Grok 4.7 Release: Benchmarks, Price, Independent Tests

§ 01What xAI shipped#

The post is titled “Introducing Grok 4.7” and dated Sep 21, 2026. Its subhead: “SpaceXAI’s most powerful model for coding and knowledge work. Twice as fast, at half the price of comparable models.” Its first paragraph is more careful than its subhead: the model “works longer on difficult tasks, checks its own work more carefully, and comes with our best-calibrated safeguards to date”, is “Served at the same price and speed as” Grok 4.6, and “is highly competitive in its class”.

01Same price as Grok 4.6, higher on every row xAI published against it

02Where 4.7 leads and where it does not, per xAI's own launch table

03Four dates from one account, and the day the company shipped

1 / 3

←→

The price row is the 4.6 row to the cent, and the docs confirm it: the two grok-4.7 lines in xAI’s pricing table match the two grok-4.6 lines. That is what “half the price of comparable models” means in the subhead: not cheaper than 4.6, cheaper than the frontier models in xAI’s own comparison, which list at $4/$20 (GPT-5.6 Sol) and $10/$50 (Fable 5.1).

§ 02xAI’s benchmark table#

The launch post carries one table with four columns: Grok 4.7 at xHigh effort, Grok 4.6 at High, GPT-5.6 Sol at Max and Fable 5.1 at Max. These are xAI’s numbers, run by xAI. The first independent read, from Artificial Analysis, landed 21 minutes after launch and sits in the next section.

Read across the rows: 4.7 beats 4.6 on all seven, with the largest jump on Terminal-Bench 4.0, which nearly doubles. Against Fable 5.1 Max it leads on EEBench and the Harvey legal benchmark, sits within a point on DeepSWE and AA Briefcase, and trails on the two coding benchmarks Cursor users care most about, on HealthBench and on GDPval. Against GPT-5.6 Sol Max it leads on four rows. The post’s own framing, “at the frontier in price-performance”, is the claim the table supports; “most powerful” is the subhead’s word and the table does not support it against Fable.

§ 03The first independent reads#

Artificial Analysis posted its Intelligence Index result at 16:38 UTC, 21 minutes after xAI’s announcement. Cursor’s own post went up at 16:22 UTC. Neither is a full third-party benchmark table yet; both are the first numbers not run by xAI.

The two reads agree on the shape. Artificial Analysis’s own framing is that “Grok 4.7’s gains come with higher token usage”: the model thinks more than twice as long per task as 4.6 to earn its score, which is what xAI’s “works longer on difficult tasks” means in tokens. At the same per-token price as 4.6, a task that used to cost one unit now costs about two. Cursor confirms the price and keeps 4.6 on the menu, which is the honest option for anyone whose tasks did not need the extra thinking.

Two platforms posted their own availability notes the same day, which is where most of the model’s first paid traffic will come from.

§ 04Week one: the independent tests#

By September 26 the first third-party results had moved past launch-day reads. None of them is run by xAI.

Snorkel’s benchmark is the most useful of these for a buyer, because it grades code quality as well as correctness: “It reaches a tasteful pass@3 of 40.0%, up from 38.9% for Grok 4.6.” That is a 1.1-point gain, far smaller than the 17.7 points xAI’s own table shows on Terminal-Bench 4.0, and a reminder that the vendor picks the rows. Arena’s changelog records the additions but not a settled rank; ranks on young entries move as votes arrive.

§ 05The Cyber Index: first on a new benchmark#

On September 28, 2026, Artificial Analysis launched the Artificial Analysis Cyber Index, a defensive-security benchmark that averages three tests on real software: finding vulnerabilities (CWE-Bench-AA), validating them (DeepsecBench-AA) and patching them without breaking the code (CyberGym-E2E-AA). Grok 4.7 at xhigh effort scored 56 percent, tied for first with Xiaomi’s MiMo-V2.6-Pro among the 17 models on the launch chart.

The frontier models from OpenAI and Anthropic sit lower for a reason the index states itself: “GPT-6 Astra, GPT-6 Sol, Claude Fable 5.1, Claude Opus 5.5, Qwen3.8 2.4T A95B, and Qwen3.8 27B refuse at least 98% of tasks, making it difficult to assess frontier model performance.” Part of Grok 4.7’s lead reflects what it attempts, not only what it solves. For a security team that is the practical question: a model that declines the patching task scores zero on it, whatever it could have done.

