Grok 4.7's awkward problem is visible in two numbers.
Artificial Analysis scores the xhigh version at 46 on its Intelligence Index. That is good.
The same testing shows the model using about 81,000 output tokens per task. Grok 4.6 at high reasoning used about 36,000. Grok 4.7 took roughly 1,197 seconds per task in the xhigh evaluation, close to 20 minutes.
SpaceXAI charges the same headline rate of $2 per million input tokens and $6 per million output tokens. Yet Artificial Analysis calculates a $3.74 average cost per Intelligence Index task at xhigh.
This is where the launch story becomes uncomfortable.
The rate card barely changed. The amount of computation Grok is willing to consume did.
Cheap tokens can produce expensive answers
AI companies have trained customers to compare models using token prices because token prices are easy to publish.
Agents make that metric increasingly misleading.
A model that costs half as much per token can still cost more to finish a job if it burns through twice the output, spends minutes reasoning, calls tools repeatedly or wanders down dead ends.
Grok 4.7 looks vulnerable to exactly this problem.
Artificial Analysis recorded 240 million output tokens across its Intelligence Index evaluation at xhigh. Its comparison class median was dramatically lower. The same testing put output speed at roughly 39 tokens a second and described the model as unusually verbose.
None of this means Grok 4.7 is weak. SpaceXAI built it for coding, knowledge work and agentic tasks, and the model is scoring well on demanding evaluations.
The complaint is more commercially damaging than that.
A stronger model should earn the additional inference it consumes.
Developers do not buy reasoning tokens because they enjoy watching a model reason. They buy completed work.
If a small improvement in reliability requires a large increase in computation, the economics deteriorate quickly in production. Multiply the difference across thousands of coding sessions or autonomous tasks and a cheap API can become an expensive system.
That matters for Grok in particular because aggressive pricing has been part of its appeal.
Benchmarks hide how irritating a model can be
A benchmark does not get impatient.
It does not care that a coding task has been running for fifteen minutes. It does not wonder why the agent has reopened the same files three times. It does not have a monthly usage cap that disappears faster than expected.
Humans do.
That is why harsh early user reaction deserves attention even when individual complaints are anecdotal. Real-world use introduces costs that benchmark tables have trouble capturing: waiting, retrying, supervising and cleaning up after bad decisions.
Grok 4.7 may turn out to justify its heavier reasoning on difficult workloads. There will certainly be problems where spending more tokens is exactly what a user wants.
But the burden has shifted.
Longer reasoning used to sound automatically desirable. The industry now has enough experience with reasoning models to know better. Extra thought is useful when it improves the answer enough to pay for itself.
Grok 4.7 appears to be testing that limit.
The disk-wipe story is wrong. The underlying safety problem is not
One dramatic claim circulating around agent failures needs correcting.
The traceable report of an AI coding agent issuing rm -rf $HOME and damaging a developer environment concerns Cursor's Composer 2.5. The developer discussing the episode explicitly contrasted it with their experience using Grok. The incident should not be presented as a Grok 4.7 disk wipe.
That correction does not rescue the design pattern that made the incident possible.
Grok 4.7 is explicitly sold for coding and agentic work. SpaceXAI's documentation supports tool calling, code execution and long-running agent loops.
Any model operating in that environment will eventually make a stupid decision.
The serious engineering question is what happens next.
A coding assistant should be free to make mistakes inside a disposable workspace. Giving the same assistant enough access to erase a developer's home directory turns a model error into an infrastructure failure.
This problem will get worse as agents become more capable because users will trust them with longer tasks and broader permissions.
Guardrails therefore have to live outside the model too. Snapshots, scoped credentials, sandboxes and approval requirements for destructive actions are mundane compared with another ten points on a benchmark. They are also what separates an irritating hallucination from a catastrophic one.
Grok 4.7's early reception exposes both sides of the agent problem at once.
These systems are being allowed to spend more compute and take more actions because labs want them to solve harder tasks. That makes mistakes more expensive in dollars and potentially more expensive on the machine.
The market will eventually stop treating model intelligence as a score divorced from those costs.
For agentic AI, the better metric is becoming painfully obvious: how much money, time and supervision does it take to get a piece of work that a human is willing to keep?
Grok 4.7 has improved its benchmark scores.
It now has to prove that the work got better by more than the bill did.