Grok 4.5 benchmarks, pricing, and where to use it

What Grok 4.5 is: SpaceXAI's model for coding and agents

Grok 4.5 key facts

Maker

SpaceXAI (its developer platform is still called the xAI API)

Release date

July 8, 2026

Built for

Coding, agentic tasks, knowledge work

Trained with

Cursor, on tens of thousands of NVIDIA GB300 GPUs

Design

Mixture-of-experts, according to Cursor

API model ID

grok-4.5, with a 500,000-token context window

Grok 4.5 is a SpaceXAI model aimed at coding, agentic tasks and knowledge work, and Cursor helped train it. An agentic task is one where the model plans several steps, calls tools and checks its own results instead of replying once. SpaceXAI called it the company's strongest model at launch in July, before two newer Grok models arrived.

SpaceXAI released Grok 4.5 on July 8, 2026

Grok 4.5 went public on July 8, 2026, through the xAI API, Grok Build and Cursor. Cursor's post and xAI's API release notes both carry that date, as did the announcement page at launch, though on September 29, 2026, the live page showed July 16. SpaceXAI is the name the Grok maker now uses on x.ai, while its developer docs still say "xAI API" and describe Grok 4.5 as SpaceXAI's model on that API.

In a post on X quoted by TechCrunch on July 8, Elon Musk called Grok 4.5 an "Opus-class" model, a nod to Anthropic's Claude Opus line. He then said SpaceXAI's internal assessment put it roughly level with Opus 4.7, only much faster. That is the maker's own comparison, not a published benchmark.

Cursor co-trained Grok 4.5 on trillions of its own tokens

Cursor says it trained Grok 4.5 together with SpaceXAI on a mix that included trillions of tokens of its own data, which record how users work with codebases and software tools. The company describes the model as a mixture-of-experts design, meaning only a few specialist sub-networks, called experts, handle each piece of input. Reinforcement learning on hard problems in realistic environments then taught it to investigate, use tools, recover from mistakes and check its results.

SpaceXAI supplied the scale: training ran across tens of thousands of NVIDIA GB300 GPUs, and the reinforcement learning stage covered hundreds of thousands of tasks centered on multi-step software engineering. Cursor also flagged a problem with its own test. An earlier snapshot of the Cursor codebase ended up in the training data by accident, which gives Grok 4.5 an advantage on CursorBench, and Cursor says the exact impact is unclear.

Grok 4.5 benchmarks: tops SWE Marathon, third on DeepSWE 1.0

DeepSWE 1.0 scores (pass@1, SpaceXAI's chart)

Fable (max), GPT-5.5 (xhigh) and Grok 4.5 sit within 4.1 points of each other (66.1 − 62.0). Opus 4.8 (max) is 15.63 points above Opus 4.7 (max): 55.75 − 40.12.

0204060

66.1Fable (max)

64.31GPT-5.5 (xhigh)

62.0Grok 4.5

55.75Opus 4.8 (max)

40.12Opus 4.7 (max)

On SpaceXAI's own charts, Grok 4.5 leads one coding benchmark, SWE Marathon, and finishes third or fourth on the other four. Its clearer edge is efficiency, since it writes far fewer tokens per task than Opus 4.8, and the independent tester Artificial Analysis found the same pattern. A token is a small chunk of text, and API bills are counted in tokens.

DeepSWE and Terminal Bench vs GPT-5.5 and Opus 4.8

DeepSWE is a software-engineering test created by Datacurve and scored as pass@1, the share of tasks a model solves on its first try. Terminal Bench checks how well an agent works in a command-line terminal, and SWE Marathon is a set of 20 long, multi-step tasks across software engineering and related technical fields. SpaceXAI's chart also includes SWE Bench Pro, another software-engineering test.

Grok 4.5 is first on SWE Marathon at 29.0%, three points ahead of Opus 4.8 (max), but third on DeepSWE 1.0 at 62.0%, behind Fable (max) and GPT-5.5 (xhigh). That SWE Marathon lead covers only the Opus and Fable models, because SpaceXAI's chart gives no GPT-5.5 score for that test.

