Grok 4.7: How True Is “Twice as Fast, Half the Price”? Where It Really Stands on Price, Overall Scores, and Specific Fields
Grok 4.7 charges one-fifth of GPT-6 Astra’s price per input token. Yet when an independent lab ran both models through the same test suite, Grok 4.7 cost more per task. xAI launched Grok 4.7 on September 21, calling it “twice as fast, at half the price of comparable models.” The low prices and the strong legal and electrical-engineering scores are real, but its independent overall score is 7 points below the two leading models it was compared with (GPT-6 Astra and Claude Fable 5.1). Here is how the verdict shifts with the yardstick you pick.
Public-source analysis · Edited by Dandan Dad · Editorial policy
- Grok 4.7
- xAI
- OpenAI
- Anthropic
- xAI introduced Grok 4.7 as “twice as fast, at half the price of comparable models.” Its API token prices are one-fifth of GPT-6 Astra’s for input and about one-eighth for output. But it uses so many more tokens that an independent lab measured a higher cost per task than Astra.
- In xAI’s own results table, it ranks first among the compared models on legal (Harvey) and electrical-engineering (EEBench) tasks. On the independent Artificial Analysis Intelligence Index, however, it scores 46, below Astra and Claude Fable 5.1 (53 each), and its measured output speed is about 1.4x theirs.
- To pick the right model, start with the kind of work you do and what a mistake costs you, not a single leaderboard.
What xAI Claims for Grok 4.7
xAI’s pitch boils down to one idea: more performance, same price. Price and speed match its predecessor, Grok 4.6, at $2 per million input tokens and $6 per million output tokens (for prompts under 200K tokens). The “twice as fast, at half the price” line compares Grok 4.7 with similar models from other companies, not with Grok 4.6.
input price
output price
(Astra 53)
What the unit-price gap means in dollars
AI APIs charge by the number of tokens, the units in which models process text. For the same number of input tokens, input that costs $1,000 on GPT-6 Astra costs $200 on Grok 4.7. For the same number of output tokens, $1,000 of output corresponds to about $120 ($6 ÷ $50, roughly 88% lower).
The catch is the phrase “for the same number of tokens.” Models use different numbers of tokens and retries to finish the same job, so a price that is one-fifth does not buy five times the work. If you run a service that processes tens of thousands of requests a day, measure real token usage on your own tasks first.
What changed from Grok 4.6
According to xAI, four things changed.
- A larger base model: Grok 4.7 uses a new, larger base model than Grok 4.6.
- Training on long tasks: xAI ran reinforcement learning longer, weighted toward problems that take hours to finish.
- Better self-checking: xAI says the model verifies its own work more carefully and manages long context better.
- Native Grok Bot support: The model was trained to understand Grok Bot, xAI’s agent environment, from the start. xAI says this makes it better at conversational tasks and general knowledge work.
xAI’s Own Results: First in Legal and Electrical Engineering
The table below shows the numbers xAI published in its announcement. xAI also compiled the scores for the other models, so they may differ from what those companies report themselves.
| Area / Benchmark | Grok 4.7 xHigh | Grok 4.6 | GPT-5.6 Sol | Fable 5.1 |
|---|---|---|---|---|
| Input price ($/M) | $2 | $2 | $4 | $10 |
| Output price ($/M) | $6 | $6 | $20 | $50 |
| CursorBench 4.0 (coding) | 46.3% | 40.4% | 41.7% | 51.8% |
| DeepSWE v1.1 (coding) | 71.0%* | 65.2% | 72.7% | 70.0% |
| EEBench (electrical engineering) | 64.0% | 53.0% | 39.4% | 56.4% |
| AA Briefcase v1.1 (multi-hour office work) | 1,657 | 1,546 | 1,487 | 1,678 |
| Terminal-Bench 4.0 (terminal work) | 37.6% | 20.3% | 37.3% | 57.9% |
| Harvey Legal Agent | 19.6% | 15.8% | 2.5% | 6.7% |
| HealthBench Professional (clinical reasoning) | 56.7% | 48.5% | 60.5% | 62.1% |
* high effort setting · Source: xAI announcement (Sep 21, 2026)
Where it clearly leads
Grok 4.7 is clearly ahead in electrical engineering (EEBench 64.0%) and legal agent work (Harvey 19.6%). Its Harvey score is about 8x GPT-5.6 Sol’s (2.5%) and about 3x Fable 5.1’s (6.7%). For people who work in these fields, that gap may matter more than any overall ranking.
Where it trails in the same table
Even in xAI’s own table, Fable 5.1 leads on coding (CursorBench 4.0), multi-hour terminal work (Terminal-Bench 4.0), and clinical reasoning (HealthBench Professional). On terminal work, the gap between Grok 4.7 (37.6%) and Fable 5.1 (57.9%) is more than 20 percentage points.
