Grok 4.5

What it is

Grok 4.5 is a model from xAI aimed at coding, agentic and knowledge work. It was announced on 16 July 2026, and xAI's release notes list it under July 2026; Wikipedia records that it reached the API on 8 July 2026, slightly ahead of the announcement.^[1][4][2]

"Agentic" here means the model is meant to work through multi-step tasks on its own — running commands, editing files, checking results — rather than answering a single question in one pass. That focus shows up in how the model was trained, served and priced.

How it was trained

xAI trained Grok 4.5 on tens of thousands of NVIDIA GB300 GPUs, and co-trained it with Cursor, a coding tool.^[1] The reinforcement learning stage covered hundreds of thousands of tasks, most of them multi-step software engineering.^[1] Reinforcement learning, in general, means training a model by rewarding the attempts that get closer to a working result, so the model learns from its own trial runs instead of only from fixed examples.

For the data, xAI invested in deduplication, quality scoring and domain-focused selection, so that the training mixture stayed high-coverage and high-signal.^[1]

The training stack was built for highly asynchronous training, so agentic rollouts — long runs in which the model attempts a task, often interacting with tools and files — could run for many hours while learning continued across tens of thousands of GPUs.^[1]

Benchmarks

xAI reported SWE-Bench Pro 64.7%, DeepSWE 1.0 62.0%, SWE Marathon 29.0% and Terminal Bench 2.1 83.3%.^[1] Benchmarks of this kind score a model against a fixed set of tasks, so the numbers are comparable across models only when the same harness and settings are used.

The post published competitor figures alongside its own: on DeepSWE 1.0, Fable scored 66.1%, GPT-5.5 64.31%, Grok 4.5 62.0% and Opus 4.8 55.75%.^[1] On SWE Marathon's pass@1 resolution rate — the share of tasks solved on the first attempt — Grok 4.5 scored 29.0%, above Opus 4.8 at 26.0% and Fable at 24.0%.^[1]

xAI also reported token efficiency: the model averaged about 15,954 output tokens per SWE-Bench Pro task, roughly 4.2 times fewer than Claude Opus 4.8 at maximum effort.^[1] Output tokens matter because they cost money and take time to generate, so reaching a solution in fewer tokens lowers both.

Serving, price and access

The model is served at about 80 tokens per second, and the API is priced at $2 per million input tokens and $6 per million output tokens, with low, medium or high reasoning effort — high by default.^[1][2] Its context window is 500,000 tokens, and prompts of 200,000 tokens or more are billed at double rates.^[3]

Grok 4.5 became the default model in the Grok Build coding tool, and it is available in Cursor on all plans, with free usage for a limited time.^[1]

It was not available in the EU at launch; xAI opened it to EU API users in July 2026, and it reached Grok.com, X and the mobile apps on 22 July 2026.^[4][2]

Where it sits

Grok 4.5 was a coding-focused release in the Grok line, trained together with a tool that people already use for software work, and placed directly inside that tool as its default model.^[1] Its successor, Grok 4.6, followed on 12 August 2026.^[4]