Grok 4.5 Arrives at a Third of the Price — and That Might Matter More Than Benchmarks
xAI's new model trails Fable 5 and GPT-5.5 on some coding tests, but at $2 per million input tokens it undercuts everyone in its class. EU