Grok 4.5 is the latest large language model from xAI, released on July 8, 2026, and positioned as the company’s smartest model to date. Built to excel at coding, agentic tasks, and knowledge work, it’s the first model xAI has shipped since the company went public and acquired the AI coding editor Cursor. Elon Musk has framed it as a direct challenger to Anthropic and OpenAI’s top-tier models, describing it as “an Opus-class model, but faster, more token-efficient and lower cost.”
In this article we’ll discuss what Grok 4.5 actually is, how it performs on the benchmarks that matter for real-world software engineering, why its pricing and token efficiency are turning heads, and what its release tells us about the increasingly crowded frontier AI race. We’ll also look at where the model is available today, what limitations still apply, and what xAI has signaled about future releases.
TL;DR Snapshot
Grok 4.5 is xAI’s new flagship model aimed squarely at developers and knowledge workers rather than casual chatbot users. Trained alongside Cursor on datasets spanning coding, science, engineering, and math, it posts competitive scores on major software engineering benchmarks, undercuts rival pricing significantly, and uses far fewer output tokens per task than comparable models. It’s available now in Grok Build, Cursor, and the xAI console, though EU users will have to wait until later in July.
Key takeaways include…
- Grok 4.5 is priced at $2 per million input tokens and $6 per million output tokens, dramatically undercutting rival offerings from Anthropic and OpenAI.
- The model is highly token-efficient, resolving SWE Bench Pro tasks with roughly 4.2 times fewer output tokens than Opus 4.8.
- Benchmark results are strong but mixed, as Grok 4.5 leads on some coding evaluations but trails Anthropic’s Claude Fable and OpenAI’s GPT 5.6 Sol on others.
Who should read this: Developers, engineering leaders, AI enthusiasts, and anyone evaluating frontier models for coding or agentic workflows.
What Is Grok 4.5?
Grok 4.5 is xAI’s newest large language model, and it represents a deliberate shift in the company’s strategy. Grok began life in 2023 as the chatbot attached to X, but with this release, xAI is presenting it as a serious production tool for coding, multi-step agentic workflows, and long-form knowledge work, according to an analysis by FullStack Labs.
The release is notable for its timing and context. As Axios reported, this is the company’s first model launch since going public and acquiring Cursor, the popular AI coding editor. That acquisition wasn’t just a business move, it shaped the model itself. According to xAI’s official announcement, Grok 4.5 was trained alongside Cursor on datasets spanning coding, science, engineering, and math, and it was trained across tens of thousands of NVIDIA GB300 GPUs. Reporting from Fello AI indicates that the model is built on xAI’s 1.5 trillion parameter V9 foundation, and incorporates real Cursor developer session data, giving it a training signal grounded in how engineers actually work.
Availability is broad but not quite universal. Per xAI’s announcement, the model is live in Grok Build, in Cursor on all plans, and through the xAI API console, with free usage offered for a limited time in Grok Build and Cursor. It isn’t yet available in the EU, though xAI expects EU availability before the end of July.
How It Performs: Benchmarks and the Fine Print
xAI’s headline claim is that Grok 4.5 excels at real engineering tasks, and the numbers it published are genuinely competitive, though they don’t exactly tell a story of dominance. According to the benchmark results in xAI’s announcement, Grok 4.5 scored 62.0% on DeepSWE 1.0, ahead of Claude Opus 4.8’s 55.75% but behind Claude Fable’s 66.1% and GPT 5.5’s 64.31% (GPT 5.6 Sol was not yet publicly available for testing). On Terminal Bench 2.1, it hit 83.3%, essentially neck and neck with GPT 5.5’s 83.4% and just behind Fable’s 84.3%. Its standout result came on SWE Marathon, where its 29.0% resolution rate topped every listed competitor, including Opus 4.8 at 26.0% and Fable at 24.0%.
The picture is less flattering elsewhere though. On DeepSWE 1.1, Grok 4.5’s 53% trailed Fable’s 70%, GPT 5.5’s 67%, and Opus 4.8’s 59%, and on SWE Bench Pro it landed at 64.7% against Fable’s 80.4% and Opus 4.8’s 69.2%. MarkTechPost’s coverage put it plainly, xAI’s prose says the model exceeds comparable leading models, but its own chart is more mixed. And importantly, Grok 4.5 is said to be an Opus-class competitor, meaning it’s not as powerful as Anthropic and OpenAI’s flagship models, Claude Fable and GPT 5.6 Sol.
