ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›ARTICLES›GROK 4.7 VS GROK 4.6: WHAT ACTUALLY CHANGED
ARTICLES · COMPARISONBEGINNER

Grok 4.7 vs Grok 4.6: What Actually Changed

Same price, same context window, a new and larger base model. What xAI’s own benchmarks show, what independent testers found, and where the fast variant runs.

READ11 min
WORDS2,267
SECTIONS9
TYPECOMPARISON
CHECKED1 OCT 26

A bigger model at the same price. Grok 4.7 was released on 21 September 2026, about six weeks after Grok 4.6, and its maker (then called SpaceXAI, now SpaceXSI) says it costs and runs the same: “Served at the same price and speed as Grok 4.6”. What changed is underneath: “Grok 4.7 uses a new, larger base model compared to Grok 4.6”.1

TL;DR — THE SHORT VERSION
  • In one line: Grok 4.7 is a modest step up from 4.6 at the same price. Independent testers found it better, but by less than xAI’s own numbers suggest.
  • Price, context window and reasoning settings are unchanged on the public API. xAI’s release notes describe both models in near-identical words.
  • The model underneath is new: a larger base model and a longer reinforcement-learning run aimed at tasks that take hours.
  • Independent tests now exist. Artificial Analysis (21 Sep 2026) put 4.7 two points above 4.6 on its overall index. On one coding test it scored 26%, where xAI reported 38.0%.2
  • It writes much more per task: about 81k output tokens against 38k for 4.6 in the same test, so the same price per token can still mean a bigger bill.
  • The comparison is not like for like: xAI tested 4.7 at its highest reasoning setting and 4.6 one step lower.
  • Grok 4.6’s own scores moved between posts, so compare numbers from one table, never across two.
  • The fast variant is Cursor and Grok Build only. It is not on the public API.
◈ PRICES ON THIS PAGE

Every price here was correct when this site last checked it, on 22 Sep 2026. Prices change often and this site no longer updates them, so check the vendor’s own page before you rely on one: xAI pricing.

What stayed the same?

On the API, almost everything a developer sets or pays for. xAI’s release note for Grok 4.7 repeats the one it wrote for Grok 4.6 in August: a 500k context window, text and image in, text out, no text output limit, and the same four reasoning levels, low, medium, high (the default) and xhigh.3

The price is the same to the cent. Both entries give it in one sentence: $2 per million input tokens, $0.50 for cached input and $6 for output, doubling once a prompt passes 200k tokens. That is a snapshot from 22 September 2026: vendor prices change, so re-check the live page before budgeting.4

TAKEAWAY

Switching from 4.6 to 4.7 is a one-word change in the model name, not a budgeting decision. Whether it is worth doing depends on behaviour, which is where the price pages cannot help.

What changed under the hood?

The training, according to the launch post. 4.7 “was trained with a longer reinforcement learning run on a harder mix of tasks, weighted toward problems that take many hours to complete.” xAI adds: “The model is better at verifying its own work and managing longer context.” It also says it trained 4.7 to understand Grok Bot, the agent product xAI launched in August.5

That continues the direction of the last release. Grok 4.6 was also pitched at long-running agents, and xAI said then that it had started to see the model checking its own work (see Grok 4.6 vs 4.5).

xAI’s models page gives 4.7 a knowledge cut-off of May 2026 and now tells developers to use it for everything except images, video and voice.6

On safety, the post claims “an entirely new safeguard stack”, a score of 62.4% on LatchBio’s biosafety benchmark, and that on xAI’s own cyber benchmark the model let through “only 3.3% of risky dual-use prompts”. Both figures come from xAI, and the second from a benchmark it built.5

TAKEAWAY

A new, larger base model is a bigger change than the version number suggests. Same price does not mean same behaviour: prompts tuned on 4.6 may need re-testing.

