agentic workflows retrieve data repeatedly during multi-step reasoning. Web IQ is built on top of Bing's global index but completely re-architected from indexing to ranking, returning structured, passage-level evidence instead of full documents.
Speed, passages, and token density
Feeding entire HTML documents into an LLM context window burns tokens and slows reasoning down. Web IQ tackles this by stripping away the fluff and returning compact passage-level evidence across web pages, news, images, and video.
Under the hood, Microsoft tapped into its DiskANN distributed vector search work to keep latency remarkably low. Tested across five global data centers without cache hits, Web IQ clocked a sub-165ms p95 latency—which Microsoft claims is roughly 2.5x faster than its closest competitors.
On quality, Microsoft evaluated the system on 3,000 blind production queries using GDSAT (grounding satisfaction), a metric tracking completeness, freshness, and authority, where it outranked competing alternatives. Microsoft also stressed that Web IQ respects standard robots protocols and publisher controls.
Why it matters
We think this is one of Microsoft's smartest architectural pivots in recent months. Until now, most developers building agentic workflows have treated web search as an afterthought, duct-taping traditional search APIs or fragile scrapers to RAG pipelines.
That setup is slow, expensive, and constantly hits context window limits. By bringing p95 latency down under 165ms, Web IQ fundamentally changes the math for multi-step reasoning: agents can now afford to search multiple times in a single loop without users giving up on the latency.
The passage-level extraction also cuts token burn right where it hurts. The potential tradeoff is accuracy; trusting an external orchestration layer to pick the right excerpts means downstream models can't cross-check the broader document if the retrieval model misses context. Still, for developers building latency-sensitive agents, Web IQ looks like a massive step forward.
Sources (1)
- blogs.bing.com(opens in a new tab)Originalblogs.bing.com