Grok 4.7

SpaceXAI's most capable proprietary model for coding, agentic tasks, and knowledge work, released September 21, 2026 after Musk walked the date back several times from an initial "four weeks after Grok 4.6." Unlike Grok 4.6, which extended the Grok 4.5 checkpoint, Grok 4.7 sits on a new, larger base model. It was trained with a longer reinforcement-learning run on a harder task mix weighted toward problems that take many hours, verifies its own work more carefully, handles longer context better, and was trained to natively understand the Grok Bot harness for conversational work. Text and image input, 500K context, knowledge cutoff May 2026, reasoning effort low through xhigh (default high), and encrypted reasoning returned on the Responses API. Served at the same price and speed as Grok 4.6: $2/M input and $6/M output tokens, with a 75% cache discount; a "Fast" variant at twice the speed and twice the price runs only in Cursor and Grok Build.

SpaceXAI's headline scores use xhigh effort: CursorBench 4.0 46.3% (Grok 4.6: 40.4%), DeepSWE v1.1 71.0% at high effort (65.2%), EEBench 64.0% (53.0%), Terminal-Bench 4.0 38.0% (20.3%), AA Briefcase v1.1 1,657 (1,546), Harvey Legal Agent Benchmark 19.6% (15.8%), and HealthBench Professional 56.7% (48.5%). The company frames it as frontier price-performance rather than frontier capability: Fable 5.1 Max still leads on CursorBench (51.8%) and Terminal-Bench (57.9%) at five times the price. A new safeguard stack is pitched as its strongest on jailbreak resistance while keeping refusals low for legitimate cybersecurity and biology work (62.4% on LatchBio's biosafety benchmark; 3.3% of risky dual-use prompts allowed on HackerBench v0.3), with invite-only red-team access for select security partners. Artificial Analysis scores the xhigh mode 46 on Intelligence Index v4.3 (#16 at filing; high effort also rounds to 46). Parameter count is undisclosed; press figures conflict (Musk floated ~2T for Grok 4.6, Decrypt reports 2.1T for 4.7).