What do the benchmarks say?

xAI’s published comparison, Grok 4.7 at xHigh against Grok 4.6 at High, read at source 22 Sep 2026
Benchmark (what it tests, per xAI)Grok 4.7Grok 4.6
Terminal-Bench 4.0 (multi-hour terminal work)38.0%20.3%
EEBench (electrical engineering)64.0%53.0%
HealthBench Professional (clinical reasoning)56.7%48.5%
CursorBench 4.0 (software engineering)46.3%40.4%
DeepSWE v1.1 (software engineering)71.0% (high effort)65.2%
Harvey Legal Agent Benchmark (legal work)19.6%15.8%
AA Briefcase v1.1 (multi-hour office work)1,6571,546
GDPval (Elo score)16951605

SpaceXAI, Introducing Grok 4.7, 21 Sep 2026, read at source 22 Sep 2026. xAI’s own table, which describes itself as “Token prices and benchmark scores for Grok 4.7, Grok 4.6, GPT-5.6 Sol, and Fable 5.1.” These are xAI’s own scores. One row, Terminal-Bench 4.0, has since been run by an independent tester, with a lower result: see the next section.

GROK 4.7 AGAINST 4.6 · xAI’S OWN SCORES
The percentage rows of xAI’s table read on 22 Sep 2026, drawn to scale, largest gap first. Grok 4.7 was run at xHigh, Grok 4.6 at High. DeepSWE is left out because its 4.7 score used a different setting.
Grok 4.7Grok 4.6
Terminal-Bench 4.0 · Grok 4.738.0%
Terminal-Bench 4.0 · Grok 4.620.3%
EEBench · Grok 4.764.0%
EEBench · Grok 4.653.0%
HealthBench Professional · Grok 4.756.7%
HealthBench Professional · Grok 4.648.5%
CursorBench 4.0 · Grok 4.746.3%
CursorBench 4.0 · Grok 4.640.4%
Harvey Legal Agent Benchmark · Grok 4.719.6%
Harvey Legal Agent Benchmark · Grok 4.615.8%
SpaceXAI, Introducing Grok 4.7, 21 Sep 2026, read at source 22 Sep 2026. xAI’s own table, which describes itself as “Token prices and benchmark scores for Grok 4.7, Grok 4.6, GPT-5.6 Sol, and Fable 5.1.” These are xAI’s own scores; the independent Terminal-Bench 4.0 result is in the next section.

In xAI’s own figures, 4.7 is ahead of 4.6 on all eight. The gap is widest on Terminal-Bench 4.0, where the score nearly doubles, and narrowest on the legal benchmark and the two software-engineering tests.

Two things make these rows less clean than they look. First, the settings differ: the table puts 4.7 at xHigh against 4.6 at High, yet xAI’s own release note says 4.6 also supports xhigh. The DeepSWE figure for 4.7 is marked as a high-effort score, so that one row is closer to like for like.7 Second, the 4.7 post carries no note on where the scores came from. The 4.6 post did.8

TAKEAWAY

The direction is consistent, and the biggest claimed gains line up with what xAI says it trained for: long, many-step work. How big the gains really are is a different question, and the independent results below give a smaller answer.

What did independent testers find?

Better than 4.6, but by less than xAI’s table suggests. When this page was first written, on 22 September, every score came from xAI. Since then, Artificial Analysis, an independent company that runs the same set of tests on many AI models, has published its own results for Grok 4.7. It tested 4.7 at its highest reasoning setting, xhigh.9

Artificial Analysis results, published 21 Sep 2026, read 1 Oct 2026
Test (what it means)Grok 4.7Compared with Grok 4.6
Intelligence Index (one overall score built from ten tests)462 points higher
Coding Agent Index (how well it does real coding jobs, run inside Grok Build)56, ranked 4th9 points higher
Output tokens per task (how much text it writes, hidden reasoning included)about 81k4.6 used about 38k
Hallucination rate (how often it makes things up; lower is better)29%4.6 scored 34%

Artificial Analysis, Benchmarking Grok 4.7, 21 Sep 2026, read at source 1 Oct 2026: “Grok 4.7 + Grok Build now ranks 4th”; about 81k output tokens, “more than double the 38k used by Grok 4.6 (xhigh)”; “Grok 4.7 (xhigh) has a lower AA-Omniscience Hallucination Rate than Grok 4.6 (high): 29% versus 34%.” The ten-test description is The Decoder’s (below). Scores and ranks change as new models are added: re-check the source.

