Kagi Assistant

Original Article ↗ Hacker News Discussion ↗

Kagi’s new paid “Assistant” feature, which layers models like GPT‑4 and Claude on top of its subscription search engine, is drawing interest as a way to combine high‑quality web results with LLM answers, lenses, and personalization. Many commenters like Kagi’s ad‑free model, search quality, and AI summaries, but are split on whether paying $25/month on top of other AI subscriptions is worthwhile compared with usage‑based services or direct API access. Others question Kagi’s growing focus on AI while core search still has issues such as occasional slowness, limited location‑aware results, and query term omission.

Pricing and Plans

  • Assistant is on the $25/mo “Ultimate” tier; not available to free users and currently separate from the $10/mo plan (with hints it may move down later).
  • Some see it as good value vs $20/mo ChatGPT alone, since it bundles Kagi search plus multiple models (OpenAI, Anthropic, others).
  • Others prefer usage-based or BYO-key setups (OpenRouter, Deepinfra, Raycast, various self‑hosted UIs) and view flat subscriptions as more expensive overall.
  • Family members can be upgraded to Ultimate individually; subscriptions are pro‑rated when upgrading/downgrading.

Assistant Capabilities and Search Integration

  • Key differentiator: tight integration with Kagi search, including:
    • Use of lenses (e.g., Programming, Forums, PDFs, Recipes, Small Web, custom domain allowlists).
    • Ability to constrain results via lenses or site: filters (e.g., only Microsoft docs, only GitHub/GitLab).
  • FastGPT/Quick Answer and web summarize are widely praised for:
    • Summarizing top results or specific pages.
    • Citing sources, which users often click through for detail/verification.
  • Current limits: no image generation; missing some direct-OpenAI features (multimodal, DALL‑E). Some note no markdown code blocks yet, making coding workflows clumsy.
  • Assistant reportedly has no hard usage cap for now, but Kagi says they’re monitoring.

Quality vs Other Tools

  • Some feel Kagi Assistant and FastGPT rival or replace separate ChatGPT/Claude subscriptions, especially for search‑grounded questions.
  • Others find Perplexity, SearchGPT, or using models directly (ChatGPT/Claude apps) more capable, particularly for complex multi‑step or large‑context tasks.
  • One thread notes Kagi’s own token/input limits and shallow retrieval (few sites, sometimes odd sources) as making it weak for large documents or codebases.

Search Experience and Performance

  • Many users strongly value: no ads, ability to block/uprank/downrank domains, lenses (especially “Forums”) and bangs; some say this alone justifies paying.
  • Others report Kagi search often matches DDG/Startpage/Google, or underperforms on:
    • Location‑aware queries.
    • “Newest X” style searches.
    • Queries where terms seem ignored or result sets are too small.
  • Performance is mixed: many report it as “fast or faster than Google”; others (notably in parts of Europe and US) see ~2s latency or occasional hangs.

AI Direction and Skepticism

  • Some view AI features as a natural, even essential, evolution of search.
  • Others worry Kagi is over‑focusing on AI while core search quality and reliability have slipped, leading a few to cancel.
  • Concerns include hallucinations and confidently wrong Quick Answers, misaligned citations, and a general desire for “just good search” without AI.