Grok 4.5 dropped on July 8, 2026, and it is the first major model release since SpaceX acquired Cursor for $60 billion and rebranded xAI as SpaceXAI. The headline: Grok 4.5 was trained jointly with Cursor, making it the first frontier model built with an IDE integration partner from the ground up.
This is not just a model update. It represents the first time a model vendor and IDE maker co-developed the training process. The result is a coding model that scores 64.7% on SWE-bench Pro, slots between Claude Sonnet 5 at 63.2% and Claude Opus 4.8 at 69.2%, while costing dramatically less per task than most competitors.
Key Specifications
| Spec | Value |
|---|---|
| Release date | July 8, 2026 |
| Developer | SpaceXAI (formerly xAI) |
| Context window | 500,000 tokens |
| SWE-bench Pro | 64.7% |
| Input pricing | $2 per million tokens |
| Cached input | $0.50 per million tokens |
| Output pricing | $6 per million tokens |
| Reasoning | Configurable |
| Training hardware | Tens of thousands of NVIDIA GB300 GPUs |
The 500K context window is one of the largest available for coding models. It allows Grok 4.5 to ingest entire codebases in a single pass, which matters for repository-wide refactoring tasks where understanding the full dependency graph is critical.
The Cursor Co-Training Story
The SpaceX acquisition of Cursor changed everything about how Grok 4.5 was built. Rather than training a general model and then fine-tuning for code, SpaceXAI embedded Cursor engineers directly in the training loop.
What this means in practice: Grok 4.5 understands IDE-specific patterns that other models learn only through post-training. Tab completion, multi-file edits, inline suggestions, and project-aware context retrieval were all part of the base training data and reward signals.
This explains why Grok 4.5 performs best in Cursor specifically. The model was trained to work in that environment. If you use Cursor as your primary editor, Grok 4.5 was built for you in a way no other model can claim.
The co-training also shaped how Grok 4.5 handles multi-file contexts. Rather than treating each file as an independent document, the model learned to track relationships between files the way a developer navigates them in an IDE. Jump-to-definition, find-references, and project-wide symbol understanding are baked into the model’s representations.
SWE-bench Pro Performance
The 64.7% SWE-bench Pro score puts Grok 4.5 in an interesting position:
| Model | SWE-bench Pro | Price Tier |
|---|---|---|
| Claude Opus 4.8 | 69.2% | $5/$25 |
| Grok 4.5 | 64.7% | $2/$6 |
| Claude Sonnet 5 | 63.2% | $1/$5 |
Grok 4.5 beats Sonnet 5 by 1.5 points while costing twice as much per token. But price per token is misleading. Grok 4.5 uses approximately 4.2x fewer output tokens than Opus 4.8 on the same SWE-bench Pro tasks. It is also more concise than Sonnet 5 in most cases.
Analyst estimates put the average cost per coding task at:
- Grok 4.5: $2.49
- GPT-5.5: $5.07
- Fable 5: $11.80
So despite higher per-token pricing than Sonnet 5, Grok 4.5 ends up comparable in per-task cost because it writes tighter code.
Elon Musk also claims Grok 4.5 is “#1 on SWE marathon,” a long-horizon coding benchmark that tests sustained performance across extended sessions. This claim awaits independent verification.
Token Efficiency
This is Grok 4.5’s standout feature. Using 4.2x fewer output tokens than Opus 4.8 means:
- Faster response times (less text to generate)
- Lower costs per task (even at higher per-token rates)
- Less noise in output (tighter, more focused code)
- Better fit for streaming workflows where latency matters
For teams doing heavy automated coding with agents running hundreds of tasks daily, this efficiency gap adds up. A team spending $10,000/month on Opus 4.8 API calls might get comparable results from Grok 4.5 for under $3,000.
The tradeoff: higher hallucination rates have been reported compared to Claude models. Grok 4.5 occasionally invents APIs or uses deprecated methods. Strong test coverage and code review remain essential regardless of which model you use.
Configurable Reasoning
Grok 4.5 supports configurable reasoning, letting you adjust how much thinking the model does per request. This is similar to extended thinking in Claude models but with finer granularity.