On the coding test with xAI’s biggest gain, the independent score was lower. xAI reported 38.0% on Terminal-Bench 4.0. Artificial Analysis measured 26%, according to The Decoder, a tech news site. The same report puts Grok 4.7 “mid-pack” overall, behind Claude Fable 5.1 and GPT-6 at 53 each.10

◈ IN PLAIN TERMS

A company testing its own model picks the test, the settings and the moment. An outside tester running the same test can get a different number. Here, independent testers got a lower coding score than xAI’s own: 26% against 38.0%. Neither number is “the truth”; the gap tells you how much the setup matters.11

The token count matters for cost. You pay for every token a model writes. At the same price per token, a model that writes twice as much to finish a job costs roughly twice as much for that job.

TAKEAWAY

Grok 4.7 is a real but modest improvement in outside testing: a little smarter overall, clearly better at agent coding, less prone to making things up, and much wordier. Test it on your own work before trusting any headline number.

What Musk and the press say, but xAI’s docs do not

These are claims, not documented facts. Two numbers about Grok circulate that xAI’s developer pages do not carry.

  • Model size. Decrypt reported on 21 September 2026 that Grok 4.7 has 2.1 trillion parameters (the internal numbers a model learns in training; more usually means a bigger model), against 1.5 trillion for 4.6. xAI’s launch post gives no parameter count, checked 1 Oct 2026.12
  • Grok 4.8. Musk says it exists. On 14 September 2026 he posted on X that Grok 4.8 is a 2.5-trillion-parameter model that would “finish training this week”. As of 1 October 2026 there is no Grok 4.8 on xAI’s models list, and Grok 4.7 is still the newest model there.13
TAKEAWAY

Treat parameter counts and next-version dates from posts as announcements of intent. The models list is where a release becomes real.

Why 4.6’s own scores changed

Put the August post next to the September one and Grok 4.6 scores differently in each. Some of that is new test versions. In August xAI reported 69.9% on CursorBench v3.2; in September it reported 40.4% on CursorBench 4.0, a newer version of the test.14

One row changed with the version number unchanged. DeepSWE v1.1 is named in both tables, both times for Grok 4.6 at High, and the score is 65.9% in August and 65.2% in September. Neither post explains the difference.15 It is a small gap, but it means a score is a result from one run of one harness on one day, not a fixed property of the model.

Our Grok 4.6 vs 4.5 page quotes the August table and stays accurate to it. Do not mix its numbers with the ones on this page.

TAKEAWAY

Only compare two models inside one table from one date. A score lifted from an older announcement may have been measured on a different version of the test.

The fast variant, and where 4.7 runs

The launch post says: “We also serve a fast variant with twice the output speed at twice the price.” The developer docs narrow where you can get it: only in Cursor and Grok Build, not through the public API.16 Grok 4.6 also had a fast tier at double the price; its post did not say where it was sold.17

Standard 4.7 runs on the xAI API, in Cursor, as the default model in Grok Build, and through OpenRouter, Vercel and Cloudflare.18 xAI’s US-only endpoint, at a 10% premium, is not a 4.7 difference: it serves both models.19

One API detail appears in the 4.7 release note and not the 4.6 one: on the Responses API, 4.7 “always returns reasoning.encrypted_content”, so client code that stores or filters responses may see an extra field.16

TAKEAWAY

If you call the public API, the fast variant is not an option for you. The standard 4.7 is what you get, at the 4.6 price.

Should you switch?

