SpaceXAI’s Grok 4.7 now has a card on AIxploria, the AI tools directory, published at 01:29 UTC on 24 September 2026. It was the first of three frontier model cards the directory added that morning, ahead of Claude Opus 5.5 and GPT-6 Sol and Luna. The card calls Grok 4.7 “a reasoning model that is twice as fast and half the cost.”
That line is the problem. The same card’s review is headlined “the same old price,” and SpaceXAI’s own pricing page shows Grok 4.7 costs exactly what Grok 4.6 did. The “half the cost” phrase comes from SpaceXAI’s marketing, which compared Grok 4.7 with rival models. The card dropped the comparison, and the rival it was measured against cut its price in half the next day.
We checked the card against SpaceXAI’s announcement, documentation, pricing page and release notes. Here is what it gets right, what it gets wrong, what it leaves out, and how Grok 4.7 compares on price with the two models that launched a day later. We covered the run-up to this release in Grok 4.7 Spotted on GCP Quotas Page Ahead of Potential Release.
Table of contents
- The Grok 4.7 Card at a Glance
- What SpaceXAI Actually Launched
- Grok 4.7 Card Claims Checked Against SpaceXAI
- Where “Half the Cost” Came From
- Grok 4.7 Pricing Against the Models That Followed
- The Benchmarks, Including the Ones the Card Skips
- Musk’s Forecasts Against the Result
- The Rank Badge on the Grok 4.7 Card
- What the Card Leaves Out
- Should You Try Grok 4.7?
- Frequently Asked Questions About Grok 4.7 and SpaceXAI
- References
The Grok 4.7 Card at a Glance
AIxploria gives every tool a card with a blurb, a longer review, tags, a rating and a rank badge. The Grok 4.7 card follows the pattern, with some notable numbers.
The basics
The card sits at aixploria.com/en/grok-4-7-spacexai-model/. Its page title promises “Reviews, Price, Info & 24 Alternatives.” It is tagged “Latest AI” and “LLM models,” marked as a Verified Tool and labelled Freemium. It was published at 01:29:49 UTC and last edited nine minutes later.
Two votes
The card shows 71 upvotes. Its rating reads 4.5 out of 5, and the page’s structured data shows that figure comes from just two votes. The Claude Opus 5.5 and GPT-6 Sol and Luna cards published the same morning each had seven.
Who makes it
The card correctly names SpaceXAI as the developer and notes the company was “founded in 2023 by Elon Musk under the xAI name.” SpaceX and xAI combined earlier this year, and SpaceXAI now publishes its developer documentation and announcements under that name. SpaceX also completed its purchase of the coding tool Cursor in August, as we reported in SpaceX Cursor Acquisition Completed. That explains why Cursor was a launch partner.
How long it took
SpaceXAI announced Grok 4.7 on 21 September. Kalshi’s market on whether it would ship before 25 September closed at 22:30 UTC that day and settled Yes. The card appeared a little over two days later, in line with other major model cards in this series.
What SpaceXAI Actually Launched
SpaceXAI’s announcement is short, and its documentation is precise. Together they are the reference for every claim on the card.
The pitch
SpaceXAI calls Grok 4.7 its “most capable model for coding and knowledge work.” It says the model “works longer on difficult tasks, checks its own work more carefully, and comes with our best-calibrated safeguards to date.” It adds, crucially, that Grok 4.7 is “served at the same price and speed as Grok 4.6.”
How it was trained
According to SpaceXAI, Grok 4.7 uses “a new, larger base model compared to Grok 4.6.” It was trained with a longer reinforcement learning run “on a harder mix of tasks, weighted toward problems that take many hours to complete.” SpaceXAI also trained it to understand the Grok Bot harness natively, for conversational tasks and general knowledge work.
The specifications
SpaceXAI’s documentation lists the model name grok-4.7, a 500,000-token context window, text and image input with text output, “no text output limit,” a knowledge cutoff of May 2026 and four reasoning levels: low, medium, high (the default) and xhigh. It supports function calling, web search, X search and code execution.
Where it runs
Grok 4.7 is available on SpaceXAI’s API, on a US regional endpoint at a 10% premium, in Grok Build, where it is the default model, in Cursor on all plans, and through OpenRouter, Vercel and Cloudflare. A Fast variant, “the same model served on faster infrastructure, billed at twice the standard token rates,” is available only in Cursor and Grok Build.
