- Grok 4.7 scores 46.3% on CursorBench 4.0’s long-running coding tasks and 19.6% on the Harvey Legal Agent Benchmark at xhigh reasoning.
- Artificial Analysis reports a 29% hallucination rate and 47% overall accuracy for Grok 4.7 on the AA-Omniscience Index.
- At a 19.6% legal task success rate, teams should plan for roughly five runs per successful outcome, at 81,000 output tokens each, per-task costs add up fast.
xAI’s Grok 4.7 was built to run for hours. On the benchmarks that actually test that, it fails more often than it succeeds.
Built for Long Tasks
xAI released Grok 4.7 on September 21, 2026, positioning it as its most capable model for extended coding and knowledge work. The model is explicitly trained to handle multi-hour problems, built on a larger base than Grok 4.6 and put through longer reinforcement learning runs on harder task mixes. According to xAI, the goal is a model that works longer on difficult problems, checks its own outputs more carefully and manages larger contexts without losing coherence.
The 500,000-token context window carries over from Grok 4.6. For agentic workflows that accumulate tool outputs, intermediate results and long instruction chains, that headroom matters. Grok 4.7 also adds configurable reasoning effort from low to xhigh, letting builders trade compute for deeper problem-solving on tasks that warrant it. Context compaction, previously available in Grok 4.5, carries forward here too, helping agents avoid context overruns in sustained loops. On paper, it is a coherent architecture for long-duration autonomy. The benchmark results are harder to reconcile with that goal.
Where the Numbers Break Down
Artificial Analysis puts Grok 4.7’s hallucination rate at 29% on the AA-Omniscience Index, down from Grok 4.6’s 34%, with accuracy essentially flat at 47% versus 48%. That’s a real improvement in calibration, but the residual rate is still a problem for the use case xAI is pitching. A model meant to run autonomously for hours, checking its own outputs, is generating fabricated or incorrect information nearly one time in three. In domains where factual precision matters, legal research, compliance, complex code review, that rate compounds rather than averages out. Early errors produce bad intermediate outputs; subsequent steps treat those outputs as valid. By the end of a long chain, the damage is well upstream of where the model finally surfaces a wrong answer.
The 500,000-token context window and configurable reasoning effort are real architectural advantages, but a 47% overall accuracy score suggests Grok 4.7 struggles with internal consistency over extended reasoning chains, not just raw task difficulty. For builders running complex multi-step agentic workflows, that means outputs need human checkpoints, which cuts into exactly the efficiency gains the agentic architecture is supposed to deliver.
Why Long Contexts Fail
Even with 500,000 tokens available, models can struggle to reliably retain information from the middle of long contexts, a pattern sometimes called the “lost in the middle” effect. Multi-hour agentic workflows generate enough interaction to push into those degradation zones, meaning an early constraint or finding may effectively be lost, corrupting later reasoning built on it.
Context compaction helps with token management, but compaction is summarisation. Some details survive compression; others do not, and what gets dropped is not always predictable. In a long agentic loop, a nuance that seemed minor at step three can become the pivot at step 30. A wrong tool call produces bad data, and subsequent steps process that bad data as if it were valid, the 29% hallucination rate suggests this kind of internal drift, not just memory limits, is what’s pulling down Grok 4.7’s scores on extended tasks.
The Real Cost Is Retries
xAI prices Grok 4.7 at $2.00 per million input tokens and $6.00 per million output tokens, with cached inputs at $0.50 per million. Competitive on the surface. The problem is token volume. At xhigh reasoning, Grok 4.7 uses roughly 81,000 output tokens per Intelligence Index task, according to Artificial Analysis’s independent testing. That is more than double Grok 4.6’s 36,000 at high reasoning, and nearly triple GPT-6 Astra’s 27,000 at max.
Do the maths on legal tasks. At a 19.6% success rate, you need an average of five runs to get one successful outcome, five runs at 81,000 output tokens each. For teams building legal AI workflows, that changes the unit economics of every task. The efficiency gains agentic systems are supposed to deliver start disappearing once retries become the default. The per-token price looks manageable; the per-successful-outcome price is what actually lands in the budget.
How to Deploy It Anyway
None of this means Grok 4.7 is not worth using. It means the deployment architecture needs to account for its failure rates rather than assume them away.
Structure long tasks as a sequence of smaller, independently verifiable sub-tasks. Orchestration frameworks are the right layer for this: each node in the graph completes a bounded piece of work, gets checked and only passes results forward if they clear validation. A failure at step four does not cascade to step 12. Pair that with monitoring that can catch drift or hallucination early, before the model burns 80,000 tokens on a path that was wrong at step two.
On reasoning effort: xhigh is not always the right call. Before committing to the highest-cost setting across an entire workflow, benchmark Grok 4.7 on your actual tasks at each reasoning level. The success rate and token consumption trade-off will look different depending on task type, and for some jobs, lower effort may deliver comparable results at a fraction of the cost. The gap between Grok 4.7’s 81,000-token xhigh runs and its lower-effort equivalents is large enough that it is worth measuring before it becomes a budget line item.