At the same price, the only cost of trying 4.7 is re-testing. Run the prompts you already rely on through both models and compare the output, rather than trusting either table. Grok 4.6 is still on xAI’s models list as of 1 October 2026, with no retirement notice, so there is no forced move yet.20 One thing to watch in your test: independent testing found 4.7 writes about twice as many output tokens per task as 4.6 (see above), so check what a typical job costs you, not just the rate card. If you need results that do not shift under you, xAI says a model name with a date on the end “refers directly to a specific model release”.21

For how Grok’s price sits against other vendors, see LLM API prices compared and Gemini vs Grok API prices. The company signing these posts is SpaceXSI (formerly SpaceXAI), the renamed xAI.22 Who owns Grok explains the SpaceX acquisition.

TAKEAWAY

Same price makes this an easy trial and a hard claim to judge. Your own prompts are the only benchmark that answers your question.

What this page could not verify

  • Most of xAI’s benchmark and all of its safety numbers. This site re-ran nothing. Of xAI’s table, only the Terminal-Bench 4.0 row has an independent result we could find (26% against xAI’s 38.0%, see above).11
  • Grok 4.7’s parameter count and Grok 4.8. Both come from press reports and Musk’s posts, not from xAI’s documentation.
  • “Twice as fast, at half the price of comparable models.” The launch post’s headline line. Which models count as comparable is xAI’s choice.
  • Grok 4.6’s knowledge cut-off. Its developer page now redirects to the Grok 4.7 page, so there is nothing to compare against.
  • Why Grok 4.6’s DeepSWE score changed. Neither post explains it.
  • When Grok 4.6 will be retired. No date was published as of 1 October 2026.
SOURCES
  1. 1SpaceXAI, Introducing Grok 4.7, 21 Sep 2026, read at source 22 Sep 2026.↑
  2. 2Artificial Analysis, Benchmarking Grok 4.7, 21 Sep 2026, read at source 1 Oct 2026. The Decoder, report, 21 Sep 2026, read at source 1 Oct 2026“Grok 4.7 scores 26 percent on Terminal-Bench 4.0’s agentic coding test”. ↑
  3. 3xAI, Release Notes, entries for 21 September and 12 August 2026, read at source 22 Sep 2026“It has a 500k context window, text and image inputs with text-only output, and no text output limit.” ↑
  4. 4xAI, Release Notes, read at source 22 Sep 2026, in both the Grok 4.7 and Grok 4.6 entries: “Pricing is $2 / $0.50 / $6 per 1M tokens (input / cached input / output) below 200k prompt tokens, and $4 / $1 / $12 above.”↑
  5. 5SpaceXAI, Introducing Grok 4.7, 21 Sep 2026, read at source 22 Sep 2026“We also trained Grok 4.7 to natively understand the Grok Bot harness”. xAI, Release Notes, 11 August 2026 entry: “Durable AI teammates that work on a persistent cloud computer, with messaging, approvals, connectors, and routines.” ↑a“topping LatchBio’s biosafety benchmark at 62.4%”. ↑b
  6. 6xAI, Models, page undated, read at source 1 Oct 2026“The knowledge cut-off date of Grok 4.7 is May 2026.” and “For everything else, including code, use Grok 4.7. It is the most capable model we’ve built.” ↑
  7. 7SpaceXAI, Introducing Grok 4.7, read at source 22 Sep 2026“An asterisk on Grok 4.7 DeepSWE marks a high-effort score.” xAI, Release Notes, 12 August 2026 entry for Grok 4.6: “Reasoning effort supports low, medium, high (default), and xhigh.” ↑
  8. 8SpaceXAI, Introducing Grok 4.6, 12 Aug 2026, read at source 22 Sep 2026“Third-party model scores are the best of self-reported or publicly available results.” ↑