Independent Overall Score: 46 vs 53
A company’s results table only includes the tests the company chose. That is why it helps to look at an outside evaluator that runs many models under the same conditions. On the Intelligence Index from Artificial Analysis, which combines ten benchmarks, Grok 4.7 scores 46. Claude Fable 5.1 and GPT-6 Astra score 53 each.
A 7-point gap is not small, and the top of the chart has only gotten more crowded since. On the same index, checked October 6, Claude Opus 5.5 (released September 22) scores 58 and GPT-6.1 Sol (released September 29) scores 52. By this measure alone, it is hard to call Grok 4.7 a top-tier model.
Same test, different scores: Terminal-Bench 4.0
- xAI announcement: Grok 4.7 at 37.6%
- Artificial Analysis measurement (all three at the xhigh setting): Grok 4.7 at 26%, GPT-6 Astra at 60%, Claude Fable 5.1 at 55%
Even on a test with the same name, results can vary widely with the test harness, settings, and number of retries.
A company’s benchmark table is a bit like a report card that only shows the subjects you did well in. It isn’t a lie, but read it knowing that some subjects were left out.
The benchmarks each company chose to highlight
Put the Grok 4.7 and GPT-6 Astra announcements side by side, and the first difference is which tests each company chose.
- xAI highlights multi-hour office work (AA Briefcase) and professional work quality (GDPval Elo).
- OpenAI highlights work automation (AutomationBench) and an agent exam (Agents’ Last Exam).
- xAI also included an unfavorable number. Its table shows a Terminal-Bench score for Grok 4.7 that is lower than Fable 5.1’s (57.9%).
As DataCamp’s comparison points out, neither announcement shows its own model losing on the other company’s chosen benchmarks. That makes it hard to rank the two models from their announcements alone.
37.6% in xAI’s announcement, 26% in an independent test. A company’s own number and an outside measurement don’t always agree.
Terminal-Bench 4.0 · xAI announcement vs Artificial AnalysisPrice: One-Fifth per Token, a Different Story per Task
| Input ($/M) | Output ($/M) | vs Grok 4.7 | |
|---|---|---|---|
| Grok 4.7 | $2 | $6 | baseline |
| GPT-6 Astra | $10 | $50 | 5x input, ~8.3x output (more expensive) |
| Claude Fable 5.1 | $10 | $50 | 5x input, ~8.3x output (more expensive) |
On price per token, Grok 4.7 is clearly cheaper. For high-volume or cost-sensitive automation, that gap can be decisive. There is also a fast variant, Grok 4.7 Fast, with twice the output speed at twice the price, but according to xAI’s docs it is available only in Cursor and Grok Build, not on the public API.
The problem is that price per token is not the cost of getting a job done. OpenAI has argued that Astra finishes the same work with fewer tokens, and independent data points the same way. Running its Intelligence Index, Artificial Analysis measured an average cost per task of $3.74 for Grok 4.7, higher than GPT-6 Astra’s $3.26 (checked October 6). Grok 4.7 generated 240 million output tokens on that test suite, about three times the median of comparable models (81 million). Cheap tokens still add up to a big bill when you use a lot of them.
The price race has moved, too. xAI’s table lists GPT-5.6 Sol at $4 input and $20 output, but GPT-6.1 Sol, which OpenAI launched at DevDay on September 29, costs $2 and $10. The same lab measured its cost per task at $0.72, about one-fifth of Grok 4.7’s, and its input price matches Grok 4.7’s.
When a cheaper model fits, and when it doesn’t
For high-volume work where a single mistake is cheap, such as a million classification jobs a month, a model with 80 to 88% lower token prices can be the rational choice even if its quality is a bit lower. For work where one mistake is expensive, such as reviewing legal documents, the quality gap is what matters.
Change the Criterion, Change the Verdict
A few years ago, spotting a frontier model, one that pushes the limits of performance, was easy: pick the latest from OpenAI, Anthropic, or Google. Grok 4.7 blurs the line: its independent overall score is 7 points below the two leading models it was compared with, its token prices are far lower, and it ranks first among the compared models on some of the tests xAI published.
| Criterion | Grok 4.7 | Verdict |
|---|---|---|
| Overall performance (Artificial Analysis Index) | 46 (GPT-6 Astra and Claude Fable 5.1: 53 each) | 7 points below the two compared leaders |
| Token price | One-fifth of Astra’s input price ($2/$10); 12% of its output price ($6/$50) | Lowest price among the models in xAI’s table (same as Grok 4.6) |
| Specific fields (legal, electrical engineering) | Harvey 19.6% (GPT-5.6 Sol 2.5%) | #1 (per xAI) |
A grouping for this article: three kinds of frontier models in 2026
This is not a standard industry classification. It is this article’s way of grouping the three criteria in the table above.
- Absolute frontier (best overall): Claude Opus 5.5, GPT-6 Astra, Claude Fable 5.1. Suited to the hardest tasks and work where mistakes are costly.
