▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=jLgpzgpsWPc
• Original title :: Elon promised this one would be good...

Theo reviews SpaceXi's Grok 4.7 and uses it as a case study for a larger argument: published benchmarks no longer capture what makes a model good or bad, and the marketing built on top of them is increasingly misleading. He likes the model — it is one of his favourites this year — but concludes it is a stepping stone that costs current-generation money for last-generation results.

The promises versus the release

Elon had pitched 4.7 for over a month as the model that would exceed everything else, trained on a uniquely good SpaceX corpus, better than 4.6 in every way, slightly slower to serve but with better token efficiency. The shipped model contradicts most of that: on Artificial Analysis it consumes roughly 81k tokens per task against 4.6 XHigh's 38k — more than Fable 5.1, and nearly triple Soul's 29k. Cost per task went from $1.86 (4.6) to $3.74, against Astra's $3.26.

Michael Truell (Cursor CEO, now inside the same company) pushed back publicly: internally they see only 5% more tokens on median requests and 20–30% on P99. Theo's counter is that "request" is the wrong unit. One prompt can fan out into many requests when tool calls are involved, and Artificial Analysis is weighting agentic workloads more heavily over time. Per-request efficiency can hold steady while per-prompt cost doubles. He puts it bluntly: either the correction is itself misleading, or the original numbers were.

Elon's own explanation for the week-long delay supports this reading — they had penalised response length too much in RL, so the model gave up early on hard tasks and under-verified its work. The extra week of RL that fixed that is very likely where the token efficiency went.

The benchmark picture

The official post cherry-picks an unusual set: Cursor Bench, DeepSWE V1.1, electrical engineering, AA Briefcase, Terminal Bench, legal work, HealthBench. Astra is absent throughout because OpenAI banned SpaceX from using its models, and the bench isn't shared so third parties can't fill the gap. Where comparisons do exist they are unflattering — Fable 5.1 scores roughly 20 points higher on Terminal Bench, and on external benches like GDPval and Ebench Astra is well ahead.

Cognition's Frontier Code test has 4.7 scoring below 4.6 in aggregate; Cognition attributes it to 4.7 "overcoping" on easier tasks. On Artificial Analysis's overall index 4.7 lands barely above GLM 5.3 Max and below Muse Spark 1.3. On the coding-agent index Grokbot scores 56 against 62 for Codex and Claude Code. Theo argues those six points are not trivial: a lower rate of doing something stupid translates super-linearly into how long you can leave an agent running unattended.

Front-end output is the weakest area by a wide margin — he calls the generated designs some of the worst he has seen from any model, including 2025 vintages. Anthropic remains the design leader.

Pricing and limits

The headline price-per-million matches 4.6, but the price doubles above 200k tokens and rises further in fast mode, so real usage is considerably more expensive. Ben Davis, Theo's podcast co-host and a committed Grok user, burned 8% of his weekly limit on $40 of usage — the $300/month plan no longer behaves like the effectively unlimited plan it used to be, while competitors give $8–12k of monthly value on $200 plans. The old economic argument for Grok (cheap, fast, generous subsidy) is gone, and what replaces it is not a premium token.

What the model is actually good at

The strengths are the ones benchmarks capture worst. The floor is higher: it stays on task, completes real work, and needs less intervention than its capability tier suggests. It is unusually inquisitive, digging deeper than peers and surfacing things Fable and Astra missed. It was also trained natively on the Grokbot harness, which Theo believes is a genuine and underrated advantage.

To test this he ran an ad-hoc bench: several models reviewed the T3 Chat codebase and proposed improvements, judged by Fable 5.1 and Astra. Astra came first on novelty and verification, Grok 4.7 a close second, Fable 5.1 third — Fable made only five suggestions against eight, but all were valid; it simply went down fewer dead ends. Cost told the other half of the story: best case Grok matched Astra and Fable, worst case it cost twice as much. His "fish slop" 3D generation test went badly — an hour and a half to build, backwards animations, broken mouse tracking, against a decent result from Astra on the identical prompt.

Verdict

A model worth playing with, especially for people already inside Cursor, but hard to recommend otherwise: more expensive, slower, higher error rate, worse code, and without the subscription subsidy that made earlier Grok models compelling. Grok 4.5's virtues — cheap, fast, efficient — have not survived into 4.6 or 4.7. Theo reads 4.7 as evidence that SpaceXi is struggling with the generational jump that Fable and Astra already made, and hopes the next release is where the promises finally land.