xAI

Grok 4.5

Provider

xAI

Status

Current

Context

500,000 tok

SWE-bench

64.7%

Price

$2 / $6 /MTok

Grok 4.5 is xAI’s current flagship model, launched on 8 July 2026 across Grok Build, the Cursor coding editor and the SpaceXAI console. It is a reasoning-first model aimed squarely at coding, agentic tasks and knowledge work, and it is the first Grok model shipped under the SpaceXAI branding that followed SpaceX’s acquisition of xAI and, in June 2026, the AI coding startup Cursor. Grok 4.5 was trained alongside Cursor — SpaceXAI describes it as “jointly trained” — which shows in its positioning as an agentic-coding model rather than the general reasoning flagship its predecessor was.

Elon Musk framed it as “an Opus-class model, but faster, more token-efficient and lower cost,” comparing it to Claude Opus 4.7. The independent read is more measured: Artificial Analysis places Grok 4.5 5th on its Intelligence Index (~54) behind Fable 5, GPT-5.6 Sol, Opus 4.8 and GPT-5.5 (it was 4th at launch, before GPT-5.6 Sol reached general availability the next day). Its clearest wins are price and token efficiency — $2 / $6 per million tokens, and roughly 4x fewer output tokens than Opus 4.8 to resolve the same coding tasks. Notably, Grok 4.5 drops the native video input that Grok 4.3 introduced and shrinks the context window from 1M to 500K, trading breadth for a sharper coding-and-agentic focus.

Quick specs

TRY GROK 4.5 →

What’s new in Grok 4.5

Grok 4.5 is a pivot rather than a straight-line upgrade over Grok 4.3: it is repositioned around coding and agents, and made two deliberate trade-offs to get there.

Trained alongside Cursor

Grok 4.5’s headline change is the Cursor tie-up. After SpaceX agreed to acquire Cursor (reported at ~$60 billion in June 2026), SpaceXAI trained Grok 4.5 jointly with the Cursor team, using reinforcement learning across hundreds of thousands of coding and agentic tasks. The model launched inside Cursor on all plans the same day, and is aimed at repository-scale software engineering and long-running agent loops rather than general chat.

Much better token efficiency

The most concrete practical gain is output-token economy. On SWE-Bench Pro, SpaceXAI reports Grok 4.5 resolving tasks in an average of 15,954 output tokens against 67,020 for Opus 4.8 (max) — about 4.2x fewer. Because output tokens dominate the bill on agentic runs, this can make Grok 4.5 dramatically cheaper per completed task than its raw $2/$6 price already suggests. Independent Artificial Analysis testing corroborates the direction (~14,000 output tokens per Intelligence Index task, ~60% fewer than Opus 4.8).

Cheaper, but a smaller window and no video

The trade-offs are real. Grok 4.5’s context window drops from Grok 4.3’s 1M to 500K, and it removes native video input — Grok 4.3 was xAI’s first model to accept video, and Grok 4.5 reverts to text-and-image input only. Reasoning runs at “high” effort by default and cannot be disabled, which helps quality but adds latency and tokens on simple calls.

Benchmark performance

SpaceXAI published only four coding/agentic benchmarks at launch, all first-party and none independently verified on day one — normal for a same-day release, but it means these are the vendor’s numbers, not settled results. The honest read is that the story changes with the harness.

The read-through: Grok 4.5 beats Opus 4.8 on two of the four benchmarks SpaceXAI chose to publish (DeepSWE 1.0 and Terminal-Bench 2.1) and loses on the other two (DeepSWE 1.1 by 6 points, SWE-Bench Pro by 4.5 points). In the same table, Fable 5 leads all four. Independent Artificial Analysis ranks Grok 4.5 4th overall on its Intelligence Index (~54), behind Fable 5, Opus 4.8 and GPT-5.5 — so “Opus-class” is fair as a rough tier, but Grok 4.5 is not the outright leader its launch framing implies. Where it does lead cleanly is efficiency and cost, not raw capability.

Pricing breakdown

Grok 4.5 keeps xAI’s aggressive pricing, with token efficiency compounding the discount on agentic work.

Cost comparison with contemporaries

Note that Grok 4.3 is nominally cheaper per token ($1.25/$2.50) and has a larger 1M window — so Grok 4.5 is not a strict upgrade on price. Its edge is quality-per-dollar on coding, where its token efficiency and higher benchmark scores can outweigh Grok 4.3’s lower sticker rate.

How to access Grok 4.5

Via API

Grok 4.5 is generally available with no waitlist on SpaceXAI’s OpenAI-compatible API as grok-4.5, so existing OpenAI integrations can switch with minimal changes. It was not available in the EU at launch, with availability expected mid-July 2026.

from openai import OpenAI
client = OpenAI(base_url="https://api.x.ai/v1", api_key="XAI_API_KEY")

resp = client.chat.completions.create(
    model="grok-4.5",
    messages=[{"role": "user", "content": "Your prompt here"}],
)
print(resp.choices[0].message.content)

Via Cursor and the Grok apps

Grok 4.5 launched inside Cursor on all plans, and in Grok Build. Consumer access in the Grok apps runs through the subscription tiers; SpaceXAI confirmed access on SuperGrok and X Premium+ at launch.

See the Grok app page for the full consumer breakdown and the xAI provider page for company context.