On Terminal Bench 2.1, Grok 4.5 lands 0.1 point behind GPT-5.5 and 4.4 points ahead of Opus 4.8. On DeepSWE 1.1, which Datacurve ran in one shared test setup (the mini-swe-agent harness), it slips to fourth, below Opus 4.8 (max).

Grok 4.5 against GPT-5.5, Opus 4.8 and Fable

The rival scores use the settings named on SpaceXAI's chart: xhigh for GPT-5.5 and max for Opus 4.8 and Fable. SpaceXAI says rival figures come from each developer's published system cards or leaderboards, so the "Grok 4.5 against GPT-5.5, Opus 4.8 and Fable" table is not one controlled run. For what changed in Opus 4.8, the model that beats Grok 4.5 on SWE Bench Pro and DeepSWE 1.1, see Claude Opus 4.8: Pricing, Fast Mode, Workflows.

Token efficiency: a quarter of Opus 4.8's output per task

On SpaceXAI's chart, Grok 4.5 averages 15,954 output tokens per SWE Bench Pro task, about 4.2 times fewer than the 67,020 that Opus 4.8 (max) uses. Opus 4.8 (max) resolves more of those tasks, 69.2% against 64.7%, so Grok 4.5 gives up 4.5 points of resolve rate in exchange for writing under a quarter of the output.

SpaceXAI also claims roughly twice the token efficiency of comparable leading models, with tasks finished in under half the steps. That broader claim names no specific rival, so treat it separately from the 4.2 times figure, which applies only to Opus 4.8 (max). SpaceXAI serves the model at 80 tokens per second, a speed it calls fast-model speed.

Independent Grok 4.5 reviews and hallucination rates

Artificial Analysis ranked Grok 4.5 fourth on its Intelligence Index at launch, but measured a hallucination rate more than double Grok 4.3's. The independent benchmarking firm gave it a launch score of 54, behind only Fable 5, GPT-5.5 and Opus 4.8, and 16 points above Grok 4.3. Running inside Grok Build, Grok 4.5 scored 76 on the firm's Coding Agent Index, level with GPT-5.5 (xhigh) in Codex.

On the firm's AA-Omniscience knowledge test, accuracy rose from 35% for Grok 4.3 to 52% for Grok 4.5, while the hallucination rate climbed from 25% to 54%. Artificial Analysis calls this a common pattern, where larger models know more and also answer with more confidence. For fact-heavy work, check the model's answers against a source.

Grok 4.5 pricing: $2 input and $6 output per million tokens

Grok 4.5 pricing (xAI docs, September 29, 2026; fast variant from Cursor's launch post)

Input

$2.00 per million tokens

Cached input

$0.30 per million tokens

Output

$6.00 per million tokens

Long requests

Higher rate above 200K tokens of context

Fast variant (Cursor, at launch)

$4 input and $18 output per million tokens

Grok 4.5 costs $2 per million input tokens and $6 per million output tokens on the xAI API, unchanged from launch as of September 29, 2026. Input is the text you send, output is what the model writes back, and the bill grows with both. Because Grok 4.5 also finishes a task in fewer tokens than Opus 4.8, the gap in cost per finished task is wider than the gap in per-token prices.

What a coding task costs on the Grok API

At list price, the output tokens for an average SWE Bench Pro task cost about $0.10 on Grok 4.5 and about $1.68 on Opus 4.8 (max). This is our own arithmetic: 15,954 tokens at $6 per million on the Grok API, which xAI's docs call the xAI API, and 67,020 at Anthropic's $25 per million rate for Opus 4.8. It counts output only, so input, cached context and tool calls add to both bills.

Artificial Analysis measured the cost of whole agent runs instead. On its Coding Agent Index, Grok 4.5 in Grok Build cost $2.49 per task, GPT-5.5 in Codex $5.07 and Fable 5 in Claude Code $11.80.

