AI

Grok 4.7 Pushes SpaceXAI Into the Top 4 of AI Labs

co-written by newsrooms.ai22. September 2026, 09:53

SpaceX’s AI division has just released Grok 4.7, and the new model moves the company into the leading group of AI labs. On the Artificial Analysis Intelligence Index, Grok 4.7 scores 46 points, two more than its predecessor Grok 4.6. That puts it fourth among the labs building frontier models. The evaluation used the highest reasoning setting, called xhigh.

For the company, which rebranded from xAI to SpaceXAI after its merger with SpaceX, this is the best placement it has achieved in the benchmark. SpaceXAI positions Grok 4.7 as its most capable model for coding and knowledge work, claiming it runs twice as fast at half the price of comparable models. The measurable gains cluster in two areas: agentic knowledge work and coding agents.

A Larger Base Model and a Longer Reinforcement Learning Run

According to the company, Grok 4.7 is built on a new and larger base model than Grok 4.6. It was trained with a longer reinforcement learning run on a harder mix of tasks, deliberately weighted toward problems that take many hours to complete. The result, SpaceXAI says, is a model that verifies its own work more carefully and handles long context more reliably.

Grok 4.7 was also trained to natively understand Grok Bot, the company’s own agent harness. That is meant to pay off in conversational tasks and general knowledge work.

A Clear Jump in Agentic Knowledge Work

The biggest gain shows up on AA-Briefcase, the private Artificial Analysis benchmark for long-horizon agentic knowledge work. Grok 4.7 scores 1,657 Elo there, 111 points above Grok 4.6 in its high configuration (1,546). Only Claude Opus 5 and Claude Fable 5.1 rank ahead, and Fable 5.1 leads by a narrow margin at 1,678 Elo.

The improvement is driven by analytical quality, which climbs from 1,690 to 1,994 Elo, while presentation quality stays roughly flat at 1,499 versus 1,519. The benchmark also checks whether submissions actually meet each task’s requirements, from completing the analysis to producing the deliverables that were asked for.

GDPval tells a similar story. It asks models to produce practical work products for professions such as lawyers, nurses and financial analysts. Grok 4.7 scores 1,695 Elo there, compared with 1,605 for Grok 4.6 (high). Fable 5.1 (max) leads with 1,735, while GPT-6 Astra (max) sits at 1,542.

The Coding Agent Overtakes GPT-5.6 Sol

On the Artificial Analysis Coding Agent Index, Grok 4.7 paired with the company’s own Grok Build agent scores 56 points, nine more than Grok 4.6. That is enough to pass GPT-5.6 Sol. Among models evaluated in their native harnesses, the combination now ranks fourth, behind Claude Fable 5.1, GPT-6 Astra and Claude Opus 5.

All three components improve: DeepSWE v1.1 rises from 65 to 73 percent, Terminal-Bench 4.0 from 18 to 33 percent, and SWE-Atlas-QnA from 58 to 63 percent. These figures sit apart from the Intelligence Index results, which standardize the harness across every model.

SpaceXAI itself highlights CursorBench 4.0, a test built around longer-running coding tasks. Grok 4.7 reaches 46.3 percent there, up from 40.4 percent for Grok 4.6, ahead of GPT-5.6 Sol (41.7 percent) and behind Fable 5.1 (51.8 percent). Measured by average cost per task, the company argues, that makes Grok 4.7 the strongest price-performance option in its class.

Strong in Engineering and Law, Weaker in Clinical Work

The domain-specific benchmarks published by SpaceXAI show a sharply uneven profile. On EEBench, which covers electrical engineering, Grok 4.7 scores 64.0 percent, ahead of Fable 5.1 (56.4 percent) and well ahead of GPT-5.6 Sol (39.4 percent). On the Harvey Legal Agent Benchmark it reaches 19.6 percent, where the competition lands at 6.7 and 2.5 percent respectively.

Clinical reasoning runs the other way. On HealthBench Professional, Grok 4.7 scores 56.7 percent and trails both GPT-5.6 Sol (60.5 percent) and Fable 5.1 (62.1 percent). On Terminal-Bench 4.0, SpaceXAI reports a jump from 20.3 to 38.0 percent, while Fable 5.1 leads that test clearly at 57.9 percent.

Elsewhere, the Numbers Barely Move

Across the remaining Intelligence Index tasks, Grok 4.7 broadly matches Grok 4.6. It gains on Terminal-Bench 4.0 (up 4.5 percentage points) and GDP.pdf (up 3.0 points), and gives some ground on AA-LCR (down 3.7 points) and AutomationBench-AA (down 1.1 points).

Reliability improves slightly. The hallucination rate on AA-Omniscience falls from 34 to 29 percent, while accuracy holds steady at 47 versus 48 percent. The overall index rises from 30 to 32.

The Cost of the Gains: Far More Tokens

Those gains come with a much heavier compute bill. Grok 4.7 burns roughly 81,000 output tokens per Intelligence Index task, against 36,000 for Grok 4.6 (high) and 27,000 for GPT-6 Astra (max). That works out to 125 and 196 percent more, respectively. Muse Spark 1.3 (max) sits in between at 60,000 tokens.

Artificial Analysis measures output speed at roughly 188 tokens per second on long prompts, with an average of about 7.1 minutes per benchmark task. Anyone deploying the model pays for that token appetite twice, once in cost and once in waiting time.

A New Safety Stack and Invite-Only Red Team Access

SpaceXAI says it built an entirely new safeguard stack for Grok 4.7 and calls it the company’s strongest model so far on refusals and jailbreak resistance. In dual-use domains such as cybersecurity and biology, the company claims the model leads on both usefulness for benign tasks and safe refusal of dangerous ones, topping LatchBio’s biosafety benchmark at 62.4 percent.

On HackerBench v0.3, its in-house test for risky and malicious cyber tasks, SpaceXAI reports that Grok 4.7 lets through only 3.3 percent of risky dual-use prompts while rarely blocking legitimate security work. The company has also begun giving selected cybersecurity partners invite-only access to the model’s red-team capabilities for defense research.

Pricing Holds, With a Fast Variant at Double the Rate

The commercial terms stay where they were. The context window remains at 500,000 tokens. Pricing starts at around €1.70 per million input tokens ($2) and around €5.10 per million output tokens ($6), with cache hits discounted to roughly 43 cents per million tokens ($0.50). For comparison, GPT-5.6 Sol runs at about €3.40 and €17 per million input and output tokens, while Fable 5.1 sits at roughly €8.50 and €42.50. SpaceXAI also offers a fast variant with twice the output speed at twice the price.

Reasoning effort is configurable from low to xhigh, with the published benchmark results produced at the highest setting. The model is available now in Cursor and Grok Build, as well as through the Grok API, third-party coding harnesses, model routers and cloud platforms.