Grok 4.5

Overview

Grok 4.5 is xAI's flagship model, released July 8, 2026. It is a mixture-of-experts system, and the most unusual thing about it is how it was trained: jointly with Cursor, on what Cursor's co-launch post describes as trillions of tokens of Cursor data.

That collaboration has a corporate shadow. SpaceX signed an agreement to acquire Anysphere, Cursor's maker, on June 16, 2026 — three weeks before this model shipped — and SpaceX had already absorbed xAI in February. The deal had not closed when Grok 4.5 launched. Read the model's positioning with that in mind.

The positioning itself has shifted. xAI and Cursor frame 4.5 as the first Grok built for more than software engineering, naming knowledge work, finance, legal, and data science. The benchmark it leads with, though, is coding.

Two things went backwards. The context window fell from Grok 4.3's 1,000,000 tokens to 500,000. And the price rose sharply: $2.00 per million input tokens and $6.00 per million output, against $1.25 and $2.50 for 4.3. Grok 4.3 was the cheap frontier model. Grok 4.5 is not.

A note on the vendor. Following the SpaceX acquisition, x.ai presents the company as SpaceXAI, and the developer documentation has followed. What has not moved is the API endpoint, still api.x.ai. This page lists the developer as xAI, the name Grok shipped under.

A note on sourcing. x.ai returns HTTP 403 to automated fetchers, so xAI's own Grok 4.5 announcement could not be read directly for this page. The strongest available sources are docs.x.ai, which is readable, and Cursor's co-launch blog post, written by a co-developer of the model. Vendor benchmark figures below reach us through press relay of the announcement, not from the announcement itself, and are marked as such.

Capabilities

• Agentic tool calling: function calling, web search, X search, and code execution are all documented capabilities.
• Configurable reasoning effort: low, medium, and high, defaulting to high. The published price corresponds to the default.
• Coding: the benchmark suite xAI leads with, and the axis on which Cursor's involvement shows.
• Token efficiency: xAI claims roughly 4.2× fewer output tokens than Claude Opus 4.8 at maximum effort on SWE-Bench Pro — about 15,954 against 67,020.
• Vision input: Artificial Analysis lists text and image input with text output. xAI's own Grok 4.5 documentation page describes only text input explicitly, so treat image support as probable rather than confirmed.

Technical Specifications

• Model ID: grok-4.5
• Architecture: Mixture of experts
• Context window: 500,000 tokens
• Input modalities: Text; images likely, per Artificial Analysis
• Output modalities: Text
• Capabilities: Function calling, web search, X search, code execution, reasoning
• Reasoning effort: low, medium, high (default high)
• API base URL: https://api.x.ai/v1

xAI has not published a knowledge cutoff for Grok 4.5, so none is listed here. The "November 2024" cutoff that appears on xAI's models page belongs to Grok 3 and Grok 4 and must not be attributed to 4.5.

Whether grok-latest has been repointed from grok-4.3 to grok-4.5 is not documented. Pin the explicit model ID if the distinction matters to you.

Performance / Benchmarks

Vendor-published, via press relay

These figures originate with the xAI and Cursor launch materials and reach this page through press coverage of them. They are not independently verified.

xAI also claims the top position on Harvey's Legal Agent Benchmark. Press summaries of the same materials note that Claude Fable 5 at maximum effort leads all four coding benchmarks above, with Grok 4.5 closest on Terminal-Bench 2.1.

Independent, from Artificial Analysis

Both index figures are measured under Intelligence Index v4.1, so the 16 point gap between them is a like-for-like comparison. Artificial Analysis places Grok 4.5 in the top tier of frontier models; its exact ordinal rank varies between the evaluator's own summaries and is not restated here. On AutomationBench-AA it reports 51.4%, ahead of Claude Fable 5 at 48.6% and Claude Opus 4.8 at 48.5%.

