Grok 4.7 is out: what the vendor table supports, and what it does not

xAI released Grok 4.7 on 21 September 2026 and called it the company’s most capable model for coding and knowledge work. It is a generator. It is live in Cursor, Grok Build, and the Grok API, with GitHub Copilot rolling the same id out gradually at provider list price. The list price did not change versus Grok 4.6: $2 per million input tokens and $6 per million output tokens.

That is enough to file a story. It is not enough to crown a model. This desk’s writeup is split on purpose. The durable page is the Grok 4.7 profile. The task split against Claude is Grok 4.7 vs Claude. The bill, including the fast tier and the Sonnet mix-up, is the pricing review. What follows is the reading of xAI’s own table, dated with the launch.

The slogan and the table disagree in a useful way

The headline says twice as fast, at half the price of comparable models. The body says the same price and speed as Grok 4.6. Keep the body. “Half the price” only becomes a checkable sentence when you use the columns xAI printed: GPT-5.6 Sol Max at $4 / $20 and Fable 5.1 Max at $10 / $50. It is not half the price of Grok 4.6, and it is not half the price of Claude Sonnet 5, which this site still dates at $2 / $10.

There is a faster Grok 4.7 at twice the output speed and twice the price. That is a second product. Turning it into $4 / $12 is our reading of the sentence, not a rate-card row. Check the live price before you automate it.

Where 4.7 is ahead of 4.6, which is the boring and real result

On every task xAI listed, 4.7 beats 4.6: CursorBench 46.3% versus 40.4%, DeepSWE 71.0% versus 65.2%, EEBench 64.0% versus 53.0%, AA Briefcase 1,657 versus 1,546, Terminal-Bench 38.0% versus 20.3%, Harvey 19.6% versus 15.8%, HealthBench 56.7% versus 48.5%. If you already pay for Grok 4.6, the upgrade case does not need a rival. Same price, higher vendor scores, new base model, longer training on multi-hour tasks, a claim of better self-checks. That case is still the vendor’s. It is the cleanest one on the page.

Where the Claude column wins anyway

Fable 5.1 Max leads CursorBench (51.8% to 46.3%), Terminal-Bench (57.9% to 38.0%), AA Briefcase (1,678 to 1,657), and HealthBench Professional (62.1% to 56.7%). Terminal-Bench is the one that should change a plan. A twenty-point gap on terminal work is not something “price-performance” erases. If your agents live in a terminal, the sheet says stay with that Claude SKU or measure your own traces before you move.

Where Grok’s column is the one to quote

EEBench is 64.0% to Fable’s 56.4% and Sol’s 39.4%. Harvey’s legal-agent benchmark is 19.6% to 6.7% and 2.5%. DeepSWE is a near-tie with Fable once you keep the high-effort asterisk, and Sol is slightly higher. Quote those rows if they match the work. Do not average them with Terminal-Bench into a single winner.

xAI also shows GDPval and names GPT-6 Astra on a chart. We are not lifting a lone GDPval score out of secondary recaps. The main table is specific. The chart can wait until it is equally specific.

What we left out on purpose

• Parameter counts and context-window sizes. They are not in the launch note this page cites.
• A setup tutorial. This site profiles models. It does not onboard API keys.
• A safety grade. xAI cites a new safeguard stack, 62.4% on LatchBio’s biosafety benchmark, and 3.3% of risky prompts allowed on HackerBench v0.3. Those are vendor results. They are on the profile as claims, not as our tests.

What to open next

If you needed the model id and the limits, stop at the profile. If a team is about to replace a Claude Max SKU because of a headline, send them to the comparison first. If finance has “half the price” in a slide, send them to the pricing review. Jev remains the page for a typed decision, not another chatbot.