Cursor prices Grok 4.7 High at $4.69 per benchmark task against $9.00 for Claude Opus 5 High, with scores of 43.9% and 44.7%. The comparison covers one coding benchmark; its commercial value is the measured completed-task cost, including the tokens and steps each agent used.
xAI released Grok 4.7 on September 21 as its strongest model for coding and knowledge work. On CursorBench 4.0 at the benchmark's High setting, Grok 4.7 scores 43.9% and costs $4.69 per task, while Claude Opus 5 scores 44.7% at $9.00. The measured cost difference is 47.9%, or $4.31 per completed benchmark task in Cursor's accounting.
The result measures one coding workload. CursorBench evaluates ambiguous, multi-file tasks drawn from real Cursor sessions, and Cursor warns that small score differences may not be statistically meaningful. The 0.8-point spread should therefore be read as a narrow coding-benchmark gap.
The same class of coding task now costs materially less in Cursor's measurement
Cursor computes task cost from the tokens each model actually used and the relevant published prices. That is more informative than comparing API tariffs in isolation because long-running coding agents accumulate cost through planning, tool calls, context growth and retries.
The benchmark also exposes those behavioral differences. Grok 4.7 High uses 56,382 tokens and 71 steps on average in the current table, versus 61,405 tokens and 86 steps for Opus 5 High. Here, the lower benchmark cost arrives with both fewer measured tokens and fewer steps, while the score remains close.
xAI says Grok 4.7 uses a larger base model than 4.6, received a longer reinforcement-learning run weighted toward tasks that take many hours, and was trained to understand the Grok Bot harness natively. Its launch data show material gains over Grok 4.6 on both CursorBench 4.0 and Terminal-Bench 4.0, consistent with the company's emphasis on longer-running software work.
Harness training makes model-only comparisons less complete
Training the model around Grok Bot is commercially relevant because coding performance increasingly comes from the model and the execution harness together. Tool schemas, memory, context-management policy, permissioning and recovery behavior can change both completion rate and cost.
A model tuned to its native harness may lose part of that advantage when moved into a third-party agent. The reverse can also happen: Cursor, Copilot or an internal engineering harness may fit a team's repositories and workflows better than the provider's own environment. CTOL's developer testing has found Grok 4.7 competitive on coding, multimodal generation and reasoning tasks, but those observations do not replace repository-specific evaluation.
Distribution is broad enough that buyers can test the question quickly. xAI's public API supports a 500,000-token context window, text and image input, and low, medium, high and xhigh reasoning settings. Below 200,000 prompt tokens, pricing is $2 per million input tokens, $0.50 for cached input and $6 per million output; the rates double above that threshold. GitHub began rolling Grok 4.7 into Copilot on September 21 across Pro, Pro+, Max, Business and Enterprise plans.
On CursorBench 4.0 High, Grok 4.7 is within 0.8 percentage points of Opus 5 at roughly half the measured task cost. That is the result the evidence supports. Engineering teams with enough agent volume should compare completed-task cost on their own repositories, including retries, review burden and latency. If that local spread resembles Cursor's, xAI's price advantage is large enough to change routing policy even when Opus retains a small benchmark edge.
Sources
xAI, "Introducing Grok 4.7," September 21, 2026. Cursor, CursorBench 4.0 results, accessed September 22, 2026. SpaceXAI developer release notes. GitHub Changelog, Grok 4.7 availability in Copilot, September 21, 2026. CTOL direct developer testing, September 21, 2026.