xAI has launched Grok 4.7, a new frontier model aimed explicitly at coding, agentic tasks and professional knowledge work. The company says the model uses a larger base than Grok 4.6, was post-trained on a harder mix of long-running tasks, manages extended context more effectively and checks its own work more carefully before committing to an answer or action.
The most consequential change for agentic research may be less visible than the benchmark charts. In its official September 21 announcement, xAI says Grok 4.7 was trained to natively understand the Grok Bot harness, the persistent-agent environment the company launched in August. Rather than treating an agent framework as an external wrapper around a general model, xAI is increasingly training the model around the environment in which persistent agents actually operate.
Grok 4.7 is available through the xAI API, Grok Build, Cursor, third-party coding harnesses, model routers and cloud platforms. Standard pricing starts at $2 per million input tokens and $6 per million output tokens, although the exact API rate changes for very long prompts and specialized serving options.
The model is being trained for work that lasts hours, not just prompts
xAI says Grok 4.7 uses a new and larger base model than Grok 4.6 and received a longer reinforcement-learning run. The task mixture was deliberately weighted toward problems that can take many hours to complete.
That is an important change in training emphasis. Conventional chatbot evaluation often rewards a model for producing one strong response to one prompt. Persistent agents face a different problem: they must maintain a plan, use tools repeatedly, preserve state and continue making sensible decisions as a task evolves.
Long-running research is a good example. An agent may begin with a broad question, search the web, inspect documents, discover conflicting evidence, revise its hypothesis, execute code and return to earlier material before producing a final report.
The quality of the first response matters less than whether the model can stay coherent across that entire trajectory.
Grok 4.7 has a 500,000-token context window
xAI’s developer documentation specifies a 500,000-token context window for Grok 4.7. The model accepts text and image input and produces text output, with no text-output limit listed in the model overview.
Large context windows can help agentic research because the model may need to retain conversation history, retrieved evidence, tool outputs and intermediate work simultaneously. But raw context size is not enough. A model also needs to manage the information effectively rather than simply accepting it.
xAI specifically says Grok 4.7 improves long-context management. That is a vendor claim rather than an independently established guarantee, but it aligns with the model’s positioning around multi-hour coding and knowledge-work tasks.
For developers, long agent loops can also use context compaction, which xAI recommends for tool-heavy workloads.
Self-verification is becoming an agent capability, not just a reasoning trick
xAI says Grok 4.7 is better at verifying its own work. In a normal chatbot exchange, self-checking may mean reconsidering a calculation or reviewing a draft before responding.
Inside an agent, verification can be more operationally important. A model may need to inspect whether a file was actually created, whether a code change passes tests, whether a retrieved source supports a claim or whether a tool returned the expected data.
A persistent agent that confidently moves forward after an incorrect intermediate step can compound the error over dozens of subsequent actions. Verification therefore becomes a form of trajectory control.
xAI has not published evidence proving that Grok 4.7 eliminates those failures in real research workflows. The meaningful change is that verification is explicitly part of the model’s training and product positioning rather than an optional prompting pattern layered on afterward.
Native Grok Bot training connects the model to persistent-agent infrastructure
Grok Bot is xAI’s system for always-on agents that have their own computer, can work across applications and can continue operating without a user supervising every step. xAI launched it in August as a persistent digital-worker environment rather than a conventional chat session.
The company now says Grok 4.7 was trained to natively understand that harness. This suggests that the model’s post-training includes the conventions and operating patterns of the environment in which those agents plan, use tools and maintain longer-running work.
That is different from taking a general-purpose model and teaching it the agent framework only through a large system prompt. A model trained around the harness can potentially internalize some of its expected interaction patterns.
The result still depends on the quality of the surrounding orchestration system, tools and permissions, but the boundary between “model” and “agent framework” is becoming less clean.
Persistent agents change what long context means
For a chatbot, a long context window often means the ability to analyze a large document or maintain a lengthy conversation. For a persistent agent, context is closer to working memory.
The model may need to remember why a task was started, which sources were already inspected, what a user approved, which attempts failed and what remains unfinished. Tool outputs and intermediate artifacts can accumulate quickly.
xAI’s API documentation recommends using a prompt cache key so requests in the same conversation are routed consistently and cache hits remain reliable. It also notes that long agent loops benefit from context compaction.
Those details show that persistent-agent performance is not determined by context-window size alone. Routing, caching and memory compression become part of the practical system architecture.
Grok 4.7 can use web search, X search and code execution through the API
xAI lists function calling, web search, X search and code execution among the tools available to Grok 4.7. These capabilities make the model directly relevant to research agents that need to move beyond static pretrained knowledge.
A web-research workflow can use search to discover current sources, code execution to process or compare data and function calls to interact with external systems. The model then has to decide when each tool is necessary and how its results should affect the plan.
This makes the model’s agentic reliability more important than raw answer quality. Tool-enabled systems create additional failure modes: unnecessary searches, malformed calls, weak source selection, incorrect interpretation of tool output and premature conclusions.
Better verification and long-context management are therefore especially relevant when the model is allowed to act rather than only respond.
The headline $2/$6 price applies to standard short-context API use
xAI advertises Grok 4.7 at prices starting at $2 per million input tokens and $6 per million output tokens. The detailed release notes add an important qualification.
Below 200,000 prompt tokens, the standard API rate is $2 per million input tokens, $0.50 per million cached input tokens and $6 per million output tokens. Above 200,000 prompt tokens, those rates rise to $4, $1 and $12 respectively.
This matters for the exact workloads Grok 4.7 targets. A persistent research agent can accumulate substantial context over a long session, meaning the lowest headline token price should not be treated as the universal cost of every agent run.
