$ cat wiki/models/grok-4-5.md
Grok 4.5
xAI's enterprise deployment of the V9-class model. 1.5 trillion parameters, Cursor-trained — entered private beta at Tesla and SpaceX on 2026-06-28. The first V9 deployment under an official Grok 4.x product name; distinct from the consumer-first Grok V9-Medium released June 16.
Spec
| Attribute | Value |
|---|---|
| Developer | xai |
| Released | 2026-07-09 |
| Announced | 2026-06-28 |
| Context window | unknown |
| Pricing | $2/M input · $6/M output |
| License | unknown |
| Availability | Grok Build (default model), Cursor (all plans), SpaceXAI console/API; not in EU (regulatory review pending as of 2026-07-09) |
| Parameters | ~1.5T (V9 foundation, same base as Grok V9-Medium) |
| Training focus | V9 foundation + Cursor developer workflow data (co-trained with Cursor; SpaceX acquired Cursor for ~$60B June 2026 — see [sources/blogs/spacex-2026-06-16-cursor-acquisition.md]) |
| Private beta | 2026-06-28, Tesla and SpaceX |
| Throughput | ~80 tokens/second |
Release Date
Benchmarks
xAI published benchmarks at public launch (source):
| Benchmark | Grok 4.5 | Claude Opus 4.8 | Winner |
|---|---|---|---|
| DeepSWE 1.0 | 62.0% | 55.75% | Grok 4.5 |
| DeepSWE 1.1 | 53% | 59% | Opus 4.8 |
| Terminal Bench 2.1 | 83.3% | 78.9% | Grok 4.5 |
| SWE-Bench Pro | 64.7% | 69.2% | Opus 4.8 |
| Token efficiency: Grok 4.5 averages 15,954 output tokens per SWE-Bench Pro task vs. Opus 4.8 at 67,020 — a 4.2× gap. xAI frames this as a major cost advantage. |
Caveat: Musk's original claim "close to, perhaps exceeding Opus" is partially supported. Grok 4.5 wins 2/4 published benchmarks against Opus 4.8. "Opus-class" (same tier) is accurate; "beats Opus" is not uniformly supported by the published data.
Use Cases
- Agentic coding in Grok Build (now default model), Cursor IDE integration
- Enterprise technical reasoning: Tesla engineering, SpaceX software (original beta use cases)
- API-based coding and reasoning workflows via SpaceXAI console
Compared To
| Model | Org | Released | Key Benchmark | Notes |
|---|---|---|---|---|
| Grok 4.5 | xAI | 2026-07-09 (GA) | DeepSWE 1.0 62%, TBench 83.3%, SWE-Pro 64.7% | 1.5T, V9+Cursor, $2/$6/Mtok |
| Grok V9-Medium | xAI | 2026-06-16 | — | Same V9 base, consumer-first |
| Claude Opus 4.8 | Anthropic | 2026-05-28 | SWE-bench Pro 69.2% | Reference for Musk's claim |
| Claude Fable 5 | Anthropic | 2026-06-09 (suspended) | SWE-bench Pro 80.3% | Current frontier (suspended) |
| Grok 5 (training) | xAI | TBD | — | 6T MoE, Colossus 2 |
Significance
Grok 4.5's private beta signals xAI's pattern of deploying models at Musk's first-party companies before external release — accelerating dogfooding feedback loops. Musk's monthly cadence announcement (remainder of 2026) represents a public commitment to aggressive release velocity. The enterprise deployment at Tesla (autonomous driving software) and SpaceX (avionics, Starship) implies high-stakes internal validation at scale.
Musk's AGI comment: Simultaneously with the Grok 4.5 announcement, Musk stated "My estimate of the probability of Grok 5 achieving AGI is now at 10% and rising." This refers to the next-generation 6T model, not Grok 4.5 — treat as speculative CEO signaling (interests.md weight: 0.2×). (source)