xAI released the latest version of its AI model, Grok 4.7, on September 21, 2026. The new model is built to handle coding, specialized knowledge work, and complex reasoning tasks that can take hours to complete, while token pricing stays the same as the previous version. The model has already rolled out across development tools and xAI's own apps, and it shows solid gains over rivals in benchmark comparisons.

A Larger Foundation Model with Expanded Training Data

Grok 4.7's foundation model now runs on roughly 2.1 trillion parameters, up about 40 percent from the 1.5 trillion in the prior version. Training reportedly involved an extended reinforcement learning process built around difficult, multi-hour problems, with a focus on improving self-verification and long-context handling. xAI also folded in engineering records from its space operations, which are said to help the model reason through problems involving physical systems.

Pricing stays flat from the previous release: $2 per million input tokens (about 320 yen) and $6 per million output tokens (about 950 yen). A faster tier is also available, offering double the output speed at double the price.

Steady Gains Across Coding and Specialized Benchmarks

On the performance side, Grok 4.7 shows clear improvement on several coding benchmarks. One software-engineering evaluation climbed from 65.2 percent to 71.0 percent, while another rose from the low 40s to 46.3 percent, edging past some GPT-family models. A terminal-operation benchmark jumped from the low 20s to 38.0 percent. The model also improved on an electrical-engineering evaluation, reaching 64.0 percent, and on a legal AI-agent benchmark, reaching 19.6 percent.

Enterprise agent evaluations rose as well, with gains of more than 100 points on one leaderboard. Still, the model trails the top competing systems on a couple of metrics, underscoring how close the race among frontier models remains. xAI says it plans further performance gains in upcoming successor models.

Rebuilt Safety Systems, Now in Dev Tools and In-Car Assistants

This release also comes with an overhauled safety stack. The model scored in the 60-percent range on a biosafety evaluation, and on an internal cybersecurity test, it reportedly blocked more than 90 percent of dangerous prompts without over-refusing legitimate security research requests. Select security partners will get invite-only access to red-teaming tools for defensive research.

Availability is broad: Grok 4.7 is live without a waitlist through developer tools and the API, and it's also built into xAI's app, its social platform integration, and in-car voice assistants.

Summary

xAI's newly released Grok 4.7 expands the foundation model and its training data to strengthen coding and specialized knowledge work. Pricing holds steady with the previous version, while benchmark scores improved across the board, pointing to wider use from developer tools to in-car assistants. With further successor models already planned, the pace of competition among frontier AI models remains worth watching.

※1 USD = 159 JPY (as of September 25, 2026)