SpaceXAI has launched Grok 4.7, a new frontier model aimed at coding, AI agents, and professional knowledge work. The API starts at $2 per million input tokens, while output is priced at $6 per million tokens.
Released on September 21, 2026, Grok 4.7 is available through the Grok API and is rolling out to Cursor, Grok Build, model routers, and other developer platforms. The launch keeps Grok 4.6's standard token rates while targeting harder, longer-running tasks.
There is an important catch, though: early independent testing suggests that Grok 4.7 can use substantially more reasoning tokens than its predecessor. The headline price is low; the cost of completing a real task may be a different story.
What is Grok 4.7?
Grok 4.7 is SpaceXAI's latest general-purpose model for coding, agentic workflows, and complex knowledge work. The company positions it less as a faster answer engine and more as a model that can stay with a difficult task, use tools, manage a large context, verify its work, and complete several steps in sequence.
The launch announcement describes the model as a coding and knowledge-work system with stronger long-context management and self-verification. SpaceXAI also says it used a larger base model, a longer reinforcement-learning run, and a training mix weighted toward problems that can take an agent hours rather than minutes.
That puts Grok 4.7 into one of the most competitive parts of the 2026 AI market: coding agents that can work inside repositories, call tools, and deliver a finished work product rather than just suggest the next line of code.
Grok 4.7 API pricing starts at $2 per million input tokens
The most attention-grabbing part of the release is the price. According to the official model documentation, standard requests below the long-context threshold are billed as follows:
| Usage | Grok 4.7 price |
|---|---|
| Input | $2 / 1M tokens |
| Cached input | $0.50 / 1M tokens |
| Output | $6 / 1M tokens |
Grok 4.7 supports a 500,000-token context window, allowing developers to provide large codebases, documents, or conversation histories in one context. Prompts above 200,000 tokens use the higher long-context rates: $4 per million input tokens, $1 per million cached input tokens, and $12 per million output tokens.
The model accepts text and image input and returns text. Developers can choose low, medium, high, or xhigh reasoning effort. It also supports function calling, web search, X search, and code execution, which makes it suitable for tool-using agents rather than conventional chat alone.
Grok 4.7 shows its biggest gains in coding
Coding is where the launch looks most convincing. In SpaceXAI's published evaluations, Grok 4.7 scored 46.3% on CursorBench 4.0, compared with 40.4% for Grok 4.6. On Terminal-Bench 4.0, it reached 38.0%, compared with 20.3% for its predecessor.
Terminal-style evaluations matter for autonomous coding agents because they test interaction with a working environment and practical task completion, not only answers to isolated programming questions. SpaceXAI also reported gains on DeepSWE, AA-Briefcase, EEBench, the Harvey Legal Agent Benchmark, and HealthBench Professional.
Those are vendor-reported results, so they should be read as company claims rather than neutral rankings. Fortunately, independent testing is already available.
Independent testing puts Grok 4.7 near the frontier
Artificial Analysis tested Grok 4.7 at xhigh reasoning effort shortly after launch. It scored 46 on the Artificial Analysis Intelligence Index, two points above Grok 4.6.
The larger improvement appeared in agentic work. On AA-Briefcase, a benchmark for long-horizon professional tasks, Grok 4.7 reached 1,657 Elo, gaining 111 points over Grok 4.6 at high effort. With Grok Build, its first-party coding agent, it scored 56 on the Artificial Analysis Coding Agent Index, compared with 47 for Grok 4.6.
| Benchmark | Grok 4.7 | Grok 4.6 |
|---|---|---|
| DeepSWE v1.1 | 73% | 65% |
| Terminal-Bench 4.0 | 33% | 18% |
| SWE-Atlas-QnA | 63% | 58% |
| Coding Agent Index | 56 | 47 |
Artificial Analysis placed Grok 4.7 with Grok Build fourth among the native coding-agent systems it had tested at launch. The harness matters here: these scores describe a model plus its coding environment, not necessarily the result a developer will get from a bare API call.
The catch: Grok 4.7 uses many more reasoning tokens
The $2 input rate sounds exceptionally cheap for a frontier model, but token price is only half of the cost equation. The other half is how much the model consumes to finish a task.
