Cursor's task-cost table shrinks Grok 4.7's 7.5x list-price gap to about 1.9x

At list price, a million input tokens plus a million output tokens costs $8 on standard Grok 4.7 and $60 on standard GPT-6 Astra or Claude Fable 5.1 ^[17]. That is 7.5 times the price for the same mix ^[24]. No long agent run looks like that mix, and the one place in the record where somebody metered completed work shows the difference: in Cursor's September tests, Grok 4.7 averaged $4.69 per task while Fable 5.1 and Opus 5 cost about $9 ^[23][25]. The 7.5x becomes about 1.9x ^[28]. Cursor's published figures do not break out tokens consumed per task.

Same test, high effort on both sides: Grok 4.7 scored 43.9% on CursorBench 4.0, Fable 5.1 49.2%, Opus 5 44.7% ^[23][25]. Buying Fable's extra 5.3 points cost $4.31 more per task, about $0.81 per point ^[29]. Cursor cautions that small score differences may fall within normal evaluation variance ^[7]. Part of the increment a team is pricing may therefore be noise in the harness.

For $4.69 to transfer, a workload would have to resemble CursorBench 4.0 tasks, run at high effort, inside Cursor's harness, with Cursor's tool set and token budget. Change any of those and the scores move. SpaceXAI's own Terminal-Bench 4.0 setup puts Grok 4.7 at 38%, up from Grok 4.6's 20.3% ^[4]. Artificial Analysis puts it at 26% on the same named benchmark, against roughly 60% for GPT-6 Astra and 55% for Fable 5.1, with the cheaper DeepSeek V4.1 Flash at 27% ^[5][10]. theneuron.ai attributes divergence like that to differences in harnesses, reasoning settings, tool configurations, token budgets and evaluation environments ^[8].

The rate card has more variables than the two headline prices. SpaceXAI's developer documentation lists a 500,000-token context window, a May 2026 knowledge cutoff, and four reasoning-effort settings: low, medium, high and xhigh ^[9]. Effort is what moves token consumption, and SpaceXAI's own launch table compared 46.3% at xhigh against Grok 4.6's 40.4% at high ^[6]. The settings alone can explain that spread. The US regional endpoint adds a 10% token-price premium ^[11]. The fast version costs twice the standard rates for twice the output speed and runs only in Cursor and Grok Build ^[13].

Caching narrows the gap from the other side. Anthropic cut Fable 5.1's cache-read pricing to $0.25 per million tokens and says caching can reduce typical workload costs by roughly 25% and highly agentic workloads by as much as 45% ^[20]. Take the best case: 45% off $60 leaves $33, still just over four times Grok's $8 ^[26]. OpenAI also offers cached-input pricing ^[21].

On the independent Artificial Analysis Intelligence Index v4.3.2, which combines ten benchmarks, Grok 4.7 scores 46 while GPT-6 and Fable 5.1 lead at 53 each ^[14]. SpaceXAI reported 71% on DeepSWE v1.1 at high effort, close to GPT-5.6 Sol's 72.7% and above the 70% listed for Fable 5.1 ^[19]. SpaceXAI also says Grok 4.7 uses an entirely new safeguard stack, and reported 62.4% on LatchBio's biosafety benchmark and 3.3% of risky dual-use prompts getting through on HackerBench v0.3, its own cyber-safety test ^[18].