Server-side tools can also have separate usage charges, so the total cost of a web-research agent includes more than model tokens.
Caching can materially change the economics of long agent loops
Long-running agents repeatedly send overlapping context back to the model. Without effective caching, the system can pay to process the same conversation history and instructions over and over.
xAI recommends setting a prompt_cache_key on the Responses API, or the equivalent conversation header for Chat Completions, so requests are routed to the same server and cache hits become more reliable.
The cached-input price below the 200,000-token threshold is $0.50 per million tokens, one quarter of the standard $2 input rate. For persistent sessions with large stable prefixes, that difference can materially affect cost.
Developers therefore need to evaluate Grok 4.7 not only by headline token price but by how their agent manages caching, context growth, compaction and tool usage over time.
Reasoning effort is configurable up to xhigh
Grok 4.7 supports low, medium, high and xhigh reasoning settings, with high listed as the default in xAI’s API documentation.
This gives developers another control over the tradeoff between speed, cost and deliberation. Routine extraction or classification steps may not require the same reasoning depth as a difficult research synthesis or debugging task.
Agent systems can potentially route different stages to different effort levels rather than running every step at maximum deliberation. The model’s standard token price does not by itself determine total workload economics because reasoning behavior affects how many tokens and tool interactions a task consumes.
xAI’s launch benchmarks frequently show Grok 4.7 at xHigh or high effort, so benchmark results should be interpreted together with the configuration under which they were measured.
xAI reports gains across coding and professional knowledge work
xAI presents Grok 4.7 as its most capable model for coding and knowledge work. On the company’s published benchmark table, Grok 4.7 improves over Grok 4.6 on CursorBench 4.0, DeepSWE v1.1, EEBench, AA Briefcase, Terminal-Bench 4.0, the Harvey Legal Agent Benchmark and HealthBench Professional.
The launch page also highlights GDPval and AA Briefcase as tests of professional tasks performed by people such as lawyers, nurses and financial analysts. xAI says Grok 4.7 improves over Grok 4.6 and performs competitively with other frontier models.
These results are useful indicators of the model’s intended capabilities, but the launch page is a vendor source. Performance in a specific organization’s research workflow can depend heavily on tools, prompts, orchestration and domain-specific evaluation.
The strongest way to use the benchmark data is as evidence of xAI’s training direction, not as proof that one model universally dominates every competing system.
The model is already available in several agent environments
Grok 4.7 is available in Cursor and Grok Build, through the xAI API, and through third-party coding harnesses, model routers and cloud platforms. xAI’s documentation also lists model gateways including OpenRouter, Vercel and Cloudflare.
In Grok Build, xAI identifies Grok 4.7 as the default model of the coding agent. Cursor makes the model available across its plans.
This distribution matters because agent models increasingly compete through the environments where work already happens. A developer may encounter Grok 4.7 inside a coding agent rather than through xAI’s consumer chat interface.
The model becomes infrastructure inside another workflow rather than a destination product users have to visit directly.
Grok 4.7 Fast trades price for output speed
xAI also offers Grok 4.7 Fast, described as the same model served on faster infrastructure. The variant provides roughly twice the output speed at twice the standard token rates.
The detailed documentation says Grok 4.7 Fast is available only in Cursor and Grok Build and is not exposed through the public xAI API.
This creates a latency-cost choice for interactive agent environments. Some coding and research tasks benefit from faster model output because human users are waiting for each step, while unattended persistent agents may care more about total task cost than second-by-second response speed.
Again, agent economics depend on the workload rather than one universal model configuration.
Long-context research still depends on retrieval discipline
A 500,000-token window can hold an enormous amount of material, but giving an agent more context does not automatically improve its research. Poorly selected evidence can consume context and distract the model just as easily as useful evidence can help it.
Research agents still need effective search, source selection and evidence management. They also need mechanisms to distinguish authoritative primary material from weak secondary pages and duplicated content.
Grok 4.7’s web-search capability and longer context provide capacity for broader investigation. Whether that produces better research depends on how the agent chooses what to retrieve and what to preserve.
The model’s stronger self-verification could help at that stage, but xAI has not published a citation-focused study demonstrating a specific improvement in source accuracy for Grok 4.7.
Native agent training may matter more than another benchmark point
The most interesting part of Grok 4.7 may therefore be architectural rather than numerical. xAI is training the frontier model with awareness of the persistent-agent harness in which it will operate.
As agents become longer lived, the model has to understand more than language. It has to work coherently inside a system of tools, memory, computers, permissions and asynchronous tasks.
That pushes model development toward environment-specific competence. The frontier model and the agent runtime begin to co-evolve instead of being designed independently.
If that pattern continues across the industry, evaluating an “AI model” in isolation will become less representative of the actual product users experience.
Grok 4.7 strengthens the layer that decides how research gets done
Grok 4.7 arrives with a larger base model, longer reinforcement learning on difficult multi-hour tasks, a 500,000-token context window, configurable reasoning and stronger xAI-reported self-verification. It is also explicitly trained to understand Grok Bot, the company’s persistent-agent harness.
For research workflows, those improvements target the layer that plans searches, remembers earlier evidence, uses tools and decides whether its work needs another verification pass. They do not guarantee better citations or eliminate the need for rigorous source selection, but they address several of the failure modes that become more important as an agent runs longer.
The starting API price of $2 per million input tokens and $6 per million output tokens makes the model economically relevant for agent builders, while the higher long-context rates and separate tool costs mean real task economics still need to be measured at the workflow level.
The broader shift is clear: frontier models are no longer being optimized only to answer difficult prompts. Grok 4.7 is being trained to inhabit the agent system itself—and to keep working, checking and remembering after the first answer would normally have ended.