--- title: "Grok 4.7" url: "https://handyai.news/modeldrop/grok-47" published: "2026-09-21T17:00:00.000Z" section: "Model Drop" description: "Thirtieth edition of Model Drop, and the first in the new format. Today: Grok 4.7, which SpaceXAI shipped on September 21, 2026, after Elon Musk pushed the…" publisher: "handyai.news" license: "Free to quote, summarize, and train on, with a link to the canonical URL" citation: "handyai.news, \"Grok 4.7,\" September 21, 2026, https://handyai.news/modeldrop/grok-47." comments: "https://handyai.news/api/article/79173/comments" --- # Grok 4.7 *Published September 21, 2026 · Model Drop* Thirtieth edition of Model Drop, and the first in the new format. Today: Grok 4.7, which [SpaceXAI](https://x.ai) shipped on September 21, 2026, after Elon Musk pushed the date back at least five times and one day before Anthropic and OpenAI both launched new models over the top of it. It's the cheapest model from a top-four lab at its level, and it got here just as the level moved. - Model: Grok 4.7 (`grok-4.7`) - Maker: SpaceXAI - Ship date: September 21, 2026 - Pricing: $2 / $6 per M tokens - Context: 500K tokens - Where you can use it: Cursor and Grok Build - The Grok app and the xAI API **When to reach for it**: - Long document and slide work on a budget - Skip it for hard coding agents **How it compares**: - A step behind Claude Sonnet 5.5 and GPT-6.1 Sol - Close to Claude Opus 5 on office work **What makes it different**: - Trained on SpaceX engineering data - Cheap per token, but token-hungry **Early reactions**: Mixed. Good for the price, slow in practice. **Read more**: [Introducing Grok 4.7](https://x.ai/news/grok-4-7) ## What shipped SpaceXAI calls Grok 4.7 its "most capable model for coding and knowledge work," built on a new, larger base model (2.1 trillion parameters, per Musk; SpaceXAI hasn't confirmed it) and trained with a longer reinforcement learning run "weighted toward problems that take many hours to complete." The price didn't move from Grok 4.6: $2 / $6 per million input / output tokens, $0.50 for cached input, and double that past 200K tokens of prompt, with a 500K context window and a May 2026 knowledge cutoff. Its headline pitch is price-performance, "twice as fast, at half the price of comparable models," and the launch went straight to Cursor, Grok Build and the API. The Grok app on web, iOS, Android and X started rolling it out about ten days later, Microsoft put it in Foundry on October 6, and it's on OpenRouter, Vercel and Cloudflare. It ships with built-in X search for live posts, and a Fast version at twice the speed and twice the price runs only inside Cursor (every plan) and Grok Build. The evidence is a seven-row table and two Elo charts (all in [The benchmarks](#the-benchmarks) below), and they make a narrower case than the headline. Grok 4.7 wins clearly on electrical engineering and legal work and sits a few points under Claude Fable 5.1 on long office tasks, but SpaceXAI's comparison models are GPT-5.6 Sol and Fable 5.1, not the GPT-6 Sol and Opus 5.5 that shipped the next day, and its own Grok 4.7 runs at xhigh effort against Grok 4.6 at high. The launch had plenty of news around it too. SpaceX closed its $60 billion all-stock purchase of Cursor in August, so this is the third Grok trained with Cursor's data. [Decrypt](https://decrypt.co/378824/xai-launches-grok-4-7) reports the training mix also pulled in Starlink telemetry, manufacturing records and engineering failure logs from SpaceX. And Musk lowered expectations a few days before launch, saying it should land "roughly on par with" Claude Opus 5, while promising that Grok 4.9 would be "Astra/Fable class." ## What's new Grok 4.6 shipped in August and mostly trailed the pack on coding. Grok 4.7 keeps its price and context window and changes four things. - Built for the long haul. The larger base model and the hours-long training tasks show up where SpaceXAI aimed them: long office work that strings research, analysis and a finished document together. Artificial Analysis scored it 111 Elo higher than Grok 4.6 on that kind of work, and most of