Grok 4.5: SpaceXAI's Opus-Class Model: Benchmarks, Pricing, and the Cursor Flywheel

On July 8, 2026, SpaceXAI (the merged entity of xAI and SpaceX) launched Grok 4.5, its most capable model to date []. Built on a new 1.5-trillion-parameter V9 foundation, the model is positioned as "Opus-class" at a fraction of the cost.

Grok 4.5 arrives with a 500K-token context window, configurable reasoning effort, and a pricing structure that undercuts both GPT-5.5 and Claude Opus 4.8 by over 60%. Early benchmarks show it trading blows with frontier models on coding and agentic tasks while using sharply fewer tokens per task.

This is also the first model released since SpaceXAI's acquisition of Cursor (June 16, 2026), and Cursor's developer data was used in supplemental training. The vertical integration story: one company now owns the compute (Colossus 2), the model (V9), and the coding tool (Cursor) is as significant as the model itself.

Grok 4.5 at a Glance

The Strategic Play: Cursor + SpaceXAI

The most important context for Grok 4.5 isn't in the benchmark numbers: it's in the corporate structure.

The Vertical Stack

SpaceXAI now owns all three layers:

1. Compute: Colossus 2 (100K H100/G100 GPUs), running 7 model training runs simultaneously including a 10-trillion-parameter Grok 5 variant
2. Model: V9 foundation at 1.5T parameters
3. Tool: Cursor (acquired June 16), with developer telemetry and coding data fed back into supplemental training

This creates a data flywheel that no other AI lab has. OpenAI doesn't own a coding tool. Anthropic doesn't own a coding tool. Google has Colab but it's not comparable to Cursor's active user base [].

The Cursor Training Data Caveat

Elon Musk confirmed on June 28 that Grok 4.5 used Cursor data in "supplemental training" []. This is both a strength and a watchpoint:

• Strength: The model was fine-tuned on real developer workflows: exact keystrokes, accept/reject patterns, edit decisions. This likely explains its strong coding agent performance despite being a general model.
• Watchpoint: Cursor's user base skews toward web development (React, TypeScript, Python). Performance on embedded systems, legacy codebases, or niche languages may be weaker than benchmarks suggest.

Benchmark Performance

Intelligence Index (Artificial Analysis)

Grok 4.5 scores 54 on the AA Intelligence Index, placing 4th behind Fable 5 (72), GPT-5.5 (62), and Opus 4.8 (60) []. This is a 16-point improvement over Grok 4.3 (38), making it the largest single-generation leap on the index.

Coding Agent Performance

In Grok Build (SpaceXAI's own coding agent harness), Grok 4.5 scores 76 on the AA Coding Agent Index, tied with GPT-5.5 in Codex and just below Fable 5 in Claude Code [].

The benchmark suite combines DeepSWE, Terminal-Bench v2, and SWE-Atlas QnA. Grok Build performs well on Terminal-Bench v2, which measures agentic terminal use.

Grok 4.5 excels on efficiency:

Grok 4.5 uses 69% fewer tokens than GPT-5.5 and 74% fewer than Fable 5 to achieve comparable coding results. At $2.49 per task, it costs less than half of GPT-5.5 and a quarter of Fable 5 [].

Agentic Knowledge Work

On Snorkel's GDPval+ benchmark (2,000 professional workplace tasks), Grok 4.5 achieved a mean pass rate of 29%, beating GPT 5.5 (22%) and Opus 4.8 (21%) []. This is the strongest signal that Grok 4.5 excels at multi-step professional workflows beyond pure coding.

On the AA Intelligence Index agentic evaluations:

• GDPval-AA v2: Elo 1543 (4th, between Opus 4.8 at 1600 and GLM-5.2 at 1513)
• τ³-Banking: 33% (tops GPT-5.5 at 31%) []

Knowledge & Hallucination

On the AA-Omniscience Index, Grok 4.5 scores 26 (up from 18 for Grok 4.3). Accuracy improved from 35% to 52%, but the hallucination rate also increased from 25% to 54% []. This is a common pattern with larger models: they know more but are more confident in what they "know," including incorrect facts.

