What Is Grok 4.5?

xAI, now operating under the SpaceXAI brand, released Grok 4.5 on July 8, 2026. The company calls it “SpaceXAI’s smartest model built for coding, agentic tasks, and knowledge work,” and Elon Musk pitched it as “an Opus-class model, but faster, more token-efficient and lower cost.”

The headline facts: $2 per million input tokens, $6 per million output, a 500k context window, 80 tokens per second, and a training run that folded in real developer sessions from Cursor, the AI code editor SpaceX agreed to acquire in June. It’s live today in Grok Build, in Cursor on all plans, and through the xAI API.

This guide covers what Grok 4.5 is, what the published numbers say (including the ones xAI’s marketing doesn’t emphasize), and how to start calling it. If you want to test the endpoint before wiring it into an app, Apidog lets you send requests to grok-4.5, inspect streaming responses, and mock the API for your team without writing a line of client code.

Grok 4.5 at a glance

Two details in this table deserve a second look.

First, the context window. Grok 4.5 ships with 500k tokens of context, while the older, cheaper grok-4.3 offers 1M. If your workload is long-document RAG or whole-repo analysis, the previous model holds more. xAI is betting you’ll trade context for capability.

Second, the EU note. The release page states Grok 4.5 “is not yet available in the EU in any SpaceXAI products or the API console,” with availability expected mid-July. If you’re building for European users, plan around that gap.

What’s new in Grok 4.5

Trained alongside Cursor

The defining fact of this release: Grok 4.5 “was trained alongside Cursor.” According to Cursor’s own announcement, the editor contributed trillions of tokens capturing “user interactions with codebases and software tools.” That means the model learned from debugging traces, multi-file edits, and the corrections developers made when an agent got something wrong, not only from static code.

This is a different data strategy than most labs use, and it raises fair questions about consent and privacy. We break those down in our companion piece on what Cursor session training means for developers.

Reinforcement learning at scale

xAI says the model trained across tens of thousands of NVIDIA GB300 GPUs, with reinforcement learning covering “hundreds of thousands of tasks, centered on multi-step software engineering.” The stack runs agentic rollouts asynchronously for hours while training continues. The goal, in xAI’s words, is “per-token intelligence”: fewer, smarter tokens instead of long wandering reasoning chains.

Office work, not only code

Grok 4.5 is now the default model in Grok Build, and xAI spends a surprising amount of the announcement on spreadsheets. The model builds multi-sheet Excel models with web research, constructs PowerPoint diagrams from native shapes, and drafts prose in Word through xAI’s Office plugins. Cursor frames it as “the first we’ve built for more than software engineering.”

The benchmarks, read honestly

xAI published four coding benchmarks. Here’s the complete picture, including the models that outscore Grok 4.5 on xAI’s own charts:

Three things stand out.

Against Claude Opus 4.8, the model Musk named as the comparison point, Grok 4.5 goes two for four. It wins DeepSWE 1.0 by about 6 points and edges Terminal Bench by 4.4, but loses DeepSWE 1.1 by 6 and SWE Bench Pro by 4.5. “Opus-class” is a fair label. “Beats Opus” is not. We run the full head-to-head in Grok 4.5 vs Claude Opus 4.8.

Anthropic’s Claude Fable 5 (max) tops all four charts, on xAI’s own page. Nobody at xAI claimed frontier supremacy, and the charts are upfront about it. Grok 4.5 competes on cost and speed, not peak capability.

Provenance matters. xAI notes that competitor figures come from “the respective developers’ published system cards or benchmark leaderboards,” and the DeepSWE evals were created by Datacurve with runs handled by outside parties. That’s more transparent than pure self-reporting, but there’s still no fully independent third-party evaluation of Grok 4.5 as of this writing. Our benchmarks deep dive tracks what’s verified and what isn’t.

Token efficiency is the real story

The most interesting number in the release isn’t a benchmark score. It’s output volume. On SWE Bench Pro, Grok 4.5 resolves tasks using an average of 15,954 output tokens. Opus 4.8 (max) uses 67,020 on the same benchmark, about 4.2x more.

Output tokens are what you pay for at $6 per million, and they’re what you wait for at 80 tokens per second. A model that solves a task in a quarter of the tokens is cheaper and faster in ways a leaderboard score doesn’t show. xAI claims “roughly 2x the token efficiency of comparable leading models” overall, solving tasks “in under half the number of steps.”

For agentic workloads, where a single session can burn hundreds of thousands of tokens across tool calls, this changes the cost math more than the sticker price does. We work through the numbers in Grok 4.5 pricing explained.

How to start using Grok 4.5

You have four paths, three of them free for now:

1. Grok Build. Grok 4.5 is the default model, and xAI is “offering free Grok 4.5 usage for a limited time.”
2. Cursor. Available on all plans across desktop, web, iOS, CLI, and the SDK, with doubled usage for the first week. Setup steps in our Cursor guide.
3. Grok CLI. Install with curl -fsSL https://x.ai/cli/install.sh | bash.
4. The API. Grab a key from the xAI console and call the Responses endpoint:

curl -s https://api.x.ai/v1/responses \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-4.5",
    "input": "Summarize the breaking changes in this OpenAPI diff."
  }'

The full walkthrough, including streaming and Apidog-based testing, lives in how to use the Grok 4.5 API. If you’d rather not spend API credits while exploring, the free paths are covered in how to use Grok 4.5 for free.

Once you’re calling the endpoint, Apidog helps you keep the integration honest: save the request with environment variables for your key, assert on response fields, A/B grok-4.5 against grok-4.3 on your real prompts, and spin up a mock of the response shape so your frontend team isn’t blocked while you tune prompts. Download Apidog free to follow along.

How Grok 4.5 fits the 2026 model landscape

Grok 4.5 lands in a crowded mid-frontier tier. It’s priced below Claude Opus 4.8 ($5 input / $25 output per million) and far below Claude Fable 5, while scoring within a few points of Opus on coding benchmarks. Against open-weight rivals like GLM 5.2, it wins the published coding evals but costs more per token.

The pitch is intelligence per dollar per second, not a crown. If your bottleneck is agent throughput and API spend, that pitch is worth testing. If you need peak reasoning on the hardest problems, the same charts xAI published point you elsewhere.

For background on the models it’s chasing, see what is Claude Opus 4.8 and our earlier coverage of Grok 4.1, which traces how xAI got here.

FAQ

Is Grok 4.5 free to use? For a limited time, yes, in Grok Build and Cursor. API usage is paid at $2/$6 per million tokens. All current free paths are listed in our free-access guide.

What is the Grok 4.5 model ID? grok-4.5, with a dot. Using grok-4-5 in an API call returns a model-not-found error.

What is the Grok 4.5 context window? 500,000 tokens. The older grok-4.3 offers 1M, so pick by workload, not recency.

Is Grok 4.5 better than Claude Opus 4.8? It wins two of the four benchmarks xAI published and loses the other two, at roughly 40% of Opus’s input price. Faster and cheaper, yes. Stronger across the board, no.

Can I use Grok 4.5 in the EU? Not yet. xAI says EU availability across products and the API console is expected in mid-July 2026.