Grok 4.7 Launched: xAI's New Frontier Coding Model at Half the Price

Introduction: A New Frontier Contender Drops Today

xAI launched Grok 4.7 on September 21, 2026, calling it the company's most capable model yet for coding and knowledge work. The pitch is straightforward: a new, larger base model trained with a longer reinforcement learning run on harder tasks, served at the same price and speed as Grok 4.6 — which already undercut most of the frontier on cost.

For developers and teams evaluating AI coding tools, the launch matters for two reasons. First, Grok 4.7 posts some of its biggest gains on long-horizon work — the multi-hour coding sessions and terminal tasks where most models still fall apart. Second, it's available immediately in Cursor and through the Grok API, so you can try it in a real workflow today rather than waiting on a waitlist.

In this post, we break down the benchmark numbers, the architectural changes, the new safety stack, and what it all means if you're choosing an AI coding assistant or API model this month.

Grok 4.7 Benchmarks: Coding, Terminal Work, and Knowledge Tasks

The headline numbers come from benchmarks that stress tasks taking many hours to complete — the closest thing the industry has to measuring real engineering work.

Three things stand out:

• Terminal-Bench nearly doubled — from 20.3% to 38.0%, jumping past GPT-5.6 Sol. Terminal work is where agentic coding lives or dies, and this is the single largest generation-over-generation gain xAI reported.
• CursorBench leadership in its price class — at 46.3%, Grok 4.7 tops the table among models at or near its price point; only far more expensive models score higher.
• Domain knowledge work is a real strength — Grok 4.7 dominates on legal (19.6% on Harvey's benchmark, versus 2.5% for GPT-5.6 Sol) and electrical engineering, suggesting the model is not just a code specialist.

It's not a clean sweep: on DeepSWE and clinical reasoning (HealthBench Professional), GPT-5.6-class models still edge ahead, and xAI notes its DeepSWE number required high-effort inference. But for the price, the profile is remarkably balanced.

What Changed Under the Hood

Grok 4.7 is a new, larger base model rather than an incremental fine-tune. xAI highlights three training decisions:

• A longer RL run on harder tasks — the reinforcement learning mix was deliberately weighted toward problems that take many hours to solve, which is consistent with the outsized Terminal-Bench and CursorBench gains.
• Self-verification — the model was trained to check its own work more carefully before returning an answer, a technique that directly targets the "confidently wrong" failure mode in long coding sessions.
• Native harness understanding — Grok 4.7 was trained to natively understand the Grok Bot harness, improving conversational consistency and general knowledge work on top of raw coding skill.

Longer-context management was also improved, which matters for agentic tools that stuff entire repositories and terminal histories into a session.

Pricing and Availability

Grok 4.7 is served at the same price as Grok 4.6: $2 per million input tokens and $6 per million output tokens. For comparison, GPT-5.6 Sol (max) lists at $4/M input and $20/M output — making Grok 4.7 roughly half the input price and less than a third of the output price. A fast variant doubles output speed at twice the price.

Availability on day one:

• Cursor — selectable as a model for agent and inline coding workflows
• Grok Build — xAI's own coding harness, free to start at x.ai/build
• Grok API — via console.x.ai, plus third-party coding harnesses, model routers, and cloud platforms

Broad same-day routing support means most teams using an API gateway or model router can evaluate it in production traffic within hours of launch — a pattern that has become the norm in the 2026 model market, where choice of model is increasingly a routing decision rather than a loyalty decision.

The New Safeguard Stack

Alongside capability, xAI shipped what it describes as an entirely new safeguard stack — notable in a year when frontier models from multiple labs have been involved in high-profile security incidents.

xAI claims Grok 4.7 is the strongest model it has tested on refusals and jailbreak resistance. Specific claims include:

• 62.4% on LatchBio's biosafety benchmark, leading on both benign utility and safe refusal in dual-use domains
• Only 3.3% of risky dual-use prompts passing on HackerBench v0.3, xAI's benchmark for malicious cyber tasks, while rarely blocking legitimate security work
• Invite-only red-team access for select cybersecurity partners doing defense research

The dual-use positioning is the interesting part: the model is tuned to refuse offensive cyber and bio tasks without becoming useless to security professionals — a balance that many earlier "safe" models got wrong in the other direction. Independent evaluation will tell how well it holds up in practice, but the fact that safety calibration is now a launch-day headline — rather than an afterthought — reflects how much the market has shifted in 2026.

What Grok 4.7 Means for AI Tool Buyers

If you're picking an AI coding assistant or API model right now, Grok 4.7 changes the calculus in three ways:

1. Price-performance at the frontier keeps improving

Frontier-class coding ability is now available at $2/$6 per million tokens. If your bill is dominated by output tokens — typical for agentic coding — the savings versus premium-tier models compound quickly across thousands of agent runs.

2. Long-horizon agentic work is the new battleground

The benchmarks that improved most (Terminal-Bench, CursorBench, AA Briefcase) all measure multi-hour autonomous work. Every frontier lab is now optimizing explicitly for the agent use case, which means tool selection should weight these numbers over single-shot coding benchmarks like HumanEval-style evals.

3. Routing beats loyalty

With Grok 4.7 in Cursor, in every major router, and behind a cheap API, there's little lock-in cost to testing it against your current default on your own tasks. The pragmatic move in late 2026: benchmark your real workload across 2–3 frontier models monthly, and let a router send each request to whichever wins on cost and quality.

You can explore and compare AI coding assistants, agent frameworks, and API providers in the AI coding tools section of our directory.

Frequently Asked Questions