Grok 4.6 is Fable now

Tech/Model update: Grok 4.6 vs top frontier models

• Grok 4.6 release is reported as achieving benchmark parity / “neck-to-neck” with:
  • Fable 5 Max (described as an OpenAI frontier competitor in the transcript)
  • Anthropic’s best model (referred to as “Opus” / “Fable” family in the subtitles)

Key training delta vs Grok 4.5

• Much longer supplemental training run for Grok 4.6
• Both are described as variants of a 1.5T parameter base model (“V9 model” / “base”)
• Grok 4.6 uses more data—especially from Cursor—including curated model-generated data for:
  • reasoning
  • advanced technical concepts
  • high-quality engineering data
• Mentions:
  • an improved optimizer
  • an updated training recipe
• Claim: this strengthens later supervised fine-tuning and reinforcement learning stages

Benchmark skepticism acknowledged

The speaker notes that benchmarks can sometimes be “gamed,” but claims early hands-on results look strong rather than failing basic tests.

---

Hands-on demo claims (Portal 2 + interactive 3D)

• The speaker runs a Grok 4.6 build (with high settings; they later mention extra high exists).
• First test: prompt a “clone” of Portal 2, resulting in:
  • generation of game mechanics (blue/orange portals)
  • creation of 3D objects, a room, lighting/reflections, and puzzle logic “in one shot”
  • portal rendering/behavior like momentum conservation
• Multimodal-ish extras via external tooling:
  • the speaker mentions adding sound effects and voiceovers using ElevenLabs

“Chamber 07” segment

• Includes:

  • reflections + lighting
  • both gates alive
  • ability to walk through portals while preserving momentum

• Overall impression: Grok 4.6 feels like a high-end model (“a beast,” per the speaker).

---

Pricing + performance modes

Quoted pricing (subtitles)

• $2 per 1M input tokens
• $6 per 1M output tokens

“Fast” variant

• Described as twice the price, but also twice the speed / usage
• Roughly “evens out” for launch-week usage limits.

Launch-week availability detail

• Using Cursor or Grok Build during launch week yields double usage for the new models.
• Speaker advice: take advantage of this.

---

Tutorial-style automation concept: Cursor-driven “AI streamer” for games

The speaker describes an ongoing test to build a fully AI-automated run:

• Game: Pokemon Red
• System: a Pokemon Red emulator
• Controls: the AI connects to a harness to control the game
• Output: generates text-to-speech narration
• Avatars: a video avatar idea is mentioned to explain what’s happening
• Interaction: AI can talk to chat
• Safety: adds censorship / anti-jailbreak so it won’t get misused

Status

• The prototype can already:
  • load the game
  • pass initial screens
  • create components like fake chat and TTS voiceover
• Note: the system may still get stuck on logic later in the run.

---

Leaderboard-style evaluation: GPT-4o-style “GPTVal”

• Mentions GPTVal as a benchmark:
  • evaluates real-world job tasks/projects across industries, including:
    • engineering
    • AutoCAD
    • hospitality/travel agent work
    • finance
    • video editing
• Claim: Grok 4.6 beats Fable 5 and other competitors.
• Why it matters: positioned as more predictive of agentic work quality, aligning with Grok’s always-on agent product.

---

Grokbot: always-on agent (digital labor), teachable workflows

What Grokbot is

• Described as:
  • always on / always online / always available
  • an agent team rather than a traditional chatbot
• Available on:
  • desktop apps: Linux / Mac / Windows
  • Android coming soon

Architecture detail

• Each agent runs with its own cloud virtual machine
• Purpose: work 24/7 even if the user’s computer is off.

Role workflow (“chief of staff”)

• Speaker suggests creating a “chief of staff” role:
  • it delegates tasks to sub-agents automatically
  • user speaks to one point of contact while it coordinates others

“Teach a task” feature

• Grokbot includes a button to teach tasks by recording the user’s actions/steps.
• Motivation: avoid dangerous behavior like an agent accidentally sending messages in chat apps (speaker cites a prior incident with another model).
• “Teach a task” lets users define safe, repeatable skills for agent execution.

Example agents built

• X news / real-time X (Twitter) monitor
  • searches X for relevant posts
  • reports whether “Elon said something,” etc.
• Speaker attempts to determine whether Grok 4.6 is live on Grokbot:
  • no confirmation of the exact model, but “likely coming very soon”
• Agent can create routines like:
  • checking 3 times/day for 10 days, then notifying the user

---

UX/product polish highlighted vs competitors

• Speaker praises Grokbot UI patterns:
  • more frequent choice boxes / quick options to reduce cognitive load
  • at-a-glance answers enabling faster decisions (less reading/typing)
• Comparison note:
  • contrasts with “Fable 5,” described as having an odd speaking style or lots of rambling/memes
  • claims Grok feels faster and more direct in active use

---

Future roadmap speculation (Grok 4.7, Grok 5)

• Expected cadence/targets mentioned:
  • Grok 4.7 expected in 3–4 weeks
  • potentially a larger 2.1T parameter model (explicitly framed as rumor/speculation)
  • Grok 5 targeted before end of year
• Reason for rapid cadence:
  • speaker claims XAI reorganized engineering to ship major updates every 2–3 weeks
• Training data rumor:
  • mentions large amounts of SpaceX company data used for supplemental training for the next release (attributed to a future model described in the transcript)

---

Pricing/competition positioning (speaker’s take)

• Speaker argues Grok 4.6 pricing is lower than comparable frontier models
• Claims XAI + Cursor partnership may help it win
• Mentions “Google lagging slightly behind” as general competitive commentary

---

Main speakers / sources (as identified in subtitles)

• Speaker: Wes Roth (ends with “My name is Wes Roth”)
• Referenced organizations/products:
  • XAI
  • Anthropic
  • “Fable” (as named in subtitles)
  • OpenAI (implied)
  • Cursor
  • Grokbot
  • Grok Build
  • Grok 4.6 / 4.5 / 4.7 / 5
  • ElevenLabs
  • WhatsApp / Slack
  • X (Twitter)