Pricing Analysis

Raw API Pricing

Grok 4.5 is aggressively priced:

At $0.31 per Intelligence Index task, Grok 4.5 sits on the cost-performance Pareto frontier, cheaper per task than GLM-5.2 and Kimi K2.6 while scoring higher [].

Cost Comparison for Coding

For coding agent workloads, the cost advantage is even more dramatic:

• Grok 4.5 (Grok Build): $2.49 per coding task
• GPT-5.5 (Codex): $5.07 per task (+104%)
• Fable 5 (Claude Code): $11.80 per task (+374%)

The Long-Context Premium

One important detail: input costs double for contexts exceeding 200K tokens []. Since the context window is 500K, heavy RAG workloads that push past 200K will see effective input prices of $4/1M, still competitive, but worth modeling before scaling.

Availability and Ecosystem

Grok 4.5 is available immediately across:

The API is OpenAI-compatible, so migration is straightforward:

curl https://api.x.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $XAI_API_KEY" \
  -d '{
    "model": "grok-4.5",
    "messages": [{"role": "user", "content": "Explain Grok 4.5 in one sentence."}],
    "reasoning_effort": "medium"
  }'

The Reasoning Dial

Grok 4.5 ships with a configurable reasoning effort parameter (reasoning_effort: none | low | medium | high). This directly controls how many tokens the model spends on internal reasoning before generating a response:

• none: Fast responses, minimal chain-of-thought
• medium (default): Balanced speed and depth
• high: Maximum reasoning, best for complex coding/analysis

In Grok Build, the harness automatically adjusts reasoning effort per task type, a noticeable improvement over manual tuning in Codex or Claude Code.

What About Grok 4.3 Users?

Grok 4.3 was released on April 30, 2026 and remains available. The upgrade path:

The context window reduction from 1M to 500K is notable. For workloads that need the full 1M context, Grok 4.3 remains the cheaper option. For quality-sensitive work, the 16-point intelligence gain justifies the price increase.

Verdict: Who Should Use Grok 4.5

Strongly Consider If:

• You run coding agent workloads at scale. The 69% token efficiency vs GPT-5.5 translates directly to lower bills. At $2.49 per coding task vs $5.07 for GPT-5.5, a team running 10,000 tasks/month saves $25,800.
• You're on Cursor. The direct integration means zero migration cost.
• You need agentic knowledge work (research, analysis, document processing). The GDPval+ results suggest Grok 4.5 handles multi-step professional workflows better than any model at its price point.
• Price sensitivity is high. Grok 4.5 delivers near-frontier quality at less than half the cost of Opus 4.8 or GPT-5.5.

Proceed With Caution If:

• Hallucination is unacceptable. The 54% hallucination rate on AA-Omniscience is concerning. For factual retrieval tasks, Opus 4.8 or GPT-5.5 with retrieval augmentation may be safer.
• You need the full 1M context. Moving from Grok 4.3's 1M to 4.5's 500K is a regression.
• You're in a regulated industry in the EU. Access is delayed until mid-July, and SpaceXAI's enterprise compliance infrastructure is less mature than OpenAI's or Anthropic's.

The Bottom Line

Grok 4.5 is the strongest price-performance play among current frontier models. It doesn't beat Opus 4.8 or GPT-5.5 on raw quality, but it costs a fraction of both while using sharply fewer tokens per task. For cost-sensitive teams running agentic or coding workloads at scale, it's the most practical choice.

The Cursor acquisition adds an unpredictable variable. If the data flywheel works as Musk intends, Grok 4.5's successor will have a training advantage no other lab can replicate. If it doesn't, this is still a very good model at a very good price.

Sources

• Artificial Analysis: Grok 4.5 brings SpaceXAI to the intelligence frontier
• SpaceXAI: Introducing Grok 4.5
• xAI Docs: Pricing
• Snorkel AI: Grok 4.5 Testing Results
• LushBinary: Grok 4.5 Developer Guide
• Digital Applied: Grok 4.5: SpaceX's 1.5T V9 Model
• Fello AI: Grok 4.5 Release Date
• Big Hat Group: xAI Weekly: Grok 4.5 Launches