Did Elon catch up? (Grok 4.7 is here)

節目時間軸

1. 0:00

   Elon Musk's prediction that Grok 4.7 would match Opus 5.0 and the model's release at the same price and speed as Grok 4.6

   • Elon tweeted about a week before release that Grok 4.7 should be roughly on par with Opus 5.0, not 5.1, setting a specific benchmark expectation.
   • Grok 4.7 launched as a notable improvement over Grok 4.6 at the same price and speed, with xAI claiming it works longer on difficult tasks and checks its own work more carefully.
   • The blog post describes Grok 4.7 as xAI's most capable model for coding and knowledge work with their best calibrated safeguards to date.
2. 1:39

   Cursorbench 4.0 results show Grok 4.7 is competitive with Opus 5 but not beating it, at roughly half the cost

   • On Cursorbench 4.0, Grok 4.7 sits just behind Opus 5 on max thinking settings but costs about half as much to run the benchmark.
   • There is a major performance delta between low thinking effort at 33% and extra-high thinking effort at 46.3%, one of the biggest effort curves outside of GPT 5.6 Soul.
   • The host notes xAI now owns Cursor, so the benchmark should be viewed with that ownership relationship in mind.
3. 3:29

   Token efficiency and steps-per-task comparisons show Fable 5.1 leading on quality while Grok 4.7 wins on cost-effectiveness

   • On average output tokens per task, Grok 4.7 is very comparable to Opus 5, and at lowest thinking effort it uses few tokens but scores poorly.
   • Fable 5.1 is the quality winner but is multiple times more expensive than Grok 4.7, so its cost per completed task is much higher despite similar token usage.
   • GPT 5.6 Soul appears most efficient on steps taken per task, which matches the host's personal experience of it having the most direct shot to task completion.

關鍵概念

• Grok 4.7— The newly released xAI model that is the central subject of the episode.
• cost per task completed— The key metric the host argues matters most when evaluating models for real workloads.
• CursorBench— The coding benchmark used to compare Grok 4.7 against Opus 5 and other frontier models.

精選金句

• Grok 4.7 should be roughly on par with Opus 50, not 5.1.

  🔥

  — Elon's own prediction sets a high bar that the host then tests, revealing whether the claim holds up.

  跳至 0:00

• It is not beating it. It's just behind it on the max thinking setting for both models but it is much less expensive about half the cost to run that benchmark.

  🤯

  — Shows that near-frontier performance at half the cost is the real value proposition, not outright superiority.

  跳至 1:46

可執行的洞察