Musk Confirms Grok 4.7 Delay, Blames Overweight RL Penalty on Reply Length

On September 12, Elon Musk confirmed on X that xAI's new model Grok 4.7 is delayed, needing "a few more days" of polish. He unusually disclosed training details: the team likely set the RL penalty on reply length too heavily (or something similar), causing the model to give up prematurely on hard tasks it could actually solve, with insufficient rigor in self-checking its work.

Confirmed

As of September 13, Grok 4.7 remains unreleased; Musk said it needs "a few more days to cook"
Musk himself attributed the delay to an overweight RL penalty on reply length, premature abandonment of solvable problems, and weak self-verification
Per @MicahBerkley's account, Grok's official account confirmed the model was planned for imminent release but still needs days of RL tuning

Unconfirmed

Musk used hedging language like "possibly"; the exact details and severity of the training bug are not corroborated by official xAI technical documentation
No new exact release date has been announced

Why it matters

This is a rare case of a lab executive publicly attributing a model's behavioral flaw directly to its training mechanism, offering unusual first-hand insight into RL training trade-offs in frontier models
@nimaowji noted this is a training-level bug rather than a missing feature, meaning the fix requires re-tuning training rather than simply adding functionality, which explains the delayed launch

2026-09-12 ~ 2026-09-13 · 7 related posts

Primary sources

Grok 4.7 delayed as Musk says it needs 'a few more days' over RL training bug — nima_owji · 2026-09-13 07:31 (UTC+08:00)
xAI delays Grok 4.7 launch by days over overly punitive RL tuning, says Musk — emmanuelvivier · 2026-09-13 16:24 (UTC+08:00)

Musk delays Grok 4.7 by days, says RL over-penalized response length — inductionheads · 2026-09-12
Grok 4.7 delayed a few days as Musk blames RL length penalty for model giving up on hard tasks — minchoi · 2026-09-12
[source] Grok 4.7 delayed as Musk says it needs 'a few more days' over RL training bug — nima_owji · 2026-09-13
Grok 4.7 Delayed a Few Days for RL Tuning Over Premature Give-Ups — MicahBerkley · 2026-09-13
[source] xAI delays Grok 4.7 launch by days over overly punitive RL tuning, says Musk — emmanuelvivier · 2026-09-13

2 near-duplicate retellings: Scobleizer · rohanpaul_ai