Grok 4.7 Card Claims Checked Against SpaceXAI
We listed every checkable claim on the card and compared it with SpaceXAI’s pages.
| Card says | SpaceXAI says | Verdict |
|---|---|---|
| “Twice as fast and half the cost” (blurb) | “Half the price of comparable models”; “same price and speed as Grok 4.6” | Misleading |
| $2 input, $6 output, same as Grok 4.6 | Same, below 200K prompt tokens | Correct, incomplete |
| 500,000-token context, text and image input | Same | Correct |
| Four reasoning levels including xhigh | Low, medium, high (default), xhigh | Correct |
| Knowledge cutoff May 2026 | Same | Correct |
| Encrypted reasoning on the Responses API | Always returned on the Responses API | Correct |
| Fast variant, 2x speed at 2x price, Cursor and Grok Build only | Same | Correct |
| Free tier in Grok Build | Same; Fast variant excluded | Correct |
| On OpenRouter, Vercel and Cloudflare | Same | Correct |
| Five benchmark scores quoted | All five match the launch table | Correct |
| Scores “measured at the xhigh effort setting” | DeepSWE figure is marked “high effort” | Partly wrong |
| “Independent measurements are still the missing piece” | Artificial Analysis had scored it by 24 September | Out of date |
| Launched one day before Opus 5.5 and GPT-6 Sol | 21 September against 22 September | Correct |
The score
Of 13 claims, nine are fully correct, one is correct but omits a price tier, one is partly wrong, one is out of date and one is misleading. The misleading one is the sentence most readers will see first. On details below the blurb, the card is careful and accurate.
Where "Half the Cost" Came From
The blurb’s claim did not come from nowhere. It is a shortened version of SpaceXAI’s own tagline, and the shortening changes its meaning.
SpaceXAI’s exact words
The announcement’s subtitle reads: “Twice as fast, at half the price of comparable models.” The same page says Grok 4.7 is “served at the same price and speed as Grok 4.6.” So SpaceXAI is not claiming Grok 4.7 is cheaper or faster than its predecessor. It is claiming it is cheaper than competitors of similar ability.
What the card did
The card’s blurb reads “twice as fast and half the cost,” without “of comparable models.” A reader will naturally assume the comparison is with Grok 4.6, the previous version. That is the one comparison SpaceXAI explicitly rules out. The card’s own review headline, “the same old price,” contradicts its blurb.
Which models SpaceXAI meant
SpaceXAI’s launch table compares Grok 4.7 with Grok 4.6, GPT-5.6 Sol and Claude Fable 5.1, listing their token prices. GPT-5.6 Sol cost $4 input and $20 output per million tokens. Grok 4.7 costs $2 and $6. Against GPT-5.6 Sol, “half the price” is accurate on input.
Why it expired within a day
On 22 September, the day after Grok 4.7 launched, OpenAI released GPT-6 Sol at $2 input and $10 output. Grok 4.7’s input price is now the same as Sol’s, not half of it. Its output is still 40% cheaper. SpaceXAI’s comparison was true for about 24 hours.
“Twice as fast” has no figure
SpaceXAI’s announcement text gives no speed measurement to support “twice as fast.” The only doubling in its pricing is the Fast variant, which offers “twice the output speed at twice the price.” Readers should treat the speed claim as unsupported until independent measurements appear.
Grok 4.7 Pricing Against the Models That Followed
The card’s pricing section is accurate as far as it goes. It leaves out the price tier that matters most for long jobs.
The 200K threshold
SpaceXAI’s pricing page lists two rows for Grok 4.7. Below 200,000 prompt tokens, it costs $2 input, $0.50 cached input and $6 output. At 200,000 prompt tokens or more, it costs $4, $1 and $12, and “requests whose prompt reaches the listed token threshold are billed at the higher rate for all tokens in the request.” The card never mentions this.
Identical to Grok 4.6 in every row
SpaceXAI’s pricing and release notes show Grok 4.6 with exactly the same rates in both rows. Grok 4.5 launched at the same $2 and $6. Three successive Grok models have held one price, which is why SpaceXAI chose to compare with rivals rather than with itself.
