Why this matters now
On July 9 — the same day OpenAI released GPT-5.6 Sol to the public — SpaceXAI quietly launched Grok 4.5, its first model since going public and acquiring the AI coding startup Cursor. It’s the first major frontier model trained jointly with a coding IDE, using trillions of tokens of real developer data from Cursor’s user base.
The headline numbers are competitive: $2/$6 per million tokens, a 1.5-trillion-parameter MoE architecture, and xAI’s claim of “Opus-class” performance. But the launch is also the messiest of the year — no model card, a withdrawn benchmark after training data contamination, and a mixed independent evaluation picture.
This post covers what actually shipped, where the benchmarks are real vs inflated, and whether Grok 4.5 changes the routing math for builders.
What shipped
| Spec | Grok 4.5 |
|---|---|
| Parameters | ~1.5T MoE (exact active count undisclosed) |
| Context window | 500K tokens (down from 1M on Grok 4.3) |
| Modalities | Text + image → text |
| API pricing | $2.00/M input, $6.00/M output |
| API model ID | grok-4.5 (+ aliases grok-4.5-latest, grok-build-latest) |
| Availability | Grok Build, Cursor, xAI API, Microsoft Office add-ins, model gateways |
| EU access | Restricted at launch |
| Model card | None published |
| Training partner | Cursor (trillions of tokens of developer data) |
At $2/$6 per million tokens, Grok 4.5 matches Terra on input ($2.50) and undercuts Sonnet 5 standard pricing ($3/$15) on both. It’s the cheapest “Opus-class” model on output by a significant margin — cheaper than Mistral Medium 3.5 on output ($6 vs $7.50).
The benchmark picture — xAI’s version vs reality
xAI published a chart showing Grok 4.5 beating Claude Opus 4.8 on 2 of 4 benchmarks and losing on the other 2. Their framing: “Opus-class.” That’s defensible for a model that splits four head-to-heads roughly evenly — it’s a different claim than “beats Opus,” which is how some coverage has spun it.
| Benchmark | Grok 4.5 | Opus 4.8 | GPT-5.6 Sol | Fable 5 |
|---|---|---|---|---|
| SWE-bench Pro | 64.7% | 69.2% | — | 80.4% |
| Terminal-Bench 2.1 | — | 74.6% | 88.8% | — |
| Token efficiency | 4.2x better | baseline | — | — |
| DeepSWE 1.1 (neutral) | 53% | — | — | 70% |
Token efficiency is the real story. On SWE-bench Pro, xAI reports Grok 4.5 resolves the average task using roughly 15,954 output tokens, compared to 67,020 for Opus 4.8 in max-effort mode — a 4.2x efficiency gap. Against GPT-5.5: 1.9M total tokens per task vs 6.2M. Against Fable 5: 1.9M vs 7.2M. This means Grok 4.5’s cost advantage compounds beyond the per-token rate — you pay less per token and use fewer of them.
On the neutral DeepSWE 1.1 harness (which tests models outside vendor-optimized conditions), Grok 4.5 drops to 53% against Fable 5’s 70%. That gap suggests the SWE-bench Pro score was achieved under conditions that favor Grok’s architecture, and the real-world spread is wider than the headline number.
The Cursor problem
The collaboration with Cursor is both the most interesting and most concerning aspect of this launch.
Grok 4.5 was trained on “trillions of tokens of real developer data” from Cursor’s users. This is unprecedented — no other major model has been trained on a coding IDE’s raw interaction data. The result should be a model that understands how developers actually work, not just how they write code.
But the launch was marred by CursorBench, Cursor’s own internal evaluation suite. Cursor disclosed that “a snapshot of its own codebase was accidentally included in Grok 4.5’s training data,” making CursorBench scores unreliable. The benchmark was withdrawn from launch materials. This doesn’t invalidate the SWE-bench Pro or DeepSWE scores — those are external evals — but it means the metric Cursor itself designed to measure Grok’s code quality is unusable.
