Anthropic
Sonnet 4.6 RECOMMENDED
~$0.06/q
Best default. Strong reading comprehension, reliable citation formatting, fast responses (3-5s). Handles both quick lookups and multi-source synthesis well.
Opus 4.6 DEEP
~$0.11/q
Deepest reasoning. Excels at complex multi-hop synthesis ("compare strategic implications across 5 topics"). Overkill for simple lookups. ~2x Sonnet cost.
Haiku 4.5 FAST
~$0.02/q
Fastest and cheapest Anthropic model. Good for quick factual questions. Weaker at nuanced synthesis and may miss citation rules on complex answers.
Sonnet 4.6 + Thinking THINKING
~$0.15/q
Same model, but reasons internally before answering (adaptive thinking). Better for complex questions where standard Sonnet gives shallow answers. Adds 5-15s latency. ~2-3x Sonnet cost due to extra reasoning tokens.
Google
Gemini 2.5 Flash VALUE
~$0.008/q
6-10x cheaper than Sonnet. Surprisingly capable for synthesis. Citation formatting may be less consistent. Great for high-volume exploration.
Gemini Flash + Thinking
~$0.015/q
Bumps reasoning effort to "medium." Marginal improvement for text synthesis — thinking helps more with math/logic tasks.
Gemini 2.5 Pro
~$0.03/q
Google's best model. Strong synthesis at a price point between Haiku and Sonnet. Good middle ground when Flash feels too loose.
OpenAI
GPT-5 Mini CHEAPEST
~$0.004/q
Cheapest option overall. Decent for simple lookups but weakest at following complex citation rules. You'll see more missed or inconsistent citations.
o4-mini (reasoning)
~$0.03/q
OpenAI's reasoning model. Always thinks. Strong analytical capability but reasoning overhead doesn't help much for text synthesis — better for math/logic.
How it works
Cost estimates assume a typical chat query: ~15K input tokens (30 source cards + system prompt + history) and ~1.2K output tokens. Prices shown as $input/$output per 1M tokens in the dropdown. The + suffix means thinking generates extra tokens at the same rate.
Strategy: Start with Sonnet 4.6 as your default. Switch to Gemini Flash for casual high-volume exploration. Reach for Opus 4.6 or Sonnet + Thinking when Sonnet's answer feels too shallow on a complex question.
Topic scope affects cost. More selected topics = more source cards in context = more input tokens. "All topics" sends up to 30 articles + 10 uploads. A single-topic filter sends fewer.
Citations: The model references your sources as numbered brackets like [1], [2]. Click any citation card below a response to open the original article. Models vary in citation accuracy — Sonnet and Opus are most reliable.