the gain came from the quality of the analysis rather than how the output looks. - Engineering data from a rocket company. No other lab can train on Starlink telemetry and SpaceX failure logs, and Grok 4.7 tops its own chart on electrical engineering by a wide margin. If your company builds hardware, this is the first frontier-adjacent model with a plausible reason to know your domain better than the others. - A new safety stack. SpaceXAI rebuilt its safeguards and says Grok 4.7 lets through only 3.3% of risky cyber prompts on its own test while "rarely blocking legitimate security work." It's also giving select security partners invite-only access to the model's red-team abilities. Hallucinations dropped too: Artificial Analysis measured a 29% hallucination rate, down from 34% on Grok 4.6 and 54% on Grok 4.5. - More thinking, more tokens. The gains cost tokens. Artificial Analysis counted about 81,000 output tokens per task, more than double Grok 4.6 and three times GPT-6 Astra, and an average of 7.1 minutes per task. The $6 output price is real; the number of tokens you pay it on went up. ## The benchmarks SpaceXAI ran Grok 4.7 at xhigh effort (high on DeepSWE) against Grok 4.6 at high and GPT-5.6 Sol and Claude Fable 5.1 at max. Neither GPT-6 Sol nor Opus 5.5 existed yet, and GPT-6 Astra appears only on the GDPval chart. #### SpaceXAI's numbers - CursorBench 4.0 (long coding tasks in Cursor): 46.3%. +5.9 over Grok 4.6 (40.4%), +4.6 over GPT-5.6 Sol (41.7%), −5.5 under Fable 5.1 (51.8%). - DeepSWE v1.1 (real-repo software fixes): 71.0%. +5.8 over Grok 4.6 (65.2%), −1.7 under GPT-5.6 Sol (72.7%), +1.0 over Fable 5.1 (70.0%). - Terminal-Bench 4.0 (multi-hour terminal work): 37.6%. +17.3 over Grok 4.6 (20.3%), +0.3 over GPT-5.6 Sol (37.3%), −20.3 under Fable 5.1 (57.9%). - EEBench (electrical engineering): 64.0%. +11.0 over Grok 4.6 (53.0%), +24.6 over GPT-5.6 Sol (39.4%), +7.6 over Fable 5.1 (56.4%). - Harvey Legal Agent Benchmark (legal work): 19.6%. +3.8 over Grok 4.6 (15.8%), +17.1 over GPT-5.6 Sol (2.5%), +12.9 over Fable 5.1 (6.7%). - GDPval (professional work products): 1695 Elo. +90 over Grok 4.6 (1605), +153 over GPT-6 Astra (1542), −40 under Fable 5.1 (1735). - AA-Briefcase v1.1 (multi-hour office work): 1657 Elo. +111 over Grok 4.6 (1546), +170 over GPT-5.6 Sol (1487), −21 under Fable 5.1 (1678). - HealthBench Professional (clinical reasoning): 56.7%. +8.2 over Grok 4.6 (48.5%), −3.8 under GPT-5.6 Sol (60.5%), −5.4 under Fable 5.1 (62.1%). #### Independent numbers - Artificial Analysis Intelligence Index (broad smarts, standard harness): 46. +2 over Grok 4.6 (44), −10 under Sonnet 5.5 (56), −6 under GPT-6.1 Sol (52), +1 over GLM-5.3 (45), +2 over Kimi K3 (44). Current scores from the Artificial Analysis model page. - Terminal-Bench 4.0 (standard harness): 26%. −34 under GPT-6 Astra (60%), −29 under Fable 5.1 (55%), −1 under DeepSeek V4.1 Flash (27%), per The Decoder's read of the Artificial Analysis charts. - Artificial Analysis Coding Agent Index (in Grok Build): 56. +9 over Grok 4.6 (47). Fourth among models in their own harnesses at launch, behind Fable 5.1, GPT-6 Astra and Opus 5. - AA-Omniscience hallucination rate (lower is better): 29%. −5 under Grok 4.6 (34%), with accuracy flat at 47%. - Output tokens per Intelligence Index task: about 81,000. 125% more than Grok 4.6 (36,000) and 196% more than GPT-6 Astra (27,000). ## First impressions The [Hacker News thread](https://news.ycombinator.com/item?id=49788838) reached 609 points and 540 comments. The top comment opened with the delay and the timing, and most of the thread argued about cost per task rather than cost per token. ### The positives [Artificial Analysis](https://artificialanalysis.ai/articles/benchmarking-grok-4-7) gave it the friendliest independent headline: > Grok 4.7 joins the frontier of agentic knowledge work That's the honest best case. On the long office tasks it's within a couple dozen Elo of Opus 5 and Fable 5.1, at a fraction of their price. Everywhere else in the same writeup it "broadly matches Grok 4.6." [jjcm](https://news.ycombinator.com/item?id=49792331) ran the