Grok 4.6: Same Scale, Stronger Agents

TL;DR. SpaceXAI / xAI released Grok 4.6 on 12 August 2026. It is a post-training refresh of Grok 4.5 on the same ~1.5T-class foundation, not a documented parameter-count jump. The company says the work targeted long-running agents and stronger first-pass interactive and visual projects. Official High scores move the Artificial Analysis Intelligence Index from 56 to 61, matching GPT-5.6 Sol on that composite. Headline API price stays $2 / $6 per 1M input/output under 200k prompt tokens. Cached input rose to $0.50. The Starlight Model Arena has not run Grok 4.6.

What is externally sourced

These facts come from first-party or named independent pages retrieved on 14 August 2026.

• Release date and pitch. xAI's announcement dated 12 August 2026 describes Grok 4.6 as building on Grok 4.5 with a focus on long-running agents and more ambitious interactive and visual work. (x.ai/news/grok-4-6)
• Training recipe, vendor-stated. Longer supplemental training than 4.5; curated model-generated reasoning and engineering data; Grok 4.5 regenerated SFT trajectories; agentic RL on coding, knowledge work, kernel optimization, web development, and CAD. (same announcement)
• API contract. Model id grok-4.6. 500,000-token context. Text and image input, text output, no text output limit. Knowledge cutoff 1 February 2026. Reasoning effort: low, medium, high (default), and new xhigh. (docs.x.ai/developers/grok-4-6, release notes)
• Pricing, vendor-stated. $2 / $0.50 / $6 per 1M tokens (input / cached input / output) below 200k prompt tokens; $4 / $1 / $12 above. A fast variant is twice the price. (release notes)
• Official High scoreboard (vendor table). AA Intelligence 61 vs 4.5's 56; GDPVal-AA v2 1753 vs 1526; AA-Briefcase 1577 vs 1313; APEX-Agents 57.5% vs 47.1%; DeepSWE 1.1 65.9% vs 54%; Terminal-Bench 3.0 26% vs 15.7%; CursorBench 3.2 69.9% vs 66.7%. (announcement)
• Independent composite. Artificial Analysis reports 61 on the Intelligence Index, in line with GPT-5.6 Sol (max), behind Claude Opus 5 (63) and Claude Fable 5 (62). Measured cost per task $0.84. On AA-Briefcase, ~53 turns / ~0.5B input tokens versus ~103 turns / ~2.0B for Claude Opus 5 max. Cached-input price rose from $0.30 on 4.5 to $0.50. (Artificial Analysis, 12 August 2026)

What is my analysis

Grok 4.6 is an agent refresh, not a new brain. The useful question is not "did xAI leapfrog the frontier?" The useful question is: for a long-running coding or knowledge-work loop, does same-scale post-training plus unchanged $2/$6 list price change the routing table?

On published composites, yes at the margin. 61 on the AA Index puts it next to GPT-5.6 Sol and still behind the Claude Opus 5 family. The larger movement is on agent and office-style evals (GDPVal, Briefcase, APEX-Agents, DeepSWE) and on turn efficiency. If AA's Briefcase turn counts hold in your harness, the cost advantage is bigger than the token price list.

The 200k prompt-token cliff matters. Long agent loops that stay under 200k stay cheap. Loops that bloat context pay $4/$12. Cache hits are less discounted than on 4.5.

I am not upgrading vendor or AA tables into Starlight winners. The SIS Model Arena still dispatches through Claude Code Agent model overrides (fable, opus, sonnet, haiku). There is no grok-4.6 receipt in tools/arena/runs/ as of this writing. The last Grok-related receipt is a June 2026 Composer lane, not 4.6.

How was Grok 4.6 trained?

Only the vendor description is available. xAI says it did not move to a new public base. It ran a longer supplemental pass, regenerated SFT with Grok 4.5, filtered traces, then ran agentic RL in coding and design environments. That is consistent with a same-scale SFT+RL release. Parameter count is still not officially disclosed. Early 2T rumors were later treated as belonging to a later Grok 4.7 line in third-party writeups; I am not treating those as confirmed.

Grok 4.6 vs Grok 4.5 vs published flagships

All figures below are vendor-stated or Artificial Analysis. They are not Starlight Model Arena results.

Read the table as a routing hint, not a championship. DeepSWE and Terminal-Bench still favor Sol/Fable on the vendor table. Grok's published edge is the combination of 61-level composite intelligence, agent-eval movement, and mid-pack token price.

What the Starlight arenas will and will not claim

A later SIS round can include Grok only after the harness can pin a non-Claude model, run the same fixtures, and write a JSON receipt. Until then, any "Grok beat Opus on Starlight tasks" sentence is false.

What I tested

I did not run a Starlight Model Arena card against Grok 4.6. I did retrieve the official announcement, API docs, release notes, and the Artificial Analysis article on 14 August 2026, and I compared those numbers to the Grok 4.5 rows in the same vendor table. This session is also running on Grok 4.6, which is an operating fact, not a scored eval.

How to route it this week

1. Default xAI flagship for new Grok API work: grok-4.6, reasoning high unless you have a latency budget for medium.
2. Use xhigh only when a long agent loop is stalling. It is a cost and latency knob, not a personality.
3. Watch the 200k cliff and set a prompt cache key so cached input actually hits.
4. Keep Claude on SIS-native evals until a Grok receipt exists.
5. Do not recycle Grok 4.3 or 4.1 coding scores as 4.6 evidence.

Sources and method

Directed question: what actually changed from Grok 4.5 to 4.6, what is vendor-claimed, and what should a builder do before the Starlight arena can score it?

Sources retrieved 14 August 2026:

1. Introducing Grok 4.6 — official announcement
2. Grok 4.6 overview — API contract
3. xAI release notes, 12 August 2026
4. Artificial Analysis: Grok 4.6 benchmarks

Method: frankx.ai research phases — directed scan, source hierarchy, claim states, human publication decision. High-confidence claims here are limited to "the named page said X on the retrieval date." Cross-model gaps that mix vendor and independent runs stay directional.

Limitations: no SIS receipt; no first-party token-level bake-off in this repo; company naming (xAI / SpaceXAI) is used as the pages themselves use it; parameter count remains undisclosed.

FAQ

Is Grok 4.6 a new base model?

Not according to xAI. The announcement describes a longer supplemental run and regenerated SFT/RL on the Grok 4.5 foundation. Treat "2T" claims as unconfirmed unless xAI publishes them.

Does Grok 4.6 beat GPT-5.6 Sol?

On the Artificial Analysis Intelligence Index, the published score is a tie at 61. On several vendor coding/terminal rows, Sol still leads. "Beat" is the wrong verb.

Should I switch every agent to Grok 4.6?

Switch new xAI workloads and any 4.5 agent loops that were dying on stamina. Do not rip out Claude or Sol routes that already have measured receipts in your harness.

Did FrankX run a Grok 4.6 vs Claude battle?

No. The public Starlight Model Arena only publishes JSON receipts. None exist for Grok 4.6.

What is the API price?

$2 input / $6 output per million tokens below 200k prompt tokens; double above that. Cached input is $0.50. Confirm on the live pricing page before budgeting.

Related: Research brief · Frontier models hub · Model Arena · LLM Hub · Grok 4.3 analysis