Perplexity Quietly Wins the Search Wars for Agents

Perplexity Quietly Wins the Search Wars for Agents

Artificial Analysis launched a new Search Index on August 28, benchmarking 18 search APIs against how well they serve AI agents — not humans scrolling results [4]. Perplexity's medium-context variant scored 80, a 47-point jump over the 33-point baseline, beating Parallel Search's 75 by a wide margin. It also posted 87% accuracy on BrowseComp, the toughest benchmark in the set, while running at roughly $0.03 per task — the cheapest in the field [5].

This matters more than it looks. As agentic workflows scale, search stops being a UI feature and becomes infrastructure — the difference between an agent that hallucinates a fact and one that grounds its answer in something real. Perplexity beating dedicated agent-search startups on both accuracy and cost is a signal that the "search API for agents" category is consolidating fast, and Perplexity just claimed pole position.

@AravSrinivas was quick to highlight the result as best-in-class, and partnerships like the Decagon integration for live web search suggest Perplexity is positioning itself as the default retrieval layer for the agent economy, not just a chatbot alternative to Google [6].

A 23-Year-Old's AI Assistant Hits $2.5B — Before Shipping Publicly

Noah Shinn's Instinct, still in private beta, raised $250M at a $2.5B valuation this week — five times its valuation from just weeks earlier, bringing total funding to $350M [7][8]. The product: a text/call-based assistant that plugs into your apps and devices to handle tasks like trip planning or subscription management. There's no public revenue number. There's also no public product, really — just private beta access and viral word of mouth.

This is a useful gut-check for anyone tracking AI valuations right now. VC enthusiasm is running well ahead of demonstrated trust, and the permissions model behind an assistant that can act across your apps is exactly the kind of thing that invites scrutiny once it's actually in the wild. Forbes flagged the "VC obsession" angle directly, and the privacy concerns around broad device/app access are not a footnote — they're the whole risk profile of this product category [8].

The contrast worth sitting with: Instinct is being valued on vision and founder pedigree, while infrastructure plays like Perplexity's search benchmark win are being valued on measured performance. Both are legitimate bets, but they're different bets — and worth knowing which one you're actually making when you evaluate vendors.

EU AI Act: Transparency Rules Bite Now, High-Risk Rules Wait Until 2027

The EU's Digital Omnibus, in force since July 27, delays high-risk AI obligations (Annex III systems) to December 2027 for standalone systems and August 2028 for embedded ones — a real breathing-room extension for companies that were staring down a much tighter 2026 deadline [9]. But transparency rules under Article 50 are live as of August 2, 2026: chatbots must disclose they're AI, and AI-generated content — including deepfakes — needs machine-readable labeling [10].

Over 180 organizations have already signed the EU's Code of Practice, and new bans on non-consensual explicit content and CSAM-generating systems took effect alongside the transparency rules. Fines for non-compliance run up to €15 million, so this isn't a "wait and see" situation even with high-risk rules pushed back [9][10].

Nordic/EU: What the Omnibus Actually Buys You

For Nordic companies building AI products, the delay is genuine relief — you get until late 2027 to sort out high-risk classification and compliance instead of scrambling this year. But don't read "delay" as "ignore." Transparency and labeling requirements are enforced now, sandboxes are being expanded, and small mid-caps get SME-like simplifications, which is a meaningful concession for the Nordic startup scale that sits just above traditional SME thresholds [9].

The practical move: treat this as a runway extension, not a reprieve. Build the auditable-provenance habits now — labeling AI-generated content, disclosing chatbot interactions — because the infrastructure you build for transparency compliance today is the same infrastructure you'll need for high-risk compliance in 2027. Retrofitting governance is always more expensive than building it in from the start.