Artificial Analysis measured roughly 81,000 output tokens per Intelligence Index task for Grok 4.7 at xhigh effort, compared with about 36,000 for Grok 4.6 at high effort. The evaluator also measured approximately 7.1 minutes per task and output speed near 188 tokens per second on long prompts.
That leads to a useful rule for comparing reasoning models: price per million tokens is not the same thing as cost per completed task. A model with a low headline rate can still cost more if it needs several times as many reasoning tokens. Developers building high-volume coding agents should measure total task cost, latency, retry rate, and successful completion on their own workloads.
Grok 4.7 Fast is available in Cursor and Grok Build
SpaceXAI is also offering Grok 4.7 Fast, the same underlying model on faster inference infrastructure. It costs roughly twice as much: $4 per million input tokens, $1 per million cached input tokens, and $12 per million output tokens.
At launch, Grok 4.7 Fast was available through Cursor and Grok Build rather than the public SpaceXAI API. That split makes the choice fairly practical: standard Grok 4.7 is more attractive for autonomous jobs where latency matters less, while Fast may make sense for interactive coding sessions where waiting is expensive.
Where can you use Grok 4.7?
The API model name is grok-4.7. It supports the Responses API and Chat Completions API, along with function calling, web search, X search, and code execution. SpaceXAI's release notes list availability through the Grok API, Cursor, Grok Build, third-party coding harnesses, model routers, and selected cloud platforms.
Cursor supports Grok 4.7 across its desktop application, web and cloud agents, CLI, SDK, automations, and iOS app. Availability and routing can change quickly, so developers should verify the live provider catalog before building a production dependency around one integration.
Grok 4.7 is increasingly an agent model, not just a chatbot
The larger story is the direction SpaceXAI is taking Grok. The launch repeatedly emphasizes long-running work: modifying a repository, researching across sources, producing a professional document, analyzing a dataset, or calling tools until a multi-step task is finished.
That is a different question from “Which chatbot gives the best answer?” It is closer to “Which model can complete the work, verify the result, and recover when the first attempt fails?” Grok 4.7's improvements in coding agents and long-horizon professional benchmarks point in that direction.
Is Grok 4.7 worth testing?
For developers, Grok 4.7 combines an attractive set of specifications: frontier-oriented agent performance, a 500K context window, multimodal input, tool use, and a starting API price of $2 per million input tokens. Early independent testing also suggests that the improvement over Grok 4.6 is real, especially in coding agents and long-horizon knowledge work.
The pricing story still needs nuance. Grok 4.7 is inexpensive per token; it is not guaranteed to be inexpensive per task. At higher reasoning settings it can consume significantly more tokens and take longer to finish. The sensible test is to run your own representative tasks and record success rate, total tokens, latency, retries, and output quality.
That makes Grok 4.7 an interesting model to benchmark in 2026—not because one headline price settles the comparison, but because it puts serious agentic capability and a very large context window into a relatively aggressive pricing tier.
Grok 4.7 FAQ
When was Grok 4.7 released?
SpaceXAI announced Grok 4.7 on September 21, 2026.
How much does the Grok 4.7 API cost?
Standard pricing starts at $2 per million input tokens, $0.50 per million cached input tokens, and $6 per million output tokens. Requests above the 200,000-token prompt threshold use higher long-context rates.
What is the Grok 4.7 context window?
Grok 4.7 supports a maximum context window of 500,000 tokens.
Is Grok 4.7 better than Grok 4.6?
SpaceXAI's benchmarks and early independent testing show improvements in several areas, with particularly large gains in coding agents and long-horizon professional tasks. The trade-off is substantially higher token consumption at xhigh reasoning effort.
Can Grok 4.7 process images?
Yes. Grok 4.7 accepts text and image input and produces text output.
Is Grok 4.7 available in Cursor?
Yes. It is available in Cursor, including its desktop app, web and cloud agents, CLI, SDK, and automations.
What is Grok 4.7 Fast?
Grok 4.7 Fast uses the same model on faster inference infrastructure at approximately twice the token rates. At launch, it was offered through Cursor and Grok Build rather than the public API.
What is Grok 4.7 designed for?
SpaceXAI is positioning it for coding, AI agents, and professional knowledge work, especially tasks that require longer reasoning, tool use, context management, and multiple steps.