- Practical frontier (low prices with solid performance): Grok 4.7, GPT-6.1 Sol, the DeepSeek V4 family, the Gemini Flash family. Suited to high-volume automation where cost control matters.
- Domain frontier (best in a specific field): Grok 4.7 (legal, electrical engineering) and models fine-tuned for a field. Suited to practitioners in that field.
By this article’s grouping, Grok 4.7 is a practical frontier model and a frontier model in some domains. It is not an absolute frontier model, but it is also hard to say flatly that it isn’t a frontier model.
Safety: xAI’s New Safeguards
xAI says Grok 4.7 ships with an entirely new safeguard stack. Everything below comes from xAI’s own announcement.
- HackerBench v0.3 (xAI’s own benchmark): xAI says the model lets only 3.3% of risky dual-use cyber prompts through while rarely blocking legitimate security work.
- LatchBio biosafety benchmark: xAI reports a first-place score of 62.4%.
- Red-team access: xAI has started giving select cybersecurity partners invite-only access to Grok 4.7’s red-team (attacker-role) capabilities for defense research.
DataCamp notes that “safe” means something different in each company’s pitch. xAI emphasizes refusing dangerous requests up front, while OpenAI emphasizes alignment: Astra staying within set limits even while it works on its own.
So, When Should You Choose Grok 4.7?
So how much of “twice as fast” holds up? It is xAI’s claim relative to comparable models, and “speed” hides three different things: how many tokens a model produces per second (output speed), how long you wait for the first answer token (TTFT, including thinking time), and how long a whole task takes (including tools, retries, and review). The Artificial Analysis page for Grok 4.7 (Xhigh) showed 88.2 output tokens per second when checked on October 6, while the comparison chart on the same page showed 63 for GPT-6 Astra and 64 for Claude Fable 5.1. That is about 1.4x on output speed, short of 2x. And a model that uses more tokens can produce them faster and still finish later. How fast your own work gets done is something you have to measure.
Strengths
- API token prices one-fifth (input) and about one-eighth (output) of GPT-6 Astra’s; cost per task depends on token usage
- Far ahead of the compared models on legal and electrical-engineering tasks (per xAI)
- Better performance at the same price as Grok 4.6
Weaknesses
- 7 points below GPT-6 Astra and Claude Fable 5.1 on an independent index that combines ten tests
- Uses so many tokens that its independently measured cost per task is higher than Astra’s
- Fable 5.1 leads on coding, terminal work, and clinical reasoning; on terminal work, xAI’s figure (37.6%) and the outside measurement (26%) differ sharply
Work through these steps when choosing.
- Define the task first. Coding, document work, and classification call for different scores. An overall index is only a reference.
- Estimate the cost of mistakes. If fixing a wrong answer costs more than the API bill, the expensive model is actually cheaper.
- Test with your real data. Even 10 to 20 cases with your own inputs, codebase, and documents will show where a model stumbles on your work. Log the tokens used and the time taken alongside the results. With so few cases, don’t treat the result as a general ranking.
- Make models easy to swap. Rankings change within months. Code tightly tied to one model becomes a burden later.
Verification notes: numbers that differ between sources
| Item | What we checked (Oct 6, 2026) |
|---|---|
| Grok 4.7 on Terminal-Bench 4.0 (xAI) | xAI’s announcement page shows 37.6%. Some outlets, including DataCamp, quoted 38.0%; this article uses the figure on xAI’s page. |
| The 57.9% on Terminal-Bench 4.0 | In xAI’s table it is Claude Fable 5.1’s score. In OpenAI’s Astra announcement it is Astra’s own score, with Fable 5.1 at 55.8%. Because the two tables disagree, this article compares models using Artificial Analysis’s independent results at the xhigh setting (Astra 60%, Fable 5.1 55%, Grok 4.7 26%). |
| Harvey legal benchmark gap | 19.6% ÷ 2.5% ≈ 7.8x, stated here as “about 8x.” |
| “Twice as fast” | xAI’s comparison with similar models. Price and speed are the same as Grok 4.6. |
| Price comparison | Based on API list prices (xAI docs: $2 input and $6 output per 1M tokens for prompts under 200K tokens; $4 and $12 above that). Cost per task depends on how many tokens each model uses. |
| “5 to 8 times more work for the same budget”? | That turns a token-price ratio into a workload claim. The same budget buys more tokens; how many tasks that covers depends on tokens used per task and retries. |
| Output cost example | $6 ÷ $50 = 0.12, so $1,000 becomes about $120. The ratio is about 8.33x, which the text rounds to “about one-eighth.” |
| Overall rank | The number of models and settings on the Artificial Analysis index keeps changing, so this article avoids labels like “#N” or “mid-pack” and gives the score gap to the two models actually compared (GPT-6 Astra and Claude Fable 5.1, 53 each). |
