Grok 4.5 is SpaceXAI’s newest and most capable AI model, launched July 8, 2026. Priced at $2 per million input tokens and $6 output, it delivers coding performance competitive with Claude Opus 4.8 at roughly one-fifth the cost, though independent benchmarks rank it fourth overall behind Claude Fable 5 and GPT-5.5.
SpaceXAI made Grok 4.5 public on July 9, 2026, one day after announcing it on July 8, positioning the model directly against Anthropic’s Claude Opus line. Elon Musk called it “an Opus-class model, but faster, more token-efficient and lower cost,” as reported by TechCrunch. It is the company’s first major model release since the SpaceX IPO on June 12, 2026 — and it landed the same week as OpenAI’s GPT-5.6 Sol, Terra, and Luna, which we covered in our GPT-5.6 launch report.
- Pricing: $2/M input, $6/M output tokens — vs. Claude Opus 4.7 at $5/$25
- Speed: served at 80 tokens per second
- Token efficiency: 4.2× fewer output tokens than Opus 4.8 (max) on SWE Bench Pro (15,954 vs 67,020 average)
- Availability: Grok Build, Cursor (all plans), SpaceXAI API console; EU access expected mid-July
- Independent ranking: 4th on the Artificial Analysis Intelligence Index (score 54), behind Claude Fable 5, GPT-5.5, and Claude Opus 4.8
What Grok 4.5 Actually Scores — The Key Numbers
Per SpaceXAI’s own published results, Grok 4.5 scores 62.0% on DeepSWE 1.0 (vs. Claude Fable max at 66.1% and GPT-5.5 xhigh at 64.3%) and 83.3% on Terminal Bench 2.1 (vs. GPT-5.5 at 84.4% and Fable max at 84.3%). On SWE Bench Pro it resolves 64.7% of tasks, behind Fable max (80.4%) and Opus 4.8 (69.2%). SpaceXAI also reports a #1 result on Harvey’s Legal Agent Benchmark, though that figure comes from SpaceXAI’s internal test on Harvey’s benchmark, per the official SpaceXAI announcement.
Quick Comparison: Grok 4.5 vs Claude Opus 4.8 vs GPT-5.5
| Benchmark | Grok 4.5 | Claude Fable max | GPT-5.5 xhigh | Claude Opus 4.8 max |
|---|---|---|---|---|
| DeepSWE 1.0 | 62.0% | 66.1% | 64.3% | 55.75% |
| Terminal Bench 2.1 | 83.3% | 84.3% | 84.4% | 78.9% |
| SWE Bench Pro (resolve) | 64.7% | 80.4% | 58.6% | 69.2% |
| API price (in/out per M tokens) | $2 / $6 | — | — | $5 / $25 (Opus 4.7) |
Scores as published by SpaceXAI within each model’s own harness; pricing verified July 2026.
Day-1 Independent Benchmarks Tell a More Complicated Story
Musk claimed a “#1 on SWE marathon” result on X, but independent Day-1 data does not support a #1 overall position. The Artificial Analysis Intelligence Index places Grok 4.5 fourth with a score of 54, behind Claude Fable 5, GPT-5.5, and Claude Opus 4.8 — though it does lead specifically on agentic tool use. More concerning: independent testing found Grok 4.5’s hallucination rate doubled to 54%, up from 25% on Grok 4.3 — a more capable model that is also more confidently wrong.
Where Grok 4.5 wins without dispute is cost. At $0.31 per task on the Intelligence Index, it is cheaper per task than any competitor — roughly 5× cheaper than Claude Sonnet 5 (max), according to The Decoder. For coding workloads, independent benchmarks put it roughly on par with GPT-5.5 in Codex at about half the per-task cost.
The Cursor Training Angle — Why This Model Codes Differently
SpaceXAI says Grok 4.5 was “trained alongside Cursor” on tens of thousands of NVIDIA GB300 GPUs, with more than 400,000 reinforcement-learning tasks centered on multi-step software engineering. Cursor confirmed the collaboration on its own blog. That gives the model direct feedback from real developer coding sessions — a data advantage that is hard to replicate, and one reason its token efficiency stands out: 4.2× fewer output tokens than Opus 4.8 means an agentic coding loop that takes hours with Opus could finish in a fraction of the time.
What This Means for Developers
Grok 4.5 is not the most capable model available — Claude Fable 5 and GPT-5.5 top it on most benchmarks. SpaceXAI is competing on capability-per-dollar, not raw capability, and on that axis the value case is real: good-but-not-frontier performance at frontier speed and a fraction of the price. Developers can try it now via Grok Build, on all Cursor plans, and through the SpaceXAI API console. It follows heavy AI infrastructure spending across the industry, including Anthropic’s $35B compute financing deal. EU users must wait until mid-July.
What We Don’t Know Yet
Whether the benchmark token efficiency holds up in real production workflows remains unverified, and the doubled hallucination rate raises questions about reliability in unsupervised agentic use. Musk’s own framing was measured: “Our internal assessment is that Grok 4.5 is roughly comparable to Opus 4.7, but much faster,” he wrote in a follow-up post cited by TechCrunch.
FAQ
What is Grok 4.5 and when was it released?
Grok 4.5 is SpaceXAI’s most capable AI model, announced July 8, 2026 and made public July 9. SpaceXAI describes it as built to excel at coding, agentic tasks, and knowledge work, and it was trained alongside the code editor Cursor.
How much does Grok 4.5 cost?
Grok 4.5 costs $2 per million input tokens and $6 per million output tokens via the SpaceXAI API. That is roughly one-fifth the price of Claude Opus 4.7 ($5/$25), and independent analysis measured it at $0.31 per task — the cheapest of any frontier-adjacent model.
Is Grok 4.5 better than Claude Opus 4.8?
It depends on the metric. Grok 4.5 beats Opus 4.8 on DeepSWE 1.0 and Terminal Bench 2.1 in SpaceXAI’s published results, but loses on SWE Bench Pro and ranks below it on the independent Artificial Analysis Intelligence Index. Grok 4.5 is significantly cheaper and more token-efficient.
Where can I use Grok 4.5?
Grok 4.5 is available in Grok Build, on all Cursor plans, and through the SpaceXAI API console as of July 9, 2026. It is not yet available in the EU; SpaceXAI expects EU availability in mid-July 2026.
What is Grok 4.5’s hallucination rate?
Independent Day-1 testing measured a 54% hallucination rate, double the 25% rate of Grok 4.3. The model is more capable overall but more likely to state incorrect information confidently, which matters for unsupervised agentic deployments.
