Grok 4.5 is so cheap compared to Fable 5 and GPT 5.5 that benchmark gaps may not matter much
Quick Answer
xAI's Grok 4.5, trained on Nvidia GB300 GPUs, offers competitive performance at a low cost of $2 per million input tokens, nearly matching Fable 5 and GPT 5.5 on some benchmarks while trailing in others.
Quick Take
xAI's Grok 4.5, trained on Nvidia GB300 GPUs, offers competitive performance at a low cost of $2 per million input tokens, nearly matching Fable 5 and GPT 5.5 on some benchmarks while trailing in others. Despite scoring 83.3% on 2.1, it falls behind on DeepSWE 1.1 with 53%, highlighting a mixed performance landscape.
Key Points
- Grok 4.5 scores 83.3% on Terminal Bench 2.1, close to GPT 5.5's 83.4%.
- On DeepSWE 1.1, Grok 4.5 scores 53%, lagging behind GPT 5.5's 67%.
- Grok 4.5 costs $2 input and $6 output per million tokens, undercutting competitors.
- The model uses 4.2 times fewer tokens than Opus 4.8 on Pro tasks.
- Grok 4.5 is currently unavailable in the EU, with a planned mid-July launch.
📖 Reader Mode
~2 min readxAI has released Grok 4.5. The model was trained on tens of thousands of Nvidia GB300 GPUs and targets coding, agentic tasks, and knowledge work.
Benchmark results paint a mixed picture. On Terminal Bench 2.1, which tests complex command-line tasks, Grok 4.5 scores 83.3%, nearly matching GPT 5.5 (83.4%) and trailing Anthropic's Fable 5 (84.3%) by just one point.
But the gaps widen elsewhere. On DeepSWE 1.1, which measures the ability to resolve real GitHub issues, Grok 4.5 hits 53%, well behind OpenAI's GPT-5.5 at 67% and Fable 5 at 70%. On SWE Bench Pro, a curated set of harder software engineering problems, it scores 64.7%, beating Opus 4.8 (69.2% with max settings) in some configurations but falling short of Fable 5's 80.4%.
| Model | DeepSWE 1.1 | Terminal Bench 2.1 | SWE Bench Pro |
|---|---|---|---|
| Fable max | 70% | 84.3% | 80.4% |
| GPT 5.5 xhigh | 67% | 83.4% | 58.6% |
| Opus 4.8 max | 59% | 78.9% | 69.2% |
| Grok 4.5 | 53% | 83.3% | 64.7% |
| GLM 5.2 | 44% | 81.0% | 62.1% |
xAI says it relied on heavy data filtering, deduplication, and domain-specific selection during training to keep data quality high. The reinforcement learning stage covered hundreds of thousands of tasks, mostly from software engineering, with automated scoring. xAI built the training infrastructure for asynchronous learning, so agentic runs could stretch over many hours while training continued in parallel.
Grok 4.5 undercuts the competition on price
Grok 4.5 costs $2 per million input tokens and $6 per million output tokens. That's already far below the competition. Opus 4.8 runs $5 input and $25 output per million tokens. Fable 5 charges $10 input and $50 output per million tokens. GPT-5.5 and GPT-5.6 sit at $5 input and $30 output.
xAI also says Grok 4.5 uses 4.2 times fewer tokens than Opus 4.8 on SWE Bench Pro tasks and delivers results at 80 tokens per second. Lower per-token pricing and fewer tokens per task make Grok 4.5 by far the cheapest option in this performance tier, assuming the performance and efficiency gains hold up in practice.
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Grok 4.5 | $2 | $6 |
| Opus 4.8 | $5 | $25 |
| GPT-5.5 / GPT-5.6 | $5 | $30 |
| Fable 5 | $10 | $50 |
The pricing strategy echoes what Chinese vendors like Zhipu and DeepSeek have been doing: get close enough on performance, then win on price.
Grok 4.5 is available now through Grok Build, Cursor, and the xAI console. Plugins are live for Word, PowerPoint, and Excel. The model isn't available in the EU yet, with xAI targeting a mid-July launch. xAI trained Grok 4.5 alongside the code editor Cursor, which SpaceX acquired in mid-June for $60 billion in stock.
— Originally published at the-decoder.com
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from The Decoder
See more →
An AI model programmed nonstop for 19 days on a single MirrorCode task that cost $2,600 to run
Epoch AI's MirrorCode benchmark reveals Claude Opus 4.7 as the leader with a 56% solve rate, reconstructing a 16,000-line toolkit in 14 hours. Despite this, all models tested struggle with the most complex tasks, highlighting limitations in current AI capabilities. The single task consumed $2,600 over 19 days, raising questions about cost-effectiveness in AI development.