§ 06VulcanBench: top three on an outside coding board#

On October 4, 2026, VulcanBench updated its Frontier v4 leaderboard with 92 runs of Grok 4.7 through Cursor’s agent CLI: 23 tasks at each of the four effort levels Cursor offers (it has no Max). Grok 4.7 took first, second and third place, 93.15 at Extra-high, 92.71 at High and 92.30 at Medium, ahead of Fable 5.1 at max (91.84) and Opus 5.5 at high (91.11). Low came 12th at 89.42. The tasks are behavioral reconstructions graded by hidden tests; the combined score weights functional correctness at 50 percent, lint and complexity at 8.5, security at 8.5 and Code quality, judged by a model panel for a human reader, at 33.

Two things set this column apart from the rest of the board. The judges changed. Every other column is scored by Muse Spark 1.3 and Grok 4.6, but Grok 4.6 “is not neutral for an xAI model”, so VulcanBench seated GPT-6.1 Sol in its place for Grok 4.7. The board itself says GPT-6.1 Sol rates Grok 4.7’s code about 6 points above Muse, so the column is “not strictly comparable with the other columns”. VulcanBench reran the check on Muse Spark 1.3 alone for every column: Grok 4.7 then scores 88.64 to 92.27, still first at Medium, High and Extra-high, and second at Low behind Fable 5.1 (89.46). The order survives the single-judge check; the exact margins depend on the panel.

The cost is missing. In the board’s words, “$/task is unavailable, not $0”: the sweep ran on a Cursor subscription and VulcanBench has no list price for Grok 4.7, so it sits off the cost chart. Time is published, and it tells the same story as the token counts above: 28.5 minutes per task at Extra-high, about level with Fable 5.1 at max (27.1) and nearly three times GPT-6 Astra at max (10.3). Medium is judged on 22 of 23 tasks, because one run hit the 3-hour bound and counts as failed.

Musk shared the board on October 6 at 15:32 UTC, quoting a @XFreeze post of the top three: “Grok 4.7 is excellent at dealing with large code repositories”. It is the first general coding board in this record to put 4.7 ahead of Fable 5.1. xAI’s own launch table had Fable 5.1 Max ahead on CursorBench and Terminal-Bench, and Snorkel’s Senior SWE-Bench put 4.7 sixth. Three coding benchmarks, three answers: which one predicts your work depends on whether your tasks look like hidden-test reconstructions, long terminal sessions, or code a senior engineer has to approve.

§ 07What Grok 4.7 costs, by where you buy it#

xAI’s own price did not move in week one: its pricing page still lists grok-4.7 at Grok 4.6’s rates, and above 200k prompt tokens both double. xAI describes Fast as “the same Grok 4.7 model served on faster infrastructure, at twice the standard token rates,” sold only through Cursor and Grok Build. OpenRouter’s catalog lists xAI’s own endpoint 20 percent under the direct rate. Price per token is only half the bill: Artificial Analysis measured Grok 4.7 using about twice the output tokens per task of Grok 4.6, so the same task costs more even where the rate is unchanged.

§ 08The rumors, graded against the launch#

This page existed because, for two months, everything known about Grok 4.7 came from one X account. Now the company has spoken, and each claim can be marked.

§ 09The dates, closed#

Every date came from Musk. Here is how they held, now that the company has shipped.

The pattern this page described on September 8 held: optimism, not fiction. The model was real, it shipped, and it shipped after the dates. 4.7’s slip lands between 4.6’s five days and 4.5’s seven weeks. The one thing that changed between the September 14 grade and the launch is the founder’s tone: “roughly on par with Opus 5.0” became, on launch day, “a strong combination of intelligence, speed & low cost” (Musk, 16:25 UTC).

§ 10The chain, dated and closed#

§ 11The founder’s record, word for word#

Kept for the record; these are the posts the tracker graded, with their sources.

