xAI Ships Grok 4.7 With New Safeguard Stack: Independent Benchmarks Confirm Gains, Flag Doubled Token Consumption

xAI has launched Grok 4.7, its most succesful mannequin thus far for coding and data work, positioning it as a frontier providing at a aggressive worth level. The mannequin is constructed on a brand new, bigger base than its predecessor, Grok 4.6, and was educated with an prolonged reinforcement studying run on a more durable mixture of duties weighted towards issues that require many hours to finish. According to the corporate, these enhancements improve the mannequin’s code technology, long-context dealing with, and self-verification capabilities, whereas native understanding of the Grok Bot harness makes it simpler at conversational duties and common data work.

Benchmark outcomes illustrate the place the features are most pronounced. On CursorBench 4.0, which stresses longer-running coding duties, Grok 4.7 posted a rating of 46.3%, up from 40.4% on Grok 4.6 and forward of GPT-5.6 Sol’s 41.7%, whereas remaining behind Fable 5.1’s 51.8%. The mannequin additionally recorded a high-effort rating of 71.0% on DeepSWE v1.1, trailing GPT-5.6 Sol’s 72.7% however edging Fable 5.1’s 70.0%. 

In skilled data work, Grok 4.7 improved on its predecessor throughout GDPval and AA Briefcase, the latter of which evaluates multi-hour workplace duties carried out by professionals resembling legal professionals, nurses, and monetary analysts, the place it scored 1,657, shut behind Fable 5.1’s 1,678. The most putting soar got here on Terminal-Bench 4.0, a measure of multi-hour terminal work, the place the mannequin practically doubled Grok 4.6’s efficiency, rising from 20.3 to 38.0%. On HackerBench v0.3, which covers dangerous cyber duties, the mannequin allowed solely 3.3% of harmful dual-use prompts by whereas hardly ever blocking legit safety work.

A distinguishing aspect of the discharge is Grok 4.7’s security stack, described by the corporate as totally new and the strongest it has examined on refusals and jailbreak resistance. The mannequin leads on LatchBio’s biosafety benchmark at 62.4%, balancing utility for benign duties with secure refusal on harmful ones in dual-use domains resembling cybersecurity and organic analysis. xAI has additionally begun granting choose cybersecurity companions invite-only entry to the mannequin’s red-team capabilities for protection analysis.

Pricing is a central a part of the launch technique. Grok 4.7 is served on the similar charges as Grok 4.6, two {dollars} per million enter tokens and 6 {dollars} per million output tokens, undercutting GPT-5.6 Sol (4 and twenty {dollars} respectively) and Fable 5.1 (ten and fifty {dollars}). A quick variant providing twice the output pace is on the market at twice the value. On CursorBench’s price-performance frontier, this locations Grok 4.7 among the many most cost-efficient choices for prolonged coding workloads.

The mannequin is on the market instantly in Cursor and Grok Build, with entry additionally offered by the Grok API, third-party coding harnesses, and mannequin routers and cloud platforms. The launch intensifies competitors within the frontier mannequin phase, the place distributors are more and more differentiating on long-duration agentic duties, security calibration, and value per accomplished activity reasonably than uncooked benchmark scores alone. For builders and enterprises evaluating coding-oriented fashions, Grok 4.7 presents a mixture of pricing stability, improved multi-hour activity efficiency, and hardened safeguards that would make it a viable different to dearer frontier choices.

Independent Analysis Confirms Frontier Status, With a Token-Consumption Caveat

Independent benchmarking agency Artificial Analysis has corroborated xAI’s claims, scoring Grok 4.7 at 46 on its Intelligence Index, a two-point achieve over Grok 4.6, a end result that locations SpaceXAI among the many prime 4 AI labs. The agency’s analysis, performed at xhigh reasoning effort, discovered the mannequin’s clearest advance in long-horizon agentic data work: on its non-public AA-Briefcase benchmark, Grok 4.7 gained 111 Elo over its predecessor to achieve 1,657, rating alongside Claude Opus 5 and Claude Fable 5.1 on the frontier. The enchancment was pushed primarily by analytical high quality, which rose sharply to 1,994 Elo from 1,690, whereas presentation high quality held roughly regular. On GDPval-AA, which measures sensible work merchandise resembling paperwork, spreadsheets, and slides, the mannequin scored 1,695 Elo, up 90 factors.

Performance features got here at a price in compute consumption. Artificial Analysis measured roughly 81,000 output tokens per Intelligence Index activity for Grok 4.7 at xhigh, greater than double the 36,000 utilized by Grok 4.6 and practically triple the 27,000 consumed by GPT-6 Astra at max effort. The agency additionally famous incremental modifications elsewhere: modest enhancements on Terminal-Bench 4.0 and GDP.pdf, alongside small regressions on AA-LCR and AutomationBench-AA. Reliability metrics improved modestly, with the AA-Omniscience hallucination price falling to 29% from 34%, whereas accuracy held near-unchanged at 47%. Technical specs stay according to the earlier technology, together with an unchanged 500,000-token context window and cache-hit pricing discounted to $0.50 per million tokens.

The submit xAI Ships Grok 4.7 With New Safeguard Stack: Independent Benchmarks Confirm Gains, Flag Doubled Token Consumption appeared first on Metaverse Post.