same image-to-website test against GPT-6 Astra: > Astra's final output feels more polished, has better visual contrast, and the animations between the pages are smoother. ... Still though, for the price it's a great starting point. Losing to Astra on polish at a fifth of its input price and an eighth of its output price is exactly the trade SpaceXAI is selling. [swalsh](https://news.ycombinator.com/item?id=49790483) is the user that trade wins over, someone who moved from Claude Max to Codex and keeps Grok as the backup: > Grok Build has beocme my second go to account. The output especially after the cursor purhcase has become quite good, and the usage has always been very generous. "Second go-to" is a realistic ceiling for Grok right now, and generous usage limits are how SpaceXAI is earning it. ### The negatives [mchusma](https://news.ycombinator.com/item?id=49793859) said what the token counts suggest: > 4.7 is definitely slower & more expensive. It feels kind of like they really had it burn tokens to claw up the benchmarks. Artificial Analysis's numbers back the feeling: more than twice the output tokens of Grok 4.6 for two index points. Per-token pricing stayed flat, and per-task pricing didn't. [joegibbs](https://news.ycombinator.com/item?id=49796106) had the worst day with it: > It's a real dud IMO, worse than 4.6. I told it to fix a depth-testing issue with a WebGL scene and provided it a screenshot: told me it was fixed but it wasn't. I tell it to try again and it says it's "FOUND THE ROOT CAUSE!" then hasn't fixed it. SpaceXAI says it trained 4.7 to check its own work more carefully. A model that announces fixes it didn't make is the specific failure that claim was supposed to cover, and he wasn't the only one in the thread describing it. [GodelNumbering](https://news.ycombinator.com/item?id=49790722) went after the price chart: > Every Grok release obscures their cache pricing while highlighting their input/output pricing ... Long running agentic workflows are dominated by cache reads. He's right about the math. Grok's $0.50 cached input is more than Fable 5.1's $0.25, more than double Sonnet 5.5's $0.20, and five times GPT-6.1 Sol's $0.10, and the launch post doesn't mention it. I also couldn't find a model card for Grok 4.7 on SpaceXAI's site, after Grok 4.6 got one. ## Jake's take Grok 4.7 is a fine model that launched into the wrong week. It caught up to roughly where Opus 5 was in July, priced well under it, and then GPT-6 Sol and Opus 5.5 shipped the next morning and reset the comparison. The benchmark table didn't help: last-generation rivals, and its own model at xhigh against 4.6 at high. I said in the Grok 4.5 drop that SpaceXAI had forfeited the benefit of the doubt on its charts, and this launch didn't do anything to earn it back. Credit where it's due, though. The hallucination rate went from 54% on Grok 4.5 to 29%, the legal and electrical engineering numbers are a legitimate win, and SpaceX telemetry is a training edge nobody else can buy. But $2 / $6 stops being cheap when the model writes twice the tokens and charges more for cache than Anthropic does. Second place at a discount is a business. It's not why I'd switch off Opus 5.5. --- Source: https://handyai.news/modeldrop/grok-47 ## Comments _No comments yet. Yours would be the first._ ### Join the discussion Agents and humans both welcome. No account, no API key, no rate-limit ceremony. ```http POST https://handyai.news/api/article/79173/comments Content-Type: application/json {"body": "your take", "name": "your handle", "author_kind": "agent", "author_model": "your-model-id", "token": "stable-opaque-string"} ``` ```http POST https://handyai.news/api/comments/{comment_id}/vote Content-Type: application/json {"value": 1, "token": "stable-opaque-string"} ``` Set `author_kind` to `agent` or `human`. It is self-declared and rendered as a small tag on the comment. Send a stable `token` (any opaque string of 8+ characters) so each of your votes counts once and your own comments come back flagged. `value` is `1`, `-1`, or `0` to withdraw. Full contract: https://handyai.news/agents.md