Independent rankings back up the “very good, not quite best” framing. According to Fello AI’s roundup, Grok 4.5 ranks fourth on the Artificial Analysis Intelligence Index with a score of 54, placing it above every open-weight model and all of Google’s Gemini models. For a company playing catch-up in the frontier race, that’s a meaningful result, especially given the pricing.
The Real Story: Price and Token Efficiency
If benchmarks are where Grok 4.5 competes, economics are where it thrives. Axios reports that the model is priced at $2 per million input tokens and $6 per million output tokens, compared to Claude Opus 4.8 at $5 per million input tokens and $25 per million output tokens. That’s a 60% discount on input and a 76% discount on output relative to the model Musk keeps comparing it to.
The cost advantage compounds with token efficiency. xAI claims the model delivers twice the token efficiency of the latest leading models at the same tasks, while also being served at speeds of 80 tokens per second. Per MarkTechPost, Grok 4.5 resolved SWE Bench Pro tasks using an average of 15,954 output tokens, while Opus 4.8 used 67,020 on the same benchmark, roughly 4.2 times more. Since output tokens are what you pay the most for, and what you wait the longest on, fewer tokens per task means both lower bills and lower latency.
TechCrunch notes that token costs have been a growing concern for AI consumers, which makes this efficiency-first positioning a shrewd one. A model that’s slightly behind the best on raw capability but dramatically cheaper per completed task is a genuinely attractive proposition for teams running agents and coding assistants at scale.
What This Means for the AI Race
Grok 4.5 arrived during a frenzied week for the industry. TechCrunch reported that OpenAI planned to release (and ultimately followed through on releasing) GPT 5.6 Sol, Terra, and Luna the very next day. The launch had previously been limited at the request of the Trump administration over security concerns.
There’s also an unusual wrinkle in xAI’s position. As Axios pointed out, Grok 4.5 was trained using the same compute capacity the company leases to competitors Anthropic and Google, meaning xAI may eventually have to choose between training its own models and collecting rent from their rivals.
Even so, xAI isn’t showing any signs of slowing down. According to Fello AI, Musk has said that Grok 4.5 isn’t yet running on xAI’s internally developed C/C++ inference stack built for the GB300 hardware, and he expects speeds to double or more once it does, with another step-change release signaled for the following month. One important open question remains for cautious adopters though. FullStack Labs observes that Grok 4.5 launched with less public detail on governance, data retention, and safety behavior than recent rivals, which may keep it relegated to lower-risk internal workflows for now.
Frequently Asked Questions
xAI is the artificial intelligence company founded by Elon Musk. It develops the Grok family of AI models and went public in 2026, acquiring the AI code editor Cursor along the way. Grok 4.5 is its newest and most capable model.
Cursor is a popular AI-powered code editor that xAI acquired earlier in 2026. Grok 4.5 was trained alongside Cursor, and real developer session data from the tool informed the model’s training, which is part of why it’s so focused on software engineering tasks.
Grok 4.5 is xAI’s latest large language model, released on July 8, 2026. It’s designed for coding, agentic tasks, and knowledge work, and it’s the company’s first release since going public and acquiring Cursor.
Agentic tasks are workflows where an AI model works autonomously over multiple steps, using tools, reading files, and making decisions to complete a larger goal rather than just answering a single question. Coding agents that fix bugs across a whole repository are a common example.
The model is priced at $2 per million input tokens and $6 per million output tokens through the xAI API. For comparison, Claude Opus 4.8 costs $5 per million input tokens and $25 per million output tokens.
Not yet. It’s available in Grok Build, in Cursor on all plans, and via the xAI API console, but it isn’t available in the EU at launch. xAI expects EU availability by the end of July 2026.
It depends on the task. It leads competitors on the SWE Marathon benchmark and ranks fourth on the Artificial Analysis Intelligence Index, but Anthropic’s Claude Fable and OpenAI’s GPT 5.6 Sol outscore it on several other evaluations. Its strongest argument is delivering near-frontier performance at a much lower cost per task.
Other Enterprise AI Articles You May Be Interested In
Meta Enters the AI Image Race: Everything You Need to Know About Muse Image
GPT-5.6 Is Finally Public: Why the Government Delayed OpenAI’s Most Powerful AI Models
Nvidia Stock Is Cheaper Than the S&P 500 for the First Time Since 2019
Microsoft Frontier Company: Why Big Tech Is Spending Billions on AI Deployment