How Grok 4.5 compares

Grok 4.5’s mid-2026 contemporaries are Claude Opus 4.8, Fable 5, GPT-5.5 and its own predecessor Grok 4.3.

vs Claude Opus 4.8

This is the comparison SpaceXAI leaned on. On the four launch benchmarks, Grok 4.5 splits 2–2 with Opus 4.8 on vendor numbers, and Artificial Analysis puts Opus 4.8 slightly ahead on its Intelligence Index. Opus 4.8 keeps the edge on SWE-Bench Pro (69.2% vs 64.7%), a 1M context and a longer independent track record. Grok 4.5’s counter is cost and token efficiency — roughly a quarter of Opus 4.8’s output price and ~4x fewer tokens per resolved task. Choose Opus 4.8 for top-end coding reliability and long context; choose Grok 4.5 for cheaper, high-volume agentic coding.

vs Fable 5

On SpaceXAI’s own table, Fable 5 leads Grok 4.5 on all four benchmarks (notably 80.4% vs 64.7% on SWE-Bench Pro) and sits above it on the Artificial Analysis Index. Fable 5 is the genuine coding ceiling; Grok 4.5 competes on price, not peak capability.

vs GPT-5.5

GPT-5.5 is roughly level with Grok 4.5 on the launch benchmarks (e.g. Terminal-Bench 2.1: 83.4% vs 83.3%) and just ahead on the Artificial Analysis Index. GPT-5.5 brings OpenAI’s agentic tooling and ecosystem; Grok 4.5 undercuts it heavily on price ($2/$6 vs $5/$30) and real-time X access.

vs Grok 4.3

Grok 4.5 supersedes Grok 4.3, but not on every axis. It gains meaningfully on coding, agentic performance and token efficiency, and adds the Cursor integration — but it removes native video input and halves the context window (1M → 500K), and its per-token price is actually higher ($2/$6 vs $1.25/$2.50). For video understanding or the cheapest long-context reasoning, Grok 4.3 remains the better pick; for coding and agents, Grok 4.5 is the upgrade.

The practical consensus

Grok 4.5 is a credible “Opus-class” coding model whose real advantage is economics — cheap tokens and unusually efficient completions — rather than a benchmark crown. It is not the outright best model in any category, but it may be the best value for high-volume agentic coding. See our best AI models and best AI for coding rankings for where it sits across the wider field.

Known limitations

Vendor-only benchmarks at launch. Every headline figure except the early Artificial Analysis read is a SpaceXAI first-party number on its own harness, unverified on day one — treat them as a ceiling.

No video input. Unlike Grok 4.3, Grok 4.5 does not accept video; it is text-and-image input only.

Smaller context window. The 500K window is half Grok 4.3’s 1M, a real regression for very-long-context work.

Not the cheapest Grok. At $2/$6 it costs more per token than Grok 4.3 ($1.25/$2.50); its value case rests on efficiency and quality, not sticker price.

EU availability delayed. Grok 4.5 was not available in the EU at launch, with availability expected mid-July 2026.

Company and safety baggage. As with the wider Grok line, SpaceXAI publishes less safety documentation than its peers, and the Grok apps’ image tools remain under scrutiny after the early-2026 deepfake scandal. See the xAI provider page.

Version history

SpaceXAI is reported to be training Grok 5 on the Colossus 2 supercomputer.

Frequently asked questions

When was Grok 4.5 released?

Grok 4.5 launched on 8 July 2026 across Grok Build, the Cursor coding editor and the SpaceXAI console. It is xAI/SpaceXAI’s current flagship model. It was not available in the EU at launch, with availability expected mid-July 2026.

How much does Grok 4.5 cost?

Via the API, Grok 4.5 is $2.00 per million input tokens and $6.00 per million output tokens, with cached input at $0.50/MTok and a higher rate above 200,000 input tokens. In the apps it is included with SuperGrok and X Premium+ subscriptions.

Is Grok 4.5 really “Opus-class”?

Roughly, as a tier. On SpaceXAI’s own benchmarks it splits 2–2 with Claude Opus 4.8, and independent Artificial Analysis ranks it 5th overall — behind Fable 5, GPT-5.6 Sol, Opus 4.8 and GPT-5.5. So it is in the same conversation as Opus, but it is not the outright leader; its clearest advantage is being much cheaper and more token-efficient.

How is Grok 4.5 different from Grok 4.3?

Grok 4.5 is retuned for coding and agentic work (and was trained alongside Cursor), and it is far more token-efficient. But it drops the native video input Grok 4.3 introduced, halves the context window to 500K, and actually costs more per token ($2/$6 vs $1.25/$2.50). For video or cheap long-context reasoning, Grok 4.3 is still the better choice.

What is Grok 4.5’s context window?

500,000 tokens, with text and image input and text output — down from Grok 4.3’s 1M-token window.

Is Grok 4.5 better than Claude or ChatGPT for coding?

It is competitive and much cheaper, but not clearly ahead. Opus 4.8 still leads SWE-Bench Pro (69.2% vs 64.7% vendor) and Fable 5 leads every launch benchmark. Grok 4.5’s case is value — strong agentic coding at roughly a quarter of Opus 4.8’s output price and ~4x fewer tokens per task. See best AI for coding for the head-to-head.

Written at launch and last verified 9 July 2026. All four headline benchmarks are SpaceXAI’s own launch claims and were not independently verified at release; the Artificial Analysis figures are the earliest independent numbers available. Pricing and availability reflect the 8 July 2026 launch and are subject to change.