What three job sizes cost
Price thresholds differ by company: 200,000 prompt tokens for Grok 4.7, 272,000 for GPT-6 Sol, and none for Claude Opus 5.5. That makes the cheapest model depend on the size of the job. The table works through three examples at each company’s published rates.
| Job size | Grok 4.7 | GPT-6 Sol | Claude Opus 5.5 |
|---|---|---|---|
| 150K input, 15K output | $0.39 | $0.45 | $0.90 |
| 250K input, 25K output | $1.30 (long rate) | $0.75 | $1.50 |
| 400K input, 40K output | $2.08 (long rate) | $2.20 (long rate) | $2.40 |
The middle band
For a 250,000-token job, Grok 4.7 has already crossed its 200,000 threshold while GPT-6 Sol has not crossed its 272,000 one. In that band, Grok 4.7 costs $1.30 against Sol’s $0.75, about 73% more. The card’s “half the cost” is not just out of date; for mid-sized jobs, the order is reversed.
The Benchmarks, Including the Ones the Card Skips
SpaceXAI published a table of seven benchmarks against Grok 4.6, GPT-5.6 Sol and Claude Fable 5.1. The card quotes five, accurately, but without the rival columns.
| Benchmark | Grok 4.7 | Grok 4.6 | GPT-5.6 Sol | Fable 5.1 |
|---|---|---|---|---|
| CursorBench 4.0 | 46.3% | 40.4% | 41.7% | 51.8% |
| DeepSWE v1.1 | 71.0% (high effort) | 65.2% | 72.7% | 70.0% |
| EEBench (not on card) | 64.0% | 53.0% | 39.4% | 56.4% |
| AA Briefcase v1.1 (not on card) | 1,657 | 1,546 | 1,487 | 1,678 |
| Terminal-Bench 4.0 | 37.6% | 20.3% | 37.3% | 57.9% |
| Harvey Legal Agent Benchmark | 19.6% | 15.8% | 2.5% | 6.7% |
| HealthBench Professional | 56.7% | 48.5% | 60.5% | 62.1% |
Where Grok 4.7 leads
Grok 4.7 improves on Grok 4.6 in every row, often sharply: Terminal-Bench 4.0 nearly doubles, from 20.3% to 37.6%. It leads all four columns on EEBench, an electrical engineering test, though a separate SpaceXAI chart shows GPT-6 Astra higher at 69.3%, and it leads on Harvey’s legal agent benchmark. SpaceXAI also reports an entirely new safeguard stack, with a top score of 62.4% on LatchBio’s biosafety benchmark and only 3.3% of risky prompts getting through on its own HackerBench v0.3.
Where it trails
GPT-5.6 Sol, the older OpenAI model, beats Grok 4.7 on DeepSWE and HealthBench. Claude Fable 5.1 beats it on CursorBench, Terminal-Bench 4.0, AA Briefcase and HealthBench. The card’s statement that “the three do not sit the same exams” is fair, but its benchmark table shows only Grok 4.7’s own column.
The effort setting
SpaceXAI’s table heads the Grok 4.7 column “xHigh,” but its DeepSWE score carries an asterisk and the note “high effort.” The card says all its scores were “measured at the xhigh effort setting.” For DeepSWE, that is not what SpaceXAI published.
Musk's Forecasts Against the Result
Elon Musk made a series of public predictions about Grok 4.7 before launch. We recorded them in our earlier coverage, and the launch lets us check them.
“Roughly on par with Opus 5.0”
On 14 September Musk wrote that Grok 4.7 “should be roughly on par with Opus 5.0, not 5.1.” On CursorBench 4.0, Grok 4.7 scores 46.3% in SpaceXAI’s table and Anthropic reports Opus 5 at 46.6%. On Artificial Analysis’s index, read on 24 September, Grok 4.7 scores 46.4 at xhigh against Opus 5’s 50.8. On Terminal-Bench 4.0 the gap is wider: 37.6% against Anthropic’s 52.3% for Opus 5. Different companies ran these tests, so the comparison is rough, but Musk’s forecast holds on coding and falls short on terminal work.
The release dates
The dates were less accurate. On 24 July Musk said Grok 4.7 would arrive in about four weeks, which pointed to around 21 August. On 12 August he said “3 to 4 weeks,” or 2 to 9 September. On 2 September he said it “comes out in 10 days,” or 12 September. It launched on 21 September.
What caused the slip
Musk explained the delay on 11 September: the model “still gives up on hard tasks (that it can do!) too early.” SpaceXAI’s launch post says Grok 4.7 now “works longer on difficult tasks” and “checks its own work more carefully.” The fix and the launch message line up.
The prediction markets
Kalshi’s contract on a release before 18 September settled No. Its contract on a release before 25 September settled Yes, after trading at 72 cents when we last looked on 17 September. For our full register of the pre-launch forecasts, see our earlier Grok 4.7 coverage. SpaceXAI’s next model is already the subject of a report we covered in Grok 4.8 to Feature 2.5T Parameters and New C++ Software Stack.