Those figures cover complete agent runs with input included, so they sit well above the output-only estimate. They also come from a different task mix (DeepSWE, Terminal-Bench v2 and SWE-Atlas QnA), so the two sets of numbers are not a like-for-like comparison.

Long prompts cost more, because xAI bills requests past 200K tokens of context at a higher rate. That matters when an agent carries a large codebase in its prompt. For how the Opus side of this comparison is priced across versions, see Claude Opus 4.6 vs 4.7 vs 4.8 pricing compared.

Grok 4.5 free trial at launch

SpaceXAI launched Grok 4.5 with free usage in Grok Build and Cursor for a limited time, without giving an end date. The offer text was still on the announcement page on September 29, 2026, but with no end date the page cannot tell you whether the offer is still running, so check your Grok Build or Cursor billing before relying on it. For Grok's consumer plans, see Grok Pricing 2026: Free, SuperGrok, and API Costs.

Cursor set its own launch terms inside its subscription plans. Individual and team plans came with a sizable Grok 4.5 allowance, doubled for the first week. Cursor also listed a fast variant at $4 input and $18 output per million tokens, twice the standard input rate and three times the output rate.

Where you can use Grok 4.5 now, including the EU

Where Grok 4.5 launched and where it stands (September 29, 2026)

Grok Build

Default model at launch

Cursor

All plans at launch; desktop, web, iOS, CLI and SDK

xAI API

grok-4.5, still listed on September 29, 2026; us-east-1 and us-west-2

EU

API console access since July 17, 2026

Newer models

Grok 4.6 (August 12, 2026), Grok 4.7 (September 21, 2026)

Grok 4.5 launched in Grok Build, in Cursor on every plan and on the xAI API, and EU users have had API console access since July 17, 2026. xAI's API docs still listed the model on September 29, 2026, but it is no longer SpaceXAI's newest. Grok 4.6 and Grok 4.7 have both shipped since launch at the same $2 input and $6 output base rates.

Grok Build, Cursor, and the xAI API

At launch, Grok 4.5 was available in three places: Grok Build, Cursor on all plans, and the SpaceXAI console for API keys. Grok 4.5 became the default model in Grok Build, SpaceXAI's own tool, and the launch post shows it building Excel models, PowerPoint slides and Word documents. In Cursor, the model launched on desktop, web, iOS, the CLI and Cursor's SDK.

On the API the model ID is grok-4.5, with a 500,000-token context window, the most text the model can work with in one request. As of September 29, 2026, xAI's docs list it in the us-east-1 and us-west-2 regions, both in the US, which matters for teams with data-residency rules. For more on SpaceXAI's own tool, see Grok Build on Every Plan: App Builder, Not the CLI.

EU access after the mid-July note

Grok 4.5 was not available in the EU at launch, and xAI's API release notes record EU access in the API console from July 17, 2026. SpaceXAI's launch note had said only that EU availability was expected in mid-July. That note no longer appears on the announcement page as of September 29, 2026.

The July 17 entry covers the API console only. It says nothing about Grok Build or Cursor, so EU teams using those products should confirm access inside each one.

Grok 4.6 and Grok 4.7 have followed Grok 4.5

SpaceXAI released Grok 4.6 on August 12, 2026, and Grok 4.7 on September 21, 2026, so Grok 4.5 is now two versions behind. Grok 4.6 builds on Grok 4.5 with more focus on long-running agents and visual, interactive work. Grok 4.7 moved to a new, larger base model and ships in Cursor, Grok Build and the Grok API.

Both successors keep Grok 4.5's $2 input and $6 output base rates for prompts under 200K tokens, and Grok 4.5 itself was still listed on xAI's API model page at that price on September 29, 2026. Those rates are below Anthropic's Opus 4.8 prices of $5 input and $25 output per million tokens. With the base rates identical, anyone choosing a Grok model today should compare Grok 4.5 with Grok 4.7.