AI/TLDR

Grok 4.5

SpaceXAI's July 2026 Grok flagship, positioned as an Opus-class coding and agent model trained alongside Cursor.

Overview

Grok 4.5 is SpaceXAI's Grok flagship, announced on July 8, 2026 and generally available from July 9, 2026 through Grok Build, the SpaceXAI console, and Cursor on all plans. The company positions it as its strongest model for coding, agentic tasks, and knowledge work, and Elon Musk described it as an Opus-class model that is faster, more token-efficient, and lower cost than Anthropic's Claude Opus 4.8.

The model is a mixture-of-experts network trained jointly with Cursor after SpaceXAI's acquisition of the coding platform, with a training corpus that includes trillions of tokens of Cursor data on real user interactions with codebases and software tools. It accepts text and image input and is targeted at long-horizon coding and agent workflows where efficient tool use matters.

Grok 4.5 is priced at $2.00 per 1M input tokens and $6.00 per 1M output tokens, with cached input reads at $0.50 per 1M tokens. SpaceXAI reports a serving speed of approximately 80 tokens per second and roughly 4× fewer output tokens per task than Claude Opus 4.8 on the SWE-Bench Pro coding evaluation.

Released2026-07-09
LicenseProprietary
WeightsAPI only
ArchitectureMixture-of-Experts
ModalitiesText, Vision
StatusAvailable

Benchmarks

  1. Terminal-Bench 2.183.3%
  2. SWE-Bench Pro (resolve rate)64.7%
  3. DeepSWE 1.0 (pass@1)62%
  4. DeepSWE 1.1 (mini-swe-agent harness)53%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

Pricing

Input$2.00 / 1M tokens
Cached input$0.50 / 1M tokens
Output$6.00 / 1M tokens

Pricing source ↗

Strengths

  • Opus-class quality on coding and agent benchmarks at $2/$6 per 1M tokens
  • Very high token efficiency — ~4× fewer output tokens than Claude Opus 4.8 on SWE-Bench Pro
  • Deep Cursor integration, with the model trained on trillions of tokens of Cursor coding data
  • Serves at ~80 tokens per second on SpaceXAI infrastructure
  • Available through Grok Build, the SpaceXAI console, and Cursor from day one

Best for

  • Reach for it for agentic coding inside Cursor and Grok Build on a Claude Opus-scale budget.
  • Reach for it for long-horizon software engineering, refactors, and multi-step agent workflows where output-token cost matters.
  • Reach for it for high-throughput coding agents that must resolve issues in a small number of steps.

How to access

ProviderModel ID
SpaceXAI console ↗grok-4-5
Cursor ↗grok-4-5

Grok (flagship) — every version

The full lineage of the Grok (flagship) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.

VersionReleasedContextLicense
Grok 4.5current2026-07-09Proprietary
Grok 4.32026-04-301MProprietary
Grok 4.202026-03Proprietary
Grok 4.12025-11-17Proprietary
Grok 42025-07-09Proprietary
Grok 32025-02-17Proprietary
Grok 22024-08-20Open weights
Grok 1.52024-05-15Proprietary
Grok 12023-11-03Apache-2.0

FAQ

When was Grok 4.5 released?

SpaceXAI announced Grok 4.5 on July 8, 2026 and made it generally available on July 9, 2026 through Grok Build, the SpaceXAI console, and Cursor on all plans. EU availability was noted to follow later in July.

How much does Grok 4.5 cost?

SpaceXAI prices Grok 4.5 at $2.00 per 1M input tokens and $6.00 per 1M output tokens, with cached input reads at $0.50 per 1M tokens. Elon Musk described the pricing as substantially lower than Claude Opus 4.8's $5/$25 per 1M tokens.

How is Grok 4.5 related to Cursor?

SpaceXAI trained Grok 4.5 jointly with Cursor after acquiring the coding platform. The training corpus includes trillions of tokens of Cursor data capturing real user interactions with codebases and software tools, and Grok 4.5 is available in Cursor on all plans from day one.

How efficient is Grok 4.5 relative to other frontier models?

On the SWE-Bench Pro coding evaluation, SpaceXAI reports Grok 4.5 uses roughly 15,954 output tokens per task versus about 67,020 for Claude Opus 4.8 (max) — approximately a 4× reduction in output tokens for comparable resolve rates. The model serves at around 80 tokens per second on SpaceXAI infrastructure.