The Rank Badge on the Grok 4.7 Card
The card is badged “#4 in LLM models,” linked to AIxploria’s category page.
The slot belongs to its predecessor
On 24 September, position four in the LLM models category was Grok 4.6, and Grok 4.6’s own card also says “#4.” The successor claims its predecessor’s place, while the new Grok 4.7 card does not appear in the first 36 positions at all.
Three cards, one position
The Claude Opus 5.5 card, published two hours after Grok 4.7’s, also claims “#4.” Three cards now say they hold the same position. We checked six cards at known positions, from #1 to #10, and all six badges matched their slots, so the badge is normally a reliable index. The three new cards are the exception.
The ranked lists have not caught up
Grok 4.7 returned zero mentions on AIxploria’s Top 100 and Ultimate List on 24 September. Grok 4.6, launched in August, returned 16 and 6. On the Top 100 page, those mentions sit in the trending panel at the top rather than the ranked list. Expect Grok 4.7 to take at least a week to follow.
What the Card Leaves Out
Beyond the blurb, the card’s main weakness is what it does not mention.
No batch discount
SpaceXAI’s pricing page lists a 20% batch discount for Grok 4.3 and the Grok 4.20 variants only. “Models not listed above have no batch discount.” Grok 4.7 is not listed. Teams used to discounted overnight processing should budget at full rates.
Tool calls cost extra
Web search, X search and code execution, which the card lists as standard, are billed separately. SpaceXAI charges $5 per 1,000 web searches, $5 per 1,000 code execution calls, and $5 per 1,000 X posts or $10 per 1,000 profiles returned by X search.
Caching advice
SpaceXAI “highly” recommends setting a prompt_cache_key so requests in a conversation reach the same server. Without it, the documentation warns, “you often pay full input price on a cache-cold server.” Cached input costs $0.50 against $2.00, so this single setting can cut input costs by 75%.
The US regional endpoint
Grok 4.7 is served at a US-only endpoint that keeps inference in the United States, priced at a 10% premium. For organisations with data residency requirements, that is a relevant detail the card does not include.
Should You Try Grok 4.7?
The card’s verdict is that “the math practically writes itself” at $2 per million input tokens. That is true for short jobs and misleading for long ones.
Good fits
Grok 4.7 is a strong option for coding agents working on prompts below 200,000 tokens, where it is the cheapest of the three models launched this week. Teams already using Cursor or Grok Build can try it at no extra effort, and Grok Build’s free tier makes a first test cheap. We explained how the coding agent works in How to Use Grok Build.
Poor fits
For jobs between 200,000 and 272,000 prompt tokens, GPT-6 Sol is cheaper. For the hardest terminal and agentic work, SpaceXAI’s own table shows Claude Fable 5.1 well ahead, and Anthropic’s figures put Claude Opus 5.5 ahead too.
Test on your own work
All three companies publish different benchmarks under different conditions. The fairest comparison is to run your own tasks through each model and measure both results and total cost. Our AI strategy team helps businesses set up exactly that kind of evaluation.
Frequently Asked Questions About Grok 4.7 and SpaceXAI
What is SpaceXAI?
SpaceXAI is the company that develops Grok, formed when SpaceX and xAI combined. It publishes its announcements and developer documentation under the SpaceXAI name.
When was Grok 4.7 released?
SpaceXAI announced Grok 4.7 on 21 September 2026, one day before Claude Opus 5.5 and GPT-6 Sol and Luna.
How much does Grok 4.7 cost?
$2 per million input tokens and $6 per million output tokens below 200,000 prompt tokens, and $4 and $12 at or above that threshold, the same as Grok 4.6. The Fast variant costs twice as much and is only available in Cursor and Grok Build.
Is Grok 4.7 really half the cost?
Not compared with Grok 4.6, which has identical prices. SpaceXAI’s claim was “half the price of comparable models,” meaning rivals such as GPT-5.6 Sol. GPT-6 Sol, released a day later, matches Grok 4.7’s input price.
Is the AIxploria rank badge accurate?
No. The card claims “#4 in LLM models,” but that position belongs to Grok 4.6, and Grok 4.7 is not in the first 36 places.
Has Grok 4.7 been independently tested?
Yes. Artificial Analysis scored it 46.4 on its intelligence index at xhigh effort by 24 September, ahead of Grok 4.6 at 44.3 but behind GPT-6 Sol, Claude Opus 5 and Claude Opus 5.5.
References
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
