Shipped · 2026-08-12

Grok 4.6
Back to the frontier, same $2/$6

SpaceXAI’s 2026-08-12 launch focuses on long-running agents and visual/interactive first passes. In Cursor, Grok Build, API, OpenRouter. First-week 2× included usage.

#Grok 4.6#AA 61#$2 / $6#500K ctx

Highlights

61

AA Intelligence

Ties Sol; behind Opus 5 max 63 and Fable 5 max 62.

$2 / $6

Input / output

Unchanged from 4.5. Cache hit rose $0.30 → $0.50. Fast variant is 2× price.

500K

Context

Same as 4.5 — not 1M.

65.9%

DeepSWE v1.1

Behind Sol 73% and Fable 70% on the official 4.6 table.

What 4.6 is

Grok 4.6 is a longer supplemental training run on top of 4.5 plus SFT data regenerated with 4.5 — not a fresh pre-train. xAI has published no parameter count for 4.6; the "2T" figure some early blogs used conflicts with later founder comments about 4.7, so this page does not state one.

Why it is trending

The most-shared figure is Artificial Analysis' efficiency comparison: on Briefcase, Grok 4.6 takes about 53 turns and 0.5B input tokens against Opus 5's ~103 turns and 2.0B. Note the two Terminal-Bench numbers are not comparable — xAI's launch table uses TB v3.0 (26%), while Artificial Analysis cites TB v2.1 (88.4%).

Timeline

2026-07-08/16

Grok 4.5 launch era.

2026-07-24/28

Musk teases 4.6 then 4.7.

2026-08-12

4.6 GA — a few days after the ~Aug 7 hint.

Confirmed vs caution

Confirmed

Launch post + eval table; AA article 2026-08-12; $2/$6; 500K; first-week 2× in Cursor/Build.

Caution / rumor

Three things get mixed up: xAI has not published a parameter count for 4.6, and the 2.1T figure circulating online comes from founder comments about 4.7; Terminal-Bench v3.0 and 2.1 are different scales and cannot share a column; and the fast variant costs 2×, it is not free speed.

How to choose

Pick 4.6

Long agents, visual apps, $2/$6 hedge vs Opus/Sol bills, Cursor/Build users.

Pick something else

Need 1M context, or max DeepSWE, or Claude/OpenAI-native stacks.

How to use it

console.x.ai / Cursor / x.ai/build. Model via /models.

On QCode

QCode does not currently carry the Grok family — there is no Grok model on /models. If you want an alternative in the same capability band, the catalogue does list claude-opus-5, gpt-5.6-sol, deepseek-v4-pro and kimi-k3.

FAQ

Is 4.6 out?

Yes. 2026-08-12.

Better than Opus 5?

AA 61 vs 63. Cheaper and fewer turns. Not a knockout.

Better than 4.5?

AA 56→61, much stronger long-agent story. Cache hit got more expensive.

Context 1M?

No. 500K.

Where is 4.7?

Not released. Separate tracker.

TB 26% looks terrible?

That is Terminal-Bench v3.0 on the launch table. AA’s TB v2.1 is 88.4%. Different tests.

Sources

x.ai/news/grok-4-6; AA 2026-08-12 article.

Use the Grok that exists

4.6 is the live id. 4.7 is a headline.