Research summary · Josh D / Researcher

Grok 4.6 – A field guide

2026-08-12 Primary: Eric Zakariasson X Article Official: cursor.com/blog + x.ai/news
Verification loop Short prompts Sync > async 4.5 vs 4.6 demos $2 / $6 per 1M

Bottom line

Eric’s take after weeks as daily driver: communication + speed stand out more than any single capability jump. Highest leverage isn’t “work harder” — it’s define done + verify in a loop (especially with browser use). Prefer short prompts with a clear preference unless you already have a detailed spec.

For Josh: treat this as practitioner guidance for the Grok 4.6 model in Cursor — distinct from the Grok Bot product already in the wiki.

Highest-leverage prompt

Verify the function and design after implementation, and keep on iterating and verifying until it's production ready.

On a spreadsheet app, a two-page spec vs three sentences produced nearly identical apps. Adding that one line made the model open the app, click real paths, check nested formulas, and fix failures. Same idea for 3D: capture frame → list what’s wrong → fix only those.

Key tips

Communication

  • Dense summaries (not task echo)
  • Quiet on small edits; narrates multi-file work
  • Running updates good enough to interrupt

Speed → sync

  • Fast + smarter → small synchronous loops
  • Async ships more but cold big diffs
  • Same session can stretch long-horizon

Prompting

  • “Work very hard” ≈ no effect
  • Long = specificity; short = model taste
  • 4.6 taste is good enough for short + preference
  • Say what done means

Where to steer

  • Web UI: self-verify works (DOM + shot)
  • 3D / video / physics: give a way to look
  • Or accept you’re the checker

Project experiments (4.5 vs 4.6)

Project4.6 edge (per Eric)
AoE2-style browser strategyIsometric 3D + HUD/minimap on first try vs flat 4.5 prototype
MSN MessengerMore polish (separate windows, winks)
Excalidraw presentation modeMore first-pass detail; verify still required
Board deckStructure/hierarchy vs number dump
SpreadsheetSpec length ≈ irrelevant; verification line decisive
Remotion launch film (X TS SDK)More compelling watch; video = widest model spread
EverydayConsole click-through API keys, QA, inbox triage, launch posts

Launch context (official)

  • Released 2026-08-12 with SpaceXAI — long-running agents + interactive/visual focus
  • Available: Cursor, Grok Build, API, OpenRouter, Vercel, Cloudflare
  • Pricing: $2 / 1M in, $6 / 1M out; fast variant ~2×
  • First week: 2× included usage in Cursor & Grok Build
  • Vendor headline: matches GPT-5.6 Sol on AA Intelligence Index (~61)
  • Secondary: docs/VB note 500k context; ≥200k prompts may bill higher — verify on console

Sources

  1. X Article — Grok 4.6 field guide (primary; API plain_text)
  2. Announcement post
  3. Cursor blog — Introducing Grok 4.6
  4. SpaceXAI — Introducing Grok 4.6
  5. Wiki: grok-4-6-field-guide.md, grok-4-6.md · raw/grok-4-6-field-guide-2026-08-12/