July 24: “Grok 4.6 in 2 weeks and Grok 4.7 in 4 weeks” (source)

July 28: “Grok 4.7 will be the 2.1T model released a few weeks later. This will be better than 4.6 in every way, except slightly slower to serve, albeit with even better token efficiency.” (source)

August 12: “Grok 4.7 is significantly better than 4.6 and should be ready in 3 to 4 weeks. Initial training is complete and now we’re adding a massive amount of SpaceX company data in supplemental training.” (source) Six minutes later: “Grok 4.7 will exceed all current models.” (source)

September 1: “Grok 4.7 comes out in 10 days” (source)

September 11: “Grok 4.7 needs a few more days to cook. We might have penalized response length too much (or something) in RL, as it still gives up on hard tasks (that it can do!) too early and isn’t yet sufficiently rigorous in checking its work.” (source)

September 14: “Grok 4.7 should be roughly on par with Opus 5.0, not 5.1. Better in some ways, worse in others. We need to fix multimodal performance. Grok 4.8 will be a noticeable improvement. Grok 4.9 is probably Astra/Fable class. Grok 5 maybe better than anything. We shall see.” (source)

September 21, 16:25 UTC: “Grok 4.7 is a strong combination of intelligence, speed & low cost” (source)

The September 11 post is the one the launch text answers: xAI’s model “works longer on difficult tasks, checks its own work more carefully”. The July 28 post is the one it contradicts: not slower to serve, the same speed. The July 21 and August 12 SpaceX-data posts are the ones it ignores.

§ 12The 4.x cadence, closed#

Grok 4 Fast and Grok 4.1 Fast are left out as speed variants of existing models. The 4.20 row uses Musk’s February 17 public-beta announcement; the 4.5 row uses the x.ai news date of July 16. The next model on the founder’s ladder, Grok 4.8, has its own tracker; the September 13 post said its training would finish that week and RL would start, which on this table’s cadence points at October at the earliest.

§ 13What this means if you run agents#

For anyone on Grok 4.6, the swap is a model id at the same price with a vendor table that says every row improves; the Terminal-Bench jump is the one that matters for long agentic work. For anyone choosing an engine, xAI’s table places 4.7 as the price-performance option below Fable 5.1 Max on coding, and the safety numbers are the first hard ones xAI has published for a Grok launch. For anyone choosing where agents run, none of this changes the decision: an engine swap is a version string, and what persists is the layer around it.

§ 14What we are watching for#

• Settled ranks. Arena added Grok 4.7 to its Text, WebDev and Agent boards between September 22 and 25; a settled rank on each, an Epoch read and a SWE-bench Verified score are still open. Every row in xAI’s own table is still xAI’s.
• Grok Bot. The Grok Bot docs now open on a “Meet grok-4.7” link; xAI has still not said whether running Bots moved to 4.7 or move automatically.
• The fast variant’s price row on docs.x.ai, and whether the 4.6 fast row is retired.
• The Opus comparison. Musk graded 4.7 against Opus 5.0; xAI’s table omits it and Artificial Analysis’s launch-day posts did not place the two side by side. VulcanBench’s October 4 board is the first to place 4.7 beside Opus 5.5 (not 5.0): ahead of Opus 5.5’s best column at Medium, High and Extra-high.

• Grok 4.8. The successor named on September 13; its tracker carries the chain.

• The week after the discount. Vercel’s 40 percent is for one week from September 21; Copilot has not published the model’s premium-request multiplier. Both prices, and whether Cursor keeps 4.6 alongside 4.7 past the first week, land here.

§ 15The record#

As of September 21, 2026, 16:40 UTC: x.ai/news/grok-4-7 returned 200 to a browser-agent read at 16:29 UTC with the launch text quoted above; docs.x.ai/developers/models.md lists grok-4.7 (500k context, $2.00/$0.50/$6.00 under 200k, $4.00/$1.00/$12.00 above, knowledge cutoff May 2026); OpenRouter’s x-ai/grok-4.7 entry carries a created time of 16:19:01 UTC. X posts are id-clocked: @SpaceXAI 16:17:53 and 16:17:54 UTC (four posts), @elonmusk 16:25:47 UTC. All benchmark figures are xAI’s, read from the launch post.

§ 16Update log#

This is a living page; when the story moves, the update lands here.

As of September 8, 2026: page opened. Latest Musk target is about September 11; no official movement to log.

September 11, 2026: the September 11 target passed with no release; Musk: “needs a few more days to cook”, blaming RL that penalized response length. First stated reason for a delay.

September 13, 2026: with 4.7 still unreleased, Musk named its successor, Grok 4.8; a separate tracker opened.

