Grok 4.6

Spec

Attribute	Value
Developer	xAI
Released	2026-08-07 (reported) — see ## Conflicting Reports, a second reported date of 2026-08-12
Announced	2026-07-18
Context window	500,000 tokens
Pricing	$2/M input · $6/M output; a request reaching 200K tokens re-prices in full at $4.00/M input · $1.00/M cached input · $12.00/M output
License	proprietary
Availability	API
**Three rows moved off unknown on 2026-08-16, and none of them from a
first-party page.** Context window, Pricing and Availability were still
unknown nine days after launch — recorded then as the finding rather than as a
gap in research. What changed is that the same specification now appears
consistently across several independent documentation write-ups, where on
2026-08-07 it appeared in none
(source). **Still no xAI
model card, pricing page or launch post has been read from this environment**, so
every value above is third-party and the underlying observation stands.

Availability is written as API on the narrowest thing the sources support — a published per-token rate card implies an API. No source read enumerates the surfaces (app, Cursor, X integration), so none are listed.

The 200K threshold is a cliff, not a tier: coverage states the entire request re-prices once it crosses, so the marginal cost of the 200,001st token is the difference applied to every token before it (source).

Release Date

• Released: 2026-08-07, as reported — one day ahead of the ~August 8 target Musk gave on July 25, and matching the "likely next week" of August 4. No first-party xAI page was read from this environment, so the date is carried on third-party reporting (source). Contested: four write-ups read on 2026-08-16 give 2026-08-12 instead — see ## Conflicting Reports
• Announced: July 18, 2026 (Elon Musk on X)
• Pre-training: expected to complete week of July 20, 2026 (Musk on X); as of July 23 xAI has not publicly confirmed completion, but the timeline implies pre-training is done or nearly done
• Current phase (as of 2026-07-23): likely in post-training — RL fine-tuning, RLHF, safety evaluations; no public update from xAI
• Launch: ~August 8, 2026; Musk stated on X (Jul 25) that 4.6 is "~2 weeks" out → ~Aug 8. Simultaneously confirmed Grok 4.7 (~4 weeks, ~Aug 22) — the first public confirmation of 4.7 as a distinct model. (source)

Benchmarks

Still no vendor benchmark table. Two figures circulate in launch coverage, both attributed to xAI by the outlet rather than sourced to an xAI page (source):

Measure	Reported	Comparison as reported
SWE Marathon	29.0%	Claude Opus 4.8 26.0%
Throughput	80 transactions/second	—
Per Eval Harness Configuration, a single-outlet figure with no harness
description and no vendor page behind it is recorded and not compared against this
wiki's other benchmark tables. Neither figure reappeared in anything read on
2026-08-16 (source).

First independent measurement — 2026-08-16

Grok 4.6 is absent from the 2026-08-09 Artificial Analysis snapshot and present in the 2026-08-16 one, which makes this the first independently produced figure for it that this repo holds (source):

Setting	Artificial Analysis Intelligence Index	Context Window	Cost per Task USD	Creator
Grok 4.6 · high	61	500k	$0.84	SpaceXAI
Artificial Analysis Intelligence Index is that publisher's composite of their
own benchmark suite, not an accuracy, and Cost per Task USD is their measured
cost of one task, not a per-token price. Creator reads SpaceXAI because that
is the value in their table; this wiki files the model under xAI.

Read against the same snapshot, 61 places it level with GPT-5.6 Sol (max) and Claude Opus 5 (high), one point behind Claude Fable 5 (with fallback) at 62, and two behind Claude Opus 5 (max/xhigh) at 63 — at $0.84 per task against Opus 5 (max)'s $2.34 and Fable 5's $3.14 (source). Its own predecessor Grok 4.5 sits at 56 for $0.36.

This is one measurement, not a benchmark table. It is a composite from a single vendor of evaluations, it says nothing about which tasks moved, and it does not substitute for the model card xAI has not published.

2026-08-23 — three more reasoning settings, and the spread is 9 points

The Sunday capture adds xhigh, medium and low rows beside the high row above (source):

Setting	Intelligence Index	Context Window	Cost per Task USD	Median Tokens/s
Grok 4.6 · high	61	500k	$0.84	69
Grok 4.6 · xhigh	60	500k	$1.04	62
Grok 4.6 · medium	59	500k	$0.67	73
Grok 4.6 · low	52	500k	$0.22	70
high scores above xhigh, at four fifths of the cost. That is the publisher's
table as captured and it is recorded as read — an inversion, not a typo this wiki can
correct, and one point is inside the range these composites move week to week
(high itself is unchanged at 61 from 2026-08-16). Nothing read states how the
settings map onto xAI's API, and xAI has still published no model card.

The high → low spread is 9 points for a 3.8× cost difference, which is the same shape as Qwen 3.8 27B's 43 → 52 in the same capture: on this leaderboard the reasoning setting is worth about as much as a model generation, and a row quoted without its setting names nothing.

