Kenhuang Academy
  • English

      • Models

      Cursor wires many vendor models into one Agent loop. The list, context limits, hidden flags, and prices can change weekly. This chapter is about how to choose. Treat in-app Models & Pricing and the table at the top of cursor.com/docs as the source of truth—not the names in this tutorial as an SLA.

      Where to see “what exists right now”

      1. The model picker next to the Agent input (some builds cycle with Ctrl+/ / Cmd+/).
      2. In-app Models & Pricing (usage pools, plans, Max Mode).
      3. The model table on the docs homepage (provider, default / max context, capabilities, notes).

      Many models are Hidden by default until you enable them. Some frontier models need Max Mode on legacy request-based plans. Long context (for example past ~200k / 272k tokens) is often surcharged; the docs call that out per row.

      Families you will keep seeing (2026)

      These are families that show up in the docs table, not a complete menu.

      FamilyProvider (as docs label it)Typical fit
      Composer 2.5CursorDaily Agent; the table lists Agent / Thinking / Images
      Claude Opus / Sonnet / FableAnthropicHeavy reasoning, long context, vision; Fable has retention / guardrail notes
      GPT-5.6 (plus Codex / Mini / Fast siblings)OpenAIAgent + reasoning; long context and Fast tiers usually cost more
      Gemini 3.xGoogleLong context, multimodal; some preview SKUs are image-first
      Grok 4.xLabeled Cursor (jointly trained with SpaceXAI)Agent + Thinking

      The same families also ship Haiku, Flash, Mini, Nano, Luna/Sol/Terra, Codex-specific SKUs, and more. Open the picker and Pricing—do not memorize this page.

      Auto versus a named model

      ChoiceMeaningWhen
      Auto (label follows your UI)Cursor picks / routes for the taskSmall edits, exploration, you do not want to manage SKUs
      Named modelYou pin a concrete IDRepro, comparisons, team policy (“Sonnet / Composer for this class of PR”)

      Auto is convenient but bad for “same model as last week’s plan.” Put this in a rule or PR template:

      For refactors that touch more than five files, use a named frontier
model (not Auto) and start in Plan Mode.

      The CLI accepts an explicit model (IDs follow current CLI help):

      agent -p "find and fix performance issues" --model "gpt-5"

      Stale IDs fail or get remapped—trust agent’s model list / /model.

      How to choose (practical)

      There is no single right model. Match task shape:

      TaskPreferWhy
      Tab-level complete, one-line editDo not pick an Agent modelTab has its own completion stack
      Explain a repo, small bugAuto or a fast / cheap named modelLatency beats max reasoning
      Multi-file feature, architectureA frontier SKU marked Agent + ThinkingWorth Plan Mode first
      Screenshots / UI / diagramsA model with ImagesOtherwise visual context is dropped
      Huge repo, long logsCheck Max context; watch over-threshold pricingNarrow the dump before you paste
      Image assetsAn image-generation model; output defaults to assets/Not the same as “a chat model that can see pictures”

      Thinking / reasoning tiers are slower and often cost more (some legacy plans count Thinking as two requests). Stay on the default effort until a plan stalls, then raise it.

      Pricing caveats (they expire)

      Traps that show up in the docs table:

      • Hidden by default: a model you never enabled will not appear in the daily picker.
      • Max Mode: required for some Opus / GPT-5.x / Gemini long-context combinations.
      • Long-context multipliers: some SKUs 2x input past 200k; others keep short-context rates up to 1M.
      • Cache: read/write multipliers differ (e.g. some Claude cache writes 1.25x, reads 0.1x).
      • Preview / joint models: tighter rate limits; behavior may change before GA.
      • Retention: some Fable setups need enterprise approval and state data is not used for training—read the model note and legal pages.

      Individuals watch usage pools; teams watch admin disables. Do not treat dollar figures in a tutorial as a quote.

      Privacy and enterprise

      Picking a model is also picking where data goes. Privacy mode, Team / Enterprise contracts, and regional endpoints (the docs mark a surcharge for some Claude Bedrock / Vertex routes) limit what you can actually select. Ask an admin if unsure; “it is in the picker” does not mean “paste customer data.”

      Working defaults

      1. Day to day: Auto or the team’s named daily model (many personal workflows use Composer 2.5).
      2. Multi-file or unclear requirements: named frontier model + Plan Mode.
      3. Bug repro: write the model name in the chat so you can compare later.
      4. Surprise bill: open Models & Pricing and check Max Mode, long context, and Thinking multipliers.

      Next

      评论