Elon Musk characterised the model in press interviews as "an Opus-class model, but faster, more token-efficient and lower cost." The cost claim holds against Claude Opus 4.8 and fails against Grok 4.3.

Grok 4.3 Is Still Here

This is the practical point for anyone choosing between them, and it is easy to miss.

xAI has announced no retirement or deprecation date for grok-4.3. It remains an active model in the documentation, with a 1,000,000 token context window at $1.25 input and $2.50 output per million tokens. Grok 4.5 arrived as a second flagship, not as a forced migration.

Choose 4.5 for reasoning depth, agentic tool use, and coding. Choose 4.3 for long-context work and cost-sensitive volume — a 1M window at $1.25 per million input tokens is a combination 4.5 does not offer at any price.

Use Cases

• Coding agents: the model's strongest published axis, and the reason Cursor ships it on all plans.
• Agentic workflows: function calling, web search, X search, and code execution in one model.
• Knowledge work beyond software: xAI names finance, legal, and data science as targets, and claims a lead on Harvey's Legal Agent Benchmark.
• Token-constrained pipelines: the claimed 4.2× output-token efficiency against Opus 4.8 matters more than headline price when output dominates the bill.
• Long-context analysis: use Grok 4.3 instead. 500K tokens is not a small window, but 4.3 offers twice it at a third of the output price.

Limitations

• Context window halved. 500K against Grok 4.3's 1M. A regression, not a tradeoff the vendor explains.
• Substantially more expensive. Input up 60%, output up 140% versus 4.3.
• No published knowledge cutoff. Do not reuse the November 2024 figure documented for Grok 3 and Grok 4.
• No readable primary announcement. x.ai blocks automated fetchers; vendor benchmark claims on this page reach us via press relay of the launch materials.
• Vision support unconfirmed. Listed by Artificial Analysis, not stated on xAI's Grok 4.5 documentation page.
• Alias behaviour undocumented. It is unclear whether grok-latest now resolves to 4.5.
• Regional availability. Reported as unavailable in the EU at launch, with mid-July 2026 given as the expected date. Press-reported, not confirmed by xAI.
• A "fast" variant at $4 input and $18 output per million tokens is referenced in Cursor's launch post but does not appear in xAI's documented pricing table. Treat it as unconfirmed.
• Training data provenance. The model was trained on trillions of tokens of Cursor data by the co-developer's own account, weeks after SpaceX agreed to acquire Cursor's maker.
• Text output only. Image generation remains in the separate grok-imagine line.

Pricing & Access

API pricing (per 1M tokens)

Grok 4.5's published price corresponds to the default high reasoning effort.

Access

• API: https://api.x.ai/v1
• Grok Build: xAI's developer surface
• Cursor: available on all plans from launch day, as co-developer
• EU: reported unavailable at launch, expected mid-July 2026

Model Retirements & Migration

The May 15, 2026 retirement remains the relevant migration event, and it targets Grok 4.3, not 4.5. On that date xAI retired eight models:

The pattern was consolidation: xAI collapsed the fast/non-fast and reasoning/non-reasoning axes into a single flagship with a reasoning-effort dial. Coding work moved to a dedicated grok-build line. Grok 4.5's arrival does not change any of these targets.

Ecosystem & Tools

• xAI developer documentation - Model reference, pricing, migration guides
• Grok 4.5 model page - Canonical specs
• OpenAI-compatible API - Reachable through OpenAI-compatible SDKs by overriding base_url to https://api.x.ai/v1
• Composer 2.5 - Cursor's own coding model, from the company that co-trained Grok 4.5
• grok-build-0.1 - xAI's dedicated coding model
• grok-imagine-image-quality - Image generation, separate from the Grok text line

Community & Resources

• Grok 4.5 model documentation - Specs, pricing, rate limits
• Cursor: Introducing Grok 4.5 - Co-launch post from the co-developer
• May 15, 2026 model retirement guide - The eight retired models and their targets
• xAI pricing
• xAI - Company site (now branded SpaceXAI)
• Grok on X