  9. 9Artificial Analysis, Benchmarking Grok 4.7, 21 Sep 2026, read at source 1 Oct 2026“Grok 4.7 scores 46 on the Artificial Analysis Intelligence Index”. ↑
  10. 10The Decoder, xAI launches Grok 4.7 at bargain prices, but benchmarks reveal a wide gap to Claude and GPT-6, 21 Sep 2026, read at source 1 Oct 2026“Grok 4.7 scores 26 percent on Terminal-Bench 4.0’s agentic coding test” and “Grok 4.7 scores 46 and lands mid-pack. Claude Fable 5.1 and GPT-6 lead with 53 each.” ↑
  11. 11The Decoder, report, 21 Sep 2026, read at source 1 Oct 2026“Grok 4.7 scores 26 percent on Terminal-Bench 4.0’s agentic coding test”; xAI’s 38.0% from its launch table above. ↑a“Grok 4.7 scores 26 percent on Terminal-Bench 4.0’s agentic coding test”. ↑b
  12. 12Decrypt via Yahoo Tech, xAI Launches Grok 4.7. It’s Bigger, But Late to the AI Frontier Party, 21 Sep 2026, read at source 1 Oct 2026“Grok 4.7 packs 2.1 trillion parameters, up 40% from the 1.5 trillion in Grok 4.6.” ↑
  13. 13TeslaNorth, Elon Maps the Grok Roadmap Through Grok 5 After 4.8 Training, 14 Sep 2026, read at source 1 Oct 2026, quoting Musk’s post“Grok 4.8, which is a 2.5T model trained with our new C++ software stack, will finish training this week and start RL”. xAI, Models, read at source 1 Oct 2026: newest entry grok-4.7, no 4.8 or 5 listed. ↑
  14. 14SpaceXAI, Introducing Grok 4.6, 12 Aug 2026, and Introducing Grok 4.7, 21 Sep 2026, both read at source 22 Sep 2026table rows “CursorBench v3.2 69.9%” (August) and “CursorBench 4.0 46.3% 40.4%” (September, Grok 4.7 then Grok 4.6). ↑
  15. 15SpaceXAI, Introducing Grok 4.6 (table row “DeepSWE v1.1 65.9%”) and Introducing Grok 4.7 (row “DeepSWE v1.1 71.0%* 65.2%”), both read at source 22 Sep 2026.↑
  16. 16xAI, Release Notes, 21 September 2026, read at source 22 Sep 2026“Grok 4.7 Fast, the same model at twice the token rates, is available only through Cursor and Grok Build, not on the public xAI API.” ↑a“On the Responses API, grok-4.7 always returns reasoning.encrypted_content”. ↑b
  17. 17SpaceXAI, Introducing Grok 4.6, read at source 22 Sep 2026“Additionally, there is a fast variant which is twice the price.” ↑
  18. 18xAI, Grok 4.7, read at source 22 Sep 2026“OpenRouter, Vercel, and Cloudflare”. ↑
  19. 19xAI, Pricing, read at source 22 Sep 2026“a 10% premium”, models “Currently grok-4.7 and grok-4.6 only”. ↑
  20. 20xAI, Models, read at source 1 Oct 2026grok-4.6 listed, no deprecation notes. xAI, Pricing, read at source 22 Sep 2026: grok-4.6 listed at “500k” context, the same rates as grok-4.7. ↑
  21. 21xAI, Models, read at source 22 Sep 2026.↑
  22. 22SpaceXSI, x.ai home page, read at source 13 Oct 2026its structured data gives the name “SpaceXSI” with “SpaceXAI” and “xAI” as alternate names; the footer reads “© 2026 SpaceXAI LLC”. ↑

SpaceXAI, Introducing Grok 4.7, 21 September 2026 — Introducing Grok 4.6, 12 August 2026. xAI developer docs: Release Notes, Models, Pricing and Grok 4.7. All read at source on 22 September 2026; the Models page and the Grok 4.7 post re-read on 1 October 2026.

Independent and press: Artificial Analysis, Benchmarking Grok 4.7, 21 September 2026 — The Decoder, xAI launches Grok 4.7 at bargain prices, 21 September 2026 — Decrypt via Yahoo Tech, xAI Launches Grok 4.7, 21 September 2026 — TeslaNorth, Elon Maps the Grok Roadmap, 14 September 2026. All read at source on 1 October 2026.

Prices are a snapshot and vendor benchmarks are marketing until someone else runs them.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE