Grok 4.7: The Agent Play Nobody Is Talking About

xAI shipped Grok 4.7 on September 21, 2026, and the most surprising thing about it is not what changed. It is what did not.

The model launches at the exact same price as Grok 4.6: $2 per million input tokens and $6 per million output tokens. In a year where every other frontier lab has raised rates, xAI held the line. That matters for teams running agents at scale, where a few dollars per million tokens is the difference between a profitable workflow and a budget hemorrhage.

What xAI Actually Shipped

Grok 4.7 is not a post-training patch. xAI says it is built on a new, larger base model, trained with a longer reinforcement-learning run weighted toward tasks that take many hours to complete. The company also claims improved self-verification and better native understanding of its Grok Bot agent harness.

The headline specs are straightforward: 500,000-token context window, May 2026 knowledge cutoff, text and image in, text out. Four reasoning levels from Low to XHigh, with High as the default. A "Fast" variant exists at twice the speed for twice the price, though it is only available through Cursor and Grok Build, not the public API.

Where you can actually use it today matters. Grok 4.7 is live in Cursor on all plans, rolling out in GitHub Copilot, and available through the xAI API. There is also a free entry point via Grok Build, though it does not include the Fast tier.

The Benchmarks Tell a Complicated Story

On xAI's own comparison table, Grok 4.7 beats Grok 4.6 on every row the company published. The biggest jump is on Terminal-Bench 4.0, where the new model scored 38.0 percent against 4.6's 20.3 percent. That is a massive leap for agentic coding workflows involving terminal commands, repositories, and multi-step execution.

Other gains are more modest but still meaningful: CursorBench 4.0 climbed from 40.4 to 46.3 percent, EEBench from 53.0 to 64.0 percent, and the Harvey legal-agent benchmark from 15.8 to 19.6 percent.

But the frontier competition did not stand still. On the independent Artificial Analysis Intelligence Index, Grok 4.7 scores 46 and lands mid-pack. Claude Fable 5.1 and GPT-6 lead with 53 each. On CursorBench specifically, Fable 5.1 Max still edges out Grok 4.7 even at XHigh effort. And on GDPval and HealthBench, GPT-5.6 Sol Max remains ahead.

Elon Musk claimed on August 12 that Grok 4.7 would "exceed all current models." He walked that back on September 14 to "roughly on par with Opus 5.0." The actual numbers land somewhere between those two statements. xAI's own launch post calls the model "highly competitive in its class," which is more honest than the founder's initial hype and more accurate than his later downgrade.

The Real Story Is About Agents, Not Benchmarks

Where Grok 4.7 gets interesting is not in synthetic scores but in what xAI optimized for. The training emphasis on long-running tasks, self-checking, and agent-harness understanding points to a bet that the next frontier is not better chat. It is better autonomous work.

Short-prompt benchmarks do not capture whether a model can inspect a codebase, implement a change, run tests, investigate failures, and iterate until the job is done. They do not measure whether a research assistant keeps citations attached to the right claims across a hundred-page document. They do not test whether a document task ends with an editable file that meets the brief.

Those are the tasks Grok 4.7 is supposedly built for. The Terminal-Bench leap suggests xAI may be onto something. Whether that translates to real workflows is the question developers will answer over the next few weeks.

Pricing as Strategy

xAI's decision to hold pricing steady while competitors charge premium rates is a genuine differentiator. Claude Opus 5 runs $5 input and $25 output per million tokens. Fable 5.1 Max is even steeper at $10/$50. Grok 4.7 undercuts both by a wide margin.

For teams building agentic systems that burn through millions of tokens per day, that cost gap is not theoretical. It is the difference between a sustainable product and a VC-subsidized demo. If Grok 4.7 is good enough for the task, the price advantage makes it the rational choice.

The question, of course, is whether "good enough" holds across enough use cases to make xAI a credible alternative to Anthropic and OpenAI. The benchmarks say it depends on the task. The market will say whether that is enough.

Safety and Access

xAI also published its best safety numbers to date: 62.4 percent on LatchBio's biosafety benchmark and only 3.3 percent of risky dual-use prompts slipping through on the company's HackerBench v0.3. Red-team access is invite-only for cybersecurity partners, which is a narrower program than Anthropic's or OpenAI's broader researcher access.

That is a step forward, but the gap in transparency remains. Anthropic recently published detailed misalignment reports. OpenAI disclosed model hacking incidents. xAI's numbers are welcome, but numbers without narrative context are harder to evaluate.

The Bottom Line

Grok 4.7 is a solid upgrade that improves where it counts for agentic workloads, keeps pricing aggressive, and avoids the feature-bloat trap. It is not the frontier leader Musk promised, and the 31-day delay from his first target date continues a pattern of optimistic projections. But it is a credible competitor in the coding-and-agents space, and the value proposition is real.

For developers, the right move is probably to test it on your actual workflows rather than trust the benchmarks. The Terminal-Bench jump is exciting, but your codebase is not Terminal-Bench. If Grok 4.7 can finish the jobs you care about at one-third the cost of Claude, that is the story. Everything else is just marketing.