What actually shipped with Grok 4.7
xAI released Grok 4.7 on 21 September 2026, according to MarkTechPost and The Decoder. It is positioned by xAI as its frontier model for coding, agentic tasks and knowledge work, rather than as a narrowly specialised coding release.
The model accepts text and images as input and has a 500,000-token context window, according to llm-stats.com. Access is available through the xAI API, Cursor and Grok Build. For developers already working through one of those routes, this is therefore an available model change rather than a product announcement without deployment options.
The release also keeps the standard token prices of Grok 4.6. That matters because the clearest published gains are in coding benchmarks, so buyers are being offered higher reported scores in those tests without a corresponding increase in the ordinary per-token rates.
The coding benchmarks show a real step up
The strongest case for Grok 4.7 comes from two coding measurements. On CursorBench 4.0, which tests longer-running coding tasks, Grok 4.7 scored 46.3%, compared with 40.4% for Grok 4.6. On DeepSWE v1.1 at high effort, it reached 71.0%, up from 65.2% for the previous model.
Those movements are meaningful within the boundaries of the benchmarks. They show Grok 4.7 performing better than Grok 4.6 on the specific coding tasks represented by CursorBench 4.0 and DeepSWE v1.1. If your main question is whether xAI’s newer model improved on its predecessor in these published coding evaluations, the answer from the reported numbers is yes.
What the scores do not establish is that every coding workload will improve by a similar amount. CursorBench 4.0 is specifically described as a benchmark for longer-running coding tasks, while the DeepSWE result is reported at high effort. Neither figure by itself tells you how Grok 4.7 behaves on your repository, toolchain, prompts or agent workflow.
The practical reading is narrower: there is evidence of progress in relevant coding tests, but buyers still need to distinguish benchmark improvement from guaranteed application-level improvement.
Grok 4.7 versus Grok 4.6 and the index leaders
Against Grok 4.6, Grok 4.7 has a straightforward advantage in the two disclosed coding benchmarks. CursorBench 4.0 rises from 40.4% to 46.3%, while DeepSWE v1.1 at high effort rises from 65.2% to 71.0%. The standard token prices remain unchanged, so the newer model offers those reported improvements without a higher ordinary input or output rate.
The broader comparison is less favourable. On the independent Artificial Analysis Intelligence Index version 4.3.2, which combines ten benchmarks, Grok 4.7 scores 46 and sits mid-pack. Claude Fable 5.1 and GPT-6 each score 53. The Decoder focused its coverage on this gap.
That contrast is useful because it prevents one type of benchmark from standing in for the whole model. Grok 4.7 can improve materially over Grok 4.6 on the published coding tests and still fall short of the leaders on a broader multi-benchmark index.
For developers, the implication is not that one result cancels out the other. They answer different questions. The coding numbers support the case that xAI improved areas it is explicitly targeting. The Artificial Analysis result shows that those gains do not amount to a documented overall benchmark lead.
What the pricing means when you buy tokens
For prompts below the long-context threshold, Grok 4.7 costs 2 US dollars per million input tokens and 6 dollars per million output tokens, with cached input priced at 0.50 dollars per million, according to llm-stats.com. Those prices are unchanged from Grok 4.6.
That continuity simplifies the upgrade calculation for an existing Grok buyer. If your workload stays within the standard pricing band, moving from 4.6 to 4.7 does not introduce a higher published unit price. You can therefore evaluate the newer model mainly on whether its output quality and task completion justify routing work to it.
The important qualification arrives at 200,000 tokens. Prompts at or above that level are billed at 4 dollars per million input tokens, 12 dollars per million output tokens and 1 dollar per million cached input. In each case, the rate is twice the standard figure.
That makes context length a cost variable, not merely a capability headline. Grok 4.7 can accept up to 500,000 tokens, but using very large prompts can move a workload into the higher pricing band. A buyer comparing models should therefore look at the prompt sizes actually sent in production, rather than treating the maximum context window as free capacity.
A large context window is not the same as better reasoning
The 500,000-token context window gives Grok 4.7 room for substantial text and image input. For coding and agentic work, that can be operationally useful when a task requires a large amount of source material or accumulated context.
But the published information does not show how performance changes as prompts approach the context limit. There is no disclosed benchmark here comparing Grok 4.7 at short and very long contexts, and no published figure in the available reporting for accuracy retention across the full 500,000-token window.
That distinction matters because context capacity describes how much input the model can accept, not how effectively it uses every part of that input. The higher rates at 200,000 tokens and above also mean that long-context use carries a direct billing consequence.
