Introducing Grok 4.5: SpaceXAI's Smartest Agentic Coding Model Hits 80 TPS and 4.2x Token Efficiency

Grok 4.5 marks SpaceXAI's strongest release to date. Built specifically to tackle software engineering, complex multi-step reasoning, and office productivity, the model was trained in close partnership with the Cursor coding editor. Serving at an extremely fast 80 tokens per second (TPS), it introduces a massive shift in API token economics.

#1Real-World Software Engineering Benchmarks

Grok 4.5 was trained on heavily filtered, domain-focused mixtures of coding, mathematical reasoning, and scientific literature. To prove its capabilities, SpaceXAI has published scores across several rigorous SWE-focused evaluations:

Benchmark Performance Data:

• DeepSWE 1.0 (pass@1): Grok 4.5 scores 62.0%, placing it ahead of Claude Opus 4.8 Max (55.75%) and Opus 4.7 Max (40.12%), while following GPT 5.5 Xhigh (64.31%) and Fable Max (66.1%).
• DeepSWE 1.1: Under Datacurve's mini-swe-agent harness, Grok 4.5 scores 53%, surpassing GLM 5.2 (44%) but trailing Fable Max (70%), GPT 5.5 Xhigh (67%), and Opus 4.8 Max (59%).
• SWE Marathon (pass@1): Grok 4.5 scores 29.0%, topping Opus 4.8 Max (26.0%), Fable Max (24.0%), and Opus 4.7 Max (16.0%).
• Terminal Bench 2.1: Grok 4.5 scores 83.3%, running neck-and-neck with Fable Max (84.3%) and GPT 5.5 Xhigh (83.4%).
• SWE Bench Pro: Grok 4.5 resolves 64.7% of tasks, outperforming Opus 4.7 Max (64.3%), GLM 5.2 (62.1%), and GPT 5.5 Xhigh (58.6%).

#24.2x Token Efficiency: Lower Output Volume, Lower Latency

Beyond raw performance, SpaceXAI focused heavily on token compression. In agentic software loops, verbose models generate thousands of redundant reasoning tokens, driving up latency and bills.

According to the published metrics, Grok 4.5 resolves SWE Bench Pro tasks using an average of only 15,954 output tokens. In comparison, Claude Opus 4.8 (max) requires an average of 67,020 output tokens. This represents a 4.2× reduction in generated tokens for equivalent tasks. When coupled with Grok 4.5's raw speed of 80 tokens per second (TPS), engineering workflows execute in a fraction of the time.

This speed optimization fits neatly into the trends we are tracking on our Top AI Coding Tools for 2026 guide.

#3Competitive Pricing Structure

SpaceXAI is pricing Grok 4.5 aggressively to compete directly for developer API usage:

• Input Tokens: $2.00 per million tokens.
• Output Tokens: $6.00 per million tokens.

Because the model is more concise, developers save not only on the base rate but also on the overall volume of generated tokens, making agentic loops highly viable at scale. This mirrors competitive strategies like the ones we reviewed in the AWS Billion Dollar Forward Deployed Unit launch.

#4Grok Build and Office Automation

Grok 4.5 is now the default reasoning engine in Grok Build. The platform can build complex, multi-sheet Excel models, perform live web research, calculate formulas, and annotate cells with stickies for future human inspection.

In PowerPoint and Microsoft Word, the model functions as a meticulous document designer. It can layout slides using native PowerPoint shapes, arrange charts dynamically, and write high-grade technical documentation.

#5Large-Scale Reinforcement Learning (RL)

SpaceXAI scaled training across tens of thousands of NVIDIA GB300 GPUs, deploying advanced asynchronous reinforcement learning (RL). RL loops cover hundreds of thousands of multi-step software engineering challenges, graded automatically. This enables Grok 4.5 to run long-running agentic rollouts over several hours, exploring code branches independently to arrive at the most optimal, bug-free solution.

#6Getting Started with Grok 4.5

Grok 4.5 is available today inside the Cursor editor (across all subscription plans), Grok Build, and the SpaceXAI developer console.

You can query the model using a standard REST payload:

curl -s https://api.x.ai/v1/responses \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "grok-4.5",
    "input": "Find and fix the bug, then explain it: function median(a){a.sort();return a[a.length/2]}"
  }'

For local command-line operations, you can download the xAI CLI using the following terminal commands:

irm https://x.ai/cli/install.ps1 | iex

curl -sLS https://x.ai/cli/install.sh | bash

Note: Grok 4.5 is not yet available in the EU at launch. Regional availability is scheduled to rollout in mid-July 2026.