September 14, 2026: Musk graded the model for the first time, “roughly on par with Opus 5.0, not 5.1”; no new date.

September 21, 2026, 16:17 UTC: Grok 4.7 released. Page flipped from tracker to release record on the same URL: spec and price table from the launch post and docs, xAI’s seven-row benchmark table and GDPval chart, the pre-launch claims graded (2.1T unconfirmed, SpaceX data absent, “exceed all current models” not met, “slower to serve” contradicted), the four dates closed (31 days from the first target), the cadence table closed at 40 days. The rumor-era hero replaced.

September 21, 2026, 19:05 UTC: first independent read folded. Artificial Analysis (16:38 UTC): Intelligence Index 46, rank 16 of 655, 81,000 output tokens per task against 36,000 for 4.6, about 188 tokens per second, hallucination rate 29 from 34 percent. Cursor’s post (16:22 UTC) confirms the 4.6 price and keeps 4.6 available. Grok Bot docs open on a “Meet grok-4.7” link, no running-Bot statement. New section, FAQ and watch list updated.

September 22, 2026, 01:30 UTC: distribution rows added. GitHub (changelog, September 21) has Grok 4.7 rolling out in Copilot; Vercel (changelog, September 21) serves it on AI Gateway, fx and eve at 40 percent off for a week. New watching item on the two prices.

September 26, 2026, 07:10 UTC: week-one section added. Arena added grok-4.7-xhigh to Code Arena: WebDev (September 22), Agent Arena (September 24) and Text Arena (September 25); Snorkel’s Senior SWE-Bench put it at 40.0 percent tasteful pass@3, sixth, $0.24 per trial; price rows for the xAI API, OpenRouter’s xAI endpoint and Fast; the Cursor Cloud Agent reports of September 22. Title now leads on benchmarks and price.

September 29, 2026, 02:15 UTC: Cyber Index section added. Artificial Analysis’s new Cyber Index (September 28) puts Grok 4.7 (xhigh) at 56 percent, tied first with MiMo-V2.6-Pro of 17 models; GPT-6 Astra, GPT-6 Sol, Fable 5.1 and Opus 5.5 refuse at least 98 percent of its patching tasks.

October 6, 2026, 16:09 UTC: VulcanBench section added. Its Frontier v4 board (updated October 4) puts Grok 4.7 through Cursor first, second and third (93.15 Extra-high, 92.71 High, 92.30 Medium), ahead of Fable 5.1 at max (91.84) and Opus 5.5 at high (91.11); GPT-6.1 Sol replaced Grok 4.6 as second judge for this column, the board calls it not strictly comparable, and on Muse Spark 1.3 alone 4.7 stays first at Medium to Extra-high. Cost per task unavailable. Musk shared it at 15:32 UTC.

§ 17Sources#

• Arena, leaderboard changelog, entries of September 22, 24 and 25, 2026. Snorkel AI, Grok 4.7 on Senior SWE-Bench, September 22, 2026. xAI, pricing and rate limits, read September 26, 2026. OpenRouter, Grok 4.7 endpoints, read September 26, 2026.

• xAI, Introducing Grok 4.7, September 21, 2026; read September 21, 2026, 16:29 UTC.

• xAI, Models and pricing, read September 21, 2026.
• @SpaceXAI, launch thread on X, September 21, 2026, 16:17 UTC; read in Chrome.
• Elon Musk, posts on X linked inline above, July 24 to September 21, 2026.
• xAI, Introducing Grok 4.6, August 12, 2026.
• Our own records: Grok 4.8 release date; What is Grok Bot; GLM-5.3-Flash and Ox Alpha. Artificial Analysis, Grok 4.7 (xhigh) and its launch-day thread, read September 21, 2026, 19:03 UTC; Cursor forum, Grok 4.7 is now live, 16:22 UTC; Grok Bot overview, read 19:03 UTC.
• GitHub, Grok 4.7 is now available in GitHub Copilot, September 21, 2026; read September 22, 2026.
• Vercel, Changelog, September 21, 2026 entries; read September 22, 2026.
• VulcanBench, Frontier v4 leaderboard, updated October 4, 2026; read October 6, 2026. @XFreeze, October 6, 2026, 10:17 UTC, and Elon Musk, October 6, 2026, 15:32 UTC, on X; read in Chrome.