Grok 4.7

Spec

Attribute	Value
Developer	xAI
Released	2026-09-21
Announced	2026-09-21
Context window	500K
Pricing	$2/M input · $0.50/M cached input · $6/M output (prompts ≤200K); $4/M · $1/M · $12/M above 200K
License	unknown
Availability	Grok API, Cursor, Grok Build, OpenRouter, Vercel, Cloudflare
Three rows need their reading stated
(source):

• Context window carries 500K and there is a competing reading. Two passes give 500,000 tokens; one states "a 1 million token context window with no output token limit". 500K is the better-corroborated figure and the alternative is in ## Conflicting Reports.
• Pricing is tiered by prompt length, and the ≤200K tier is unchanged from Grok 4.6 — 3 passes for the base rates, 1 for the long-prompt tier. A fast variant at twice the output speed for twice the price is named in 1 pass with no rate given.
• License is unknown, not proprietary. Nothing read names a licence, terms of service or weight policy. No Catalogue id row is present: the model is listed on OpenRouter, but openrouter.ai is unreachable from this sandbox, so no catalogue string was copied and none is guessed.

No max output figure appears in anything read, which is why no row claims one.

Release Date

2026-09-21 (3 passes). Announced and released the same day.

The window history matters because cadence is this lab's central competitive claim. Musk gave "~4 weeks" on 2026-07-25 (≈2026-08-22), then "10 days" on 2026-09-02 (2026-09-12), then said on 2026-09-11 that the model "needs a few more days to cook" with no new date offered. The model arrived nine days after the third window closed and thirty days after the first (source).

Benchmarks

All vendor-reported as relayed by coverage. No harness, effort level, run count or variance is named for any row in anything read, which is the standing gap Eval Harness Configuration tracks.

Benchmark	Grok 4.6	Grok 4.7	Passes
CursorBench 4.0	40.4%	46.3%	2
EEBench	53.0%	64.0%	2
Harvey Legal Agent Benchmark	15.8%	19.6%	2
Against the comparison set (2 passes):

Benchmark	Grok 4.7	GPT-5.6 Sol	Claude Fable 5.1
Harvey Legal Agent Benchmark	19.6%	2.5%	6.7%
EEBench	64.0% (top of table)	not stated	below Grok
CursorBench 4.0	46.3%	not stated	above Grok
Summary claims: ahead of GPT-5.6 Sol on five of seven benchmarks, behind on
DeepSWE and HealthBench Professional; Fable 5.1 leads on
CursorBench, AA Briefcase, Terminal-Bench and HealthBench Professional.

The Harvey result is the only one worth reading as a capability claim, and it needs reading carefully. A 19.6% score that is 7.8× the nearest competitor's 2.5% is either a genuine discontinuity or a harness artefact, and nothing read permits telling which. A benchmark on which the entire field scores under 20% is one where the ordering is not yet stable, and this wiki records the spread rather than the ranking.

The comparison set is one generation behind the market. Grok 4.7 is measured against GPT-5.6 Sol and Fable 5.1 — models superseded the following day by GPT-6 Sol, GPT-6 Luna and Claude Opus 5.5. Nothing read compares Grok 4.7 to any of the three, so this release enters the wiki with no figure that can be set against the 2026-09-22 wave.

Use Cases

Stated positioning is coding and knowledge work (1 pass). The benchmark selection is the more informative statement of scope: CursorBench and DeepSWE for software engineering, EEBench for electrical engineering, the Harvey Legal Agent Benchmark for legal work, HealthBench Professional for clinical. Two of the four rows xAI leads on are domain-professional agent benchmarks rather than general coding, which is a different shape of claim from every prior Grok launch on this wiki.

Input accepts text, images, and files including PDFs, returning text; function calling via tools / tool_choice and structured outputs via JSON schema in response_format (1 pass each).

Compared To

• Grok 4.6 — the direct predecessor, 2026-08-07. Same price, 40% more parameters, +5.9 CursorBench, +11.0 EEBench, +3.8 Harvey. The clean reading is more capability per dollar; the honest one is that a 40% parameter increase bought single-digit gains on two of three published rows.
• Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna — the releases of 2026-09-22, one day later. No shared benchmark exists between Grok 4.7 and any of them in anything read. The comparison available is structural: Grok 4.7 at $2/$6 sits between GPT-6 Sol's $2/$10 and GPT-6 Luna's $0.10/$0.50, and below Claude Opus 5.5's $4/$20.
• Gemini 3.8 Flash — the other current model tracked here with a tiered long-prompt price, for the same reason: both charge more above a context threshold rather than publishing one rate.

Conflicting Reports

Context window: 500K against 1M

Two passes give 500,000 tokens. One pass states "a 1 million token context window with no output token limit". A third describes "a 256K context window with no text output limit" and attributes that itself to an earlier version. 500K is carried in ## Spec as the better-corroborated reading; neither alternative is adopted (source).

Whether Grok 4.7 beats Fable 5.1 on DeepSWE

One pass lists DeepSWE among the benchmarks Grok 4.7 wins against Fable 5.1. Another states Grok 4.7 falls behind GPT-5.6 Sol on DeepSWE. These are not strictly contradictory — different comparators — but no pass gives any DeepSWE figure for Grok 4.7, so neither claim is adopted and the row is absent from ## Benchmarks rather than filled from an ordering (source).

The parameter count this wiki could not settle

Grok 4.6 carries 2T announced against 1.5T in launch coverage under its own ## Conflicting Reports. This release is reported at 2.1T, a 40% increase over 1.5T — which adopts the lower of the two Grok 4.6 readings without saying so. Nothing read settles the Grok 4.6 figure, so the "40%" is recorded as xAI's arithmetic rather than as this wiki's (source).

Sources

• sources/blogs/xai-2026-09-21-grok-4-7.md
• sources/x/2026-09-11-musk-grok-4-7-delay.md
• MarkTechPost — SpaceXAI Releases Grok 4.7
• officechai — Grok 4.7 benchmarks
• Cryptopolitan — Grok 4.7 and the Harvey legal test
• SpaceXAI — Introducing Grok 4.7 — surfaced in search, not fetched