SpaceXAI launched Grok 4.5 today, calling it the company’s smartest model yet and claiming victory on the two axes that matter most to developers writing cheques: speed and cost. The model runs at 80 tokens per second, uses roughly 4x fewer tokens than competitors to complete the same tasks, and is priced at $2 per million input tokens and $6 per million output tokens. But the benchmark fine print tells a more interesting story.
🔍 THE BOTTOM LINE
Grok 4.5 is the fastest and cheapest frontier coding model you can buy right now. It is not the smartest. On the hardest engineering benchmarks — SWE Bench Pro, DeepSWE, Terminal Bench — Anthropic’s Claude Fable 5 and GPT-5.5 still beat it head-to-head. Grok wins on throughput and price; Claude wins on quality. The market is splitting.
What Grok 4.5 Actually Does Well
According to SpaceXAI’s announcement, Grok 4.5 was trained on tens of thousands of NVIDIA GB300 GPUs with a heavy emphasis on data quality — deduplication, quality scoring, domain-focused curation. The reinforcement learning pipeline covers hundreds of thousands of multi-step software engineering tasks with automated grading.
The headline numbers: 80 TPS serving speed, 4.2x fewer output tokens than Claude Opus 4.8 on SWE Bench Pro (15,954 vs 67,020), and the lowest per-reply cost of any frontier model. A build-off test by TryAI measured Grok 4.5 at 110 tokens/sec throughput with a 0.44-second first-token latency — roughly double every other frontier model in the field.
Grok 4.5 is now the default model in Grok Build (SpaceXAI’s CLI tool), available in Cursor on all plans, and accessible via the SpaceXAI API console. It can build Excel models, PowerPoint diagrams, and Word documents through Office plugins.
Where It Falls Behind
The benchmark tables in SpaceXAI’s own announcement are revealing. On SWE Bench Pro resolve rate, Fable 5 (max) leads at 80.4%, Opus 4.8 at 69.2%, and Grok 4.5 at 64.7%. On DeepSWE 1.0, Fable 5 again tops at 66.1%, GPT-5.5 at 64.31%, and Grok 4.5 at 62.0%. On DeepSWE 1.1, Grok 4.5 drops to 53%, behind Fable (70%) and GPT-5.5 (67%).
The TryAI build-off confirmed the pattern in practice. When asked to build a 3D Rubik’s Cube — the hardest stateful task — Grok 4.5 produced a blank void on its first attempt and needed a retry. Claude Opus 4.8 and Fable 5 both nailed it first try. Grok did win on a particle gravity sandbox and tied on Breakout, showing its real strength: fast, clean, common-pattern code generation rather than novel algorithmic problem-solving.
The Cursor Partnership
Grok 4.5 was trained alongside Cursor, the AI code editor. This is a meaningful shift — SpaceXAI is not just selling a model, it is co-developing with the tool that developers actually use. The training pipeline was tuned for real-world agentic workflows, not just benchmark scores.
This mirrors the broader industry trend we covered when Cognition’s SWE-1.7 reached near-frontier coding performance — coding models are increasingly differentiated by their integration with real development environments, not raw benchmark numbers.
Pricing War
The pricing gap is stark. At $2/M input and $6/M output, Grok 4.5 costs roughly half what GPT-5.5 charges and a fraction of what Claude Fable 5 costs. Combined with 4x token efficiency, the effective cost per completed task can be 8-10x cheaper than Anthropic’s top model.
For high-volume code generation — boilerplate, test scaffolding, documentation, common UI patterns — this is a genuine competitive advantage. For novel algorithmic work, hard debugging, or tasks requiring deep reasoning, Claude Fable 5 remains the better choice despite the premium.
What This Means for the Market
The frontier coding model space now has three distinct positions:
- Claude Fable 5 / Opus 4.8 — highest quality, highest cost. The model you use when correctness matters more than cost.
- GPT-5.5 — balanced middle. Fast on short answers, capable on most tasks, but flubs the hardest ones.
- Grok 4.5 — speed and price leader. Best for volume codegen where throughput matters.
SpaceXAI’s pitch is “intelligence per unit of time and cost.” That is not the same as “most intelligent.” The market is buying both, and the gap between them is the story.
Note: Grok 4.5 is not yet available in the EU. EU availability is expected mid-July.
❓ FAQ
Is Grok 4.5 better than Claude for coding? Not on the hardest tasks. Fable 5 beats Grok 4.5 on SWE Bench Pro (80.4% vs 64.7%), DeepSWE, and Terminal Bench. Grok wins on speed (80 TPS vs ~47 TPS) and cost (roughly 8-10x cheaper per completed task for high-volume work).
How does the Cursor partnership work? Grok 4.5 was co-trained with Cursor’s team and is available as the default model in Cursor on all plans. The training data and RL pipeline were tuned for real agentic coding workflows inside the editor, not just synthetic benchmarks.
Why is token efficiency a big deal? Grok 4.5 uses 4.2x fewer output tokens than Opus 4.8 to complete the same SWE Bench Pro task (15,954 vs 67,020). At per-token pricing, this compounds: you pay for fewer tokens at a lower per-token rate. The cost advantage is multiplicative.
Should NZ developers switch? If your workload is high-volume boilerplate or API integration, Grok 4.5’s speed and cost profile make it worth testing. If you are solving novel problems or need reliable first-try output on complex stateful applications, Claude Fable 5 remains the safer bet despite the cost.
🔍 THE BOTTOM LINE
Grok 4.5 is the model that finally makes the cost-quality tradeoff explicit. You can have the fastest and cheapest, or the smartest — not both. For an industry that spent 2026 pretending frontier models were all roughly equivalent, this is the release that forced the market to pick a side.