Beyond the data contamination: xAI published no model card for Grok 4.5, breaking its own pattern from Grok 4 and 4.1. Independent evaluator Artificial Analysis clocked the model’s hallucination rate at 54% on its Omniscience eval — up from 25% on the prior Grok version. That’s a red flag for production use.
Where it fits in your stack
| Use this if… | Best pick before Grok 4.5 | Does Grok 4.5 change this? |
|---|---|---|
| Highest autonomous coding accuracy | Claude Fable 5 (80.4% SWE-bench) | No — still trailing by 15+ points |
| Best value agentic coding | Claude Sonnet 5 ($3/$15) | Yes — Grok 4.5 at $2/$6 with 4.2x token efficiency undercuts on both price and total spend |
| High-volume coding with moderate accuracy | GPT-5.6 Terra ($2.50/$15) | Yes — Grok 4.5 is cheaper on output and more token-efficient |
| Token-sensitive workloads (huge output) | — | Grok 4.5 wins on token efficiency — 4.2x fewer output tokens than Opus 4.8 |
| Self-hosted / open-weight | Mistral / GLM-5.2 | No — Grok 4.5 is not open-weight |
The strongest argument for Grok 4.5 is the cost-per-task math. At $2/$6 with 4.2x fewer output tokens, a coding task that costs $0.45 on Sonnet 5 might cost $0.12 on Grok 4.5 — a 73% reduction. The caveat is hallucination risk (54% on one independent eval) and no EU availability.
The SpaceXAI context
Grok 4.5 is Elon Musk’s first model launch since SpaceXAI’s public listing and its $2.5B acquisition of Cursor. The company now leases compute to Anthropic and Google — its direct competitors — for an estimated $920M/month from Google alone. The irony: Anthropic’s models may be running on SpaceXAI hardware while competing with Grok in the same API marketplace.
xAI’s valuation post-IPO is undisclosed but significantly below OpenAI ($300B), Anthropic ($200B), and Zhipu ($128B). Grok 4.5 needs to win on price and efficiency, not headline benchmarks — and the data suggests that’s a viable strategy, as long as the hallucination rate comes down in future versions.
Decision framework
Use Grok 4.5 if:
- You’re cost-optimizing a high-volume coding pipeline and can absorb moderate hallucination risk
- You’re already in the Cursor ecosystem — it’s the native model
- Token spend is your primary constraint (large output tasks)
Wait or skip if:
- You need guaranteed reliability — the 54% hallucination rate on independent evals is concerning
- You’re in the EU — restricted at launch
- You need a model card, safety evaluation, or transparent documentation — none published
- Your use case requires the highest accuracy tier — Fable 5, Opus 4.8, and Sol still lead
The bottom line: Grok 4.5 is the most cost-efficient “Opus-class” model on the market at $2/$6 with 4.2x token efficiency, but the lack of a model card, a withdrawn benchmark from its own training partner, and a doubling of the hallucination rate from the prior version make it a high-risk deployment for production. It’s a model worth testing and monitoring — not a default.
Related reading
- GPT-5.6 Sol Goes Public
- The Agent Model Price Guide, July 2026
- Claude Sonnet 5: Agentic Coding at Opus-Level for Half the Price
- Tencent Hy3: A 295B Open MoE Model Built for Agents
Sources
- TechCrunch — SpaceXAI releases Grok 4.5, which Elon describes as an ‘Opus-class’ model
- Axios — SpaceXAI launches new model, Grok 4.5
- KingY.ai — Grok 4.5 benchmarks, pricing, context, and the Opus claim
- Hany AI — Grok 4.5: what it’s actually good at (and where it isn’t)
- Substack — Model Drop: Grok 4.5 (Jake Handy)
- Unrot — Top 10 AI News July 10
- AI Tools Recap — AI News July 10
About the author
Charles Jasthyn De La Cueva is a full-stack developer and the founder of Open TechStack. He writes about AI engineering, developer tools, and practical model evaluation — grounded in real workflows, not press releases.