Both tables above name the setting in the row label, and that is deliberate. scripts/claim-check.py normalises a row label by stripping anything in parentheses, so Grok 4.6 (high) and Grok 4.6 (low) are the same label to it — one silently overwrites the other, and the surviving value gets compared against whichever Grok 4.6 … row happens to be last in the cited snapshot. On 2026-08-23 that produced three reported conflicts on this page (61 against 52, $0.22 against $0.84, 70 against 69) where every figure on both sides was correct and only the pairing was wrong. Writing the setting outside parentheses makes each label distinct, so the check declines to compare these rows instead of comparing them incorrectly. It does mean the check no longer verifies this table — the snapshot's own labels still collapse to grok 4.6 — and that is the honest state: a row-label comparison cannot see a reasoning setting.

Use Cases

Documentation read on 2026-08-16 describes the model as built for long-running agents, agentic coding, and interactive or visual work, taking text and image in and producing text out, with low / medium / high / xhigh reasoning effort levels (source).

That is a positioning statement carried by third parties, not a capability measurement — no task-level evaluation behind any of it was read, and the effort levels mirror the pattern Claude Sonnet 5 and Claude Opus 5 already ship.

Compared To

Model	Developer	Params	Status
Grok 4.5	xAI	1.5T	Released 2026-07-08
Grok 4.6	xAI	2T announced / 1.5T reported	Released 2026-08-07 (reported)
Kimi K3	Moonshot	2.8T MoE	Released 2026-07-16
MiniMax M3 Pro	MiniMax	2.7T (unconfirmed)	Q3 2026 (unconfirmed)

Conflicting Reports

Parameter count. On 2026-07-18 Musk announced the model as 2 trillion parameters, "33% larger" than Grok 4.5's 1.5T, and this page recorded that figure (source). Launch coverage on 2026-08-07 instead describes 1.5T on the same "V9" foundation as Grok 4.5, with the gains coming from improved supervised fine-tuning and reinforcement learning rather than scale (source).

No source read here reconciles the two, and they are different claims about what the release is — a larger model, or the same model post-trained harder. Under the source-priority rule the vendor statement outranks third-party reporting, so the Compared To table keeps both figures rather than replacing Musk's with an outlet's. This resolves the moment xAI publishes a model card. Nothing read on 2026-08-16 states a parameter count at all, so this conflict is untouched by that capture (source).

Release date. This page has carried 2026-08-07 since 2026-08-08, from launch coverage dated that day (source). Four separate documentation write-ups read on 2026-08-16 date the release to 2026-08-12 (source).

Both sides are third-party, so the source-priority rule does not separate them, and no first-party xAI page is reachable from this environment to settle it. The Released row keeps 2026-08-07 — it is the earlier claim, it is what this page has published for eight days, and replacing a dated claim with a later one of equal standing would be a silent correction of the kind Qwen 3.8 27B exists to warn against. The 08-12 date is recorded here instead. Note that the model's absence from the 2026-08-09 Artificial Analysis snapshot is consistent with either date and settles neither.

Sources

• Elon Musk on X (July 18, 2026) → (source)
• Grok 4.6 reported launch, 2026-08-07 — third-party only; no first-party page read
• Grok 4.6 specification as reported, read 2026-08-16 — context window, pricing, modalities, effort levels and a second release date; third-party only
• Artificial Analysis leaderboard, 2026-08-16 — first independent measurement held here
• Dataconomy: https://dataconomy.com/2026/07/20/musk-xai-grok-4-6-2t-model-training-next-week/
• WION News: https://www.wionews.com/world/musk-s-answer-to-china-s-ai-monster-2-trillion-parameter-grok-4-6
• Kie.ai: https://kie.ai/blog/what-is-grok-4-6

Key Facts (as of announcement)

• Parameter count: 2 trillion (vs. Grok 4.5's 1.5T — 33% larger)
• Performance claim: "better than [Grok 4.5] in every way" (Musk, unverified)
• Speed target: "keeping speed and token efficiency close" to Grok 4.5
• Motivation: Musk framed it as a direct response to Moonshot AI's Kimi K3 (2.8T MoE, announced July 16, 2026)
• Monthly cadence: consistent with xAI's June 28 announcement of monthly releases for the remainder of 2026

⚠️ No benchmarks, pricing, context window, or official launch date have been published. Treat performance claims as pre-release promotional statements until verified against third-party evals.

Written at announcement. Partly overtaken on 2026-08-16: a pricing table and a 500K context window are now documented by several third parties, and Artificial Analysis supplies one independent index figure — see ## Spec and ## Benchmarks. Still standing: no xAI model card, no vendor benchmark table, no parameter count, and now two competing release dates. The warning was about first-party silence, and xAI has not broken it.