xAI just caught up (Grok 4.6 is here)

Tech/product overview: xAI Grok 4.6 (“Grok 46”)

• Positioning/goal: Grok 4.6 is described as an update to Grok 4.5, with a focus on long-running agents and more ambitious interactive/visual work (e.g., tasks across many steps).
• Training approach: Not a new pre-training—a new post-training run on top of Grok 4.5. The video claims this post-training includes:
  • Cursor-related RL/post-training contributions (from xAI acquiring Cursor).
  • Improved optimizer/training recipe, better SFT/RL pipeline foundation.
  • Regenerating SFT trajectories using Grok 4.5, then filtering problematic traces via model-based checks.
  • Training across agentic RL tasks for knowledge work, coding, and domain-specific environments (e.g., kernel optimization, webdev, computer-aided design).
• Reliability for agents: Emphasis that Grok 4.6 stays on track in long tasks and can coordinate multiple sub-agents without “losing what it’s supposed to do.”

Benchmarks & claims (with caveats)

• The speaker cites Artificial Analysis “Intelligence Index” results:
  • Grok 4.6 score ~61, described as neck-and-neck with “56 Soul”, with other frontier models nearby.
  • A claimed improvement vs Grok 4.5: +5 points over a short period.
• Important caution: The speaker argues these benchmarks are not fully trustworthy for real-world usefulness.
  • Example: they say Opus 5 “looks good day one” but becomes worse in ongoing code-merging work.
• Examples of benchmark discussion:
  • “Huge bump” mentioned for agentic reasoning/technical evaluations (including claims like ~54% to ~66 in one benchmark set).
  • Cursor-bench details: claims of potential contamination/overlap (“accidentally trained some… data”), complicating fair comparison.

Cost/performance (“Pareto frontier”) details

• Strong emphasis on cost per use and “speed vs intelligence” tradeoffs:
  • Pricing cited: $2 per million input tokens and $6 per million output tokens (framed as much cheaper than Opus 5).
  • Claimed task cost: ~84 cents per task (vs higher costs for other models).
• However, the reviewer says token efficiency regressed compared to Grok 4.5:
  • More tokens per run (+30%+), making runs slower/more expensive in practice even if raw $/token is similar.
• Overall impression:
  • Grok 4.6 remains a strong “frontier-ish” model for intelligence/agentic ability, but it no longer holds the “uniquely fast + uniquely cheap” advantage the reviewer liked in Grok 4.5.

Official tooling: “Grok Build CLI” and “Cursor”

• The reviewer highlights the official Grok Build CLI as a key workflow component and uses it heavily during testing.
• Availability:
  • Grok 4.6 is said to be available in Cursor and Grok Build.
• Intro offer:
  • 2x usage for the first week (per the video).

Safety/security evaluation and audits

• The reviewer credits Grok 4.6 with strong security audit behavior, including:
  • Finding issues and advising on endpoint locking, trust/safety UI, and launch-readiness steps.
  • Running a security audit in a “real codebase” scenario.
• Mentioned safeguards and expanded testing:
  • Capability testing, safeguard calibration, third-party testing
  • Evaluation loops like model checking its own work on longer runs.

Real-world use cases tested

1) App/game generation experiments (“Fish Slop”)

• 2D/straightforward gameplay port: Grok 4.6 produced working-ish results, but with gameplay/controls issues (sizing, movement feel, pacing).
• 3D port:
  • Claimed to fail outright at first (black screen)
  • Then attempted a fix, but produced serious 3D direction/placement errors and poor visual pass quality.
• The reviewer contrasts this with other models:
  • Some cheaper/open-weight models reportedly produced more functional 3D results.
• Conclusion:
  • Grok 4.6 is not top-tier for 3D game port quality; it feels last-generation-like compared with faster-moving competitors.

2) Engineering work with “T3 Code” / Cursor integration

• The reviewer tested Grok 4.6 against T3 Code’s implementation of Cursor bindings.
  • Specifically, they were using an outdated adapter (“ACP”) for the Cursor CLI and wanted migration to the Cursor SDK.
• Grok 4.6 performed:
  • An audit of gaps between ACP and the Cursor SDK contract.
  • Proposed migration steps and surfaced real integration constraints, including:
    • SDK contract differences: limitations around approve/deny UI behavior, login flows, and control-plane interactions.
    • Missing behaviors the model expected from ACP, requiring engineering workarounds.
• Iterative process described:
  • The model generated/updated plan files, produced a PR, then was “babysit” with follow-up changes.
• UI/event handling quirks:
  • Mentions “events are broken” in parts of the Grok Build integration.

Reviewer’s scoring/qualitative conclusions

• The reviewer provides a category-by-category comparison vs Fable and Soul, then rates Grok 4.6:
  • Cost: viewed as a regression due to worse token efficiency / longer runs.
  • Intelligence: seen as a meaningful bump.
  • Speed: a notable regression (partly from more tokens generated).
  • Thoroughness: slightly improved, but still sometimes misses.
  • Orchestration/sub-agents: decent improvements, but not enough to offset speed/cost regression.
• Final take:
  • Grok 4.6 is promising (especially for agentic reliability and intelligence growth), but the reviewer says they’re less likely to use it heavily than Grok 4.5.
  • They’re more excited for Grok 4.7 and expect Grok to “catch up fast.”

Sponsors / aside (MX video infrastructure)

• The video includes a sponsor segment about MX (robots product) for video app infrastructure:
  • Automations for captioning, moderation, summaries, and asking questions about videos.
  • Designed to avoid building the complex video-processing pipeline for user uploads.
  • Used/endorsed by companies mentioned in the ad (examples listed include brands with Cursor/Dropbox-like framing).

Main speakers / sources (as stated or implied)

• Speaker/reviewer: the channel narrator (mentions being paid for X/Twitter usage; references their own testing and ratings).
• Model/system sources referenced: xAI Grok (4.5, 4.6, 4.7), Cursor, Artificial Analysis Intelligence Index, and other models such as:
  • “GPT/OpenAI Opus 5,” “Fable 5,” and “Soul” (referred to as 56 Soul)
  • “Gemini 3.6 Flash/Terra” and “Kimmy K3” in comparisons.