For buyers, the sensible test is therefore workload-specific. If you routinely need large prompts, the expanded window may be valuable enough to justify the higher long-prompt rates. If your prompts remain much smaller, the maximum context specification may have little effect on either your application design or your bill.
When switching to Grok 4.7 makes sense
The strongest switching case is for an existing Grok 4.6 user focused on coding, agentic tasks or knowledge work. Grok 4.7 is the model xAI now positions for those workloads, its two disclosed coding scores are higher, and standard token prices have not increased. For the consumer side of the stack, our Grok review covers the assistant itself.
Availability through the xAI API, Cursor and Grok Build also means developers have several disclosed routes for using it. The reporting does not disclose other access channels, so buyers should not assume broader availability beyond those named options.
A more cautious assessment applies if your decision depends on broad benchmark leadership. The Artificial Analysis Intelligence Index places Grok 4.7 at 46, while Claude Fable 5.1 and GPT-6 each score 53. That does not determine performance on your own task, but it does mean the available independent composite result does not support describing Grok 4.7 as the overall leader.
There is also no disclosed latency benchmark in the available figures. Teams for which response time is a primary routing criterion would need their own measurements rather than inferring speed from the release positioning.
In short, Grok 4.7 gives existing Grok users concrete reasons to test an upgrade, especially in coding. The evidence for replacing another model across all workloads is much less complete.
The wider week was about capital and control
Grok 4.7 arrived during a week in which two other developments showed how quickly the economics and governance of AI are moving. The same week also brought Anthropic’s cheaper Claude Opus 5.5, part of the same price-performance push.
Bloomberg reported on 21 September that SoftBank Group had launched a bond sale consisting of 10 billion US dollars plus 1 billion euros of senior unsecured notes, about 11.15 billion dollars in total. The sale spans five tranches with maturities from 3.5 to 7.5 years and is intended to fund the final 10-billion-dollar instalment of the 30 billion dollars SoftBank agreed in February 2026 to invest in OpenAI across April, July and October. Pricing was expected on 24 September and settlement on 29 September. Tech Startups and TechNode also covered the deal.
If completed as planned, the offering would be the largest non-financial corporate bond sale on record from the Asia-Pacific and Japan region, exceeding a 10.93-billion-dollar 7-Eleven deal from 2021. The scale provides context for the capital being committed around frontier AI, even as individual model buyers remain focused on token costs.
Governance was moving at the same time. On 21 September, the UN’s Independent International Scientific Panel on AI published its first thematic brief. As UN News reported, it invoked the precautionary principle and urged governments to put safeguards around AI agents before their risks are fully understood. Co-chair Yoshua Bengio said “the traditional model of safeguarding is unravelling”. The panel will feed into the Global Dialogue on AI Governance in May 2027.
For developers evaluating Grok 4.7, these events do not change its benchmark scores or price sheet. They do show the setting in which frontier models are now shipping: better task performance, vast financing commitments and growing pressure for safeguards are advancing at the same time.
FAQ
How much does Grok 4.7 cost?
According to llm-stats.com, standard pricing is 2 US dollars per million input tokens, 6 dollars per million output tokens and 0.50 dollars per million cached input. Prompts at or above 200,000 tokens are billed at 4 dollars input, 12 dollars output and 1 dollar cached input per million.
Is Grok 4.7 better than Grok 4.6 for coding?
On the two disclosed coding benchmarks, it scores higher. Grok 4.7 reaches 46.3% on CursorBench 4.0 versus 40.4% for Grok 4.6, and 71.0% on DeepSWE v1.1 at high effort versus 65.2%.
Does Grok 4.7 lead the major AI benchmarks?
Not on the disclosed Artificial Analysis Intelligence Index version 4.3.2. Grok 4.7 scores 46, while Claude Fable 5.1 and GPT-6 each score 53, placing Grok 4.7 mid-pack on that ten-benchmark index.
What is Grok 4.7’s context window?
Grok 4.7 has a 500,000-token context window and accepts text and images as input, according to llm-stats.com. Prompts at or above 200,000 tokens move into a higher pricing band.
Where can developers access Grok 4.7?
The disclosed access routes are the xAI API, Cursor and Grok Build. Other access channels have not been announced.
Is Grok 4.7 faster than Grok 4.6?
No latency or throughput benchmark is disclosed in the available reporting, so a measured speed advantage cannot be established from these figures. Developers for whom response time matters would need to evaluate that separately.