Use cases by reasoning level:
- Low reasoning: Tab completions, simple renames, formatting fixes
- Medium reasoning: Standard feature implementation, bug fixes
- High reasoning: Architecture decisions, complex debugging, system design
The cost difference is significant. Low-reasoning requests use minimal output tokens. High-reasoning requests can use 10x more tokens but produce better outputs for hard problems. By matching reasoning level to task complexity, you avoid paying for deep thinking on trivial edits.
Where to Access Grok 4.5
Three channels:
- Grok Build - SpaceXAI’s own coding agent platform
- All Cursor plans - Built-in across Free, Pro, and Business tiers
- SpaceXAI developer console - Direct API access for custom integrations
The Cursor integration is where individual developers will get the best experience. The model understands Cursor’s context retrieval, multi-file edit format, and inline suggestion patterns natively.
For teams building custom tooling, the SpaceXAI developer console provides standard REST API access. Pricing is consistent across all channels.
A mid-July EU release is planned. Until then, EU-based teams need to route through non-EU endpoints or wait for the regional deployment.
Pricing in Context
Here is how Grok 4.5 fits the broader AI API pricing landscape:
| Model | Input/M | Output/M | Cached/M | Avg Task Cost |
|---|---|---|---|---|
| Grok 4.5 | $2.00 | $6.00 | $0.50 | $2.49 |
| Sonnet 5 | $1.00 | $5.00 | $0.30 | ~$2.50 |
| GPT-5.5 | varies | varies | varies | $5.07 |
| Opus 4.8 | $5.00 | $25.00 | varies | ~$10+ |
| Fable 5 | varies | varies | varies | $11.80 |
The cached input pricing ($0.50/M) is particularly attractive. With a 500K context window, you can cache a massive codebase and only pay for incremental outputs on subsequent queries. This makes Grok 4.5 extremely cost-effective for iterative development on the same codebase.
Known Limitations
Be honest about the weak spots:
Hallucination rates are higher than Claude models. The model sometimes generates code using non-existent functions, deprecated APIs, or fictional NPM packages. This is the most commonly reported issue.
EU availability is delayed until mid-July 2026. European developers are temporarily locked out of direct access.
No standalone terminal agent. Unlike Claude Code or similar tools, Grok 4.5 via API does not include autonomous agent capabilities. You need Grok Build or Cursor for the agent experience.
Cursor dependency for peak performance. The joint training means the model shines brightest in Cursor. In other editors or raw API usage, you may not see the full quality gains.
SWE marathon claim unverified. Elon Musk’s “#1 on SWE marathon” assertion has not been independently confirmed.
Who Should Use Grok 4.5
Strong fit:
- Cursor users (the model was literally built for this editor)
- Teams optimizing for cost per task (not cost per token)
- Projects needing 500K context for large codebase ingestion
- Developers who value concise outputs over verbose explanations
- High-volume automated coding pipelines
Consider alternatives:
- Zero hallucination tolerance (use Sonnet 5 or Opus 4.8)
- Terminal-first workflows (Claude Code ecosystem is stronger)
- EU-based teams needing immediate access
- Need maximum accuracy regardless of cost (use Opus 4.8)
For the full comparison landscape, see our best AI coding tools guide.
FAQ
Is Grok 4.5 better than Claude Sonnet 5?
On SWE-bench Pro, yes: 64.7% vs 63.2%. In practice, the gap is small. Grok 4.5 has a slight accuracy edge and a significant advantage in Cursor. Sonnet 5 has lower hallucination rates and works well everywhere. See our full comparison.
What does “trained jointly with Cursor” actually mean?
Cursor engineers were embedded in the training process. Reward signals, training data selection, and evaluation criteria were shaped by how code is edited in Cursor. This is different from post-training fine-tuning. It was part of base model development.
Can I use Grok 4.5 outside of Cursor?
Yes. The SpaceXAI developer console provides standard API access, and Grok Build offers its own IDE experience. But peak performance is in Cursor because that is the training environment.
How does the 500K context compare to competitors?
It is among the largest for coding models. Opus 4.8 and Sonnet 5 offer 200K. GPT-5.6 varies by tier. The 500K window means loading entire service boundaries without chunking.
Is the $2.49 per-task cost accurate?
It is an analyst estimate based on average SWE-bench Pro task token usage. Your actual costs depend on task complexity, reasoning level, and caching behavior. Simple tasks cost less. Complex tasks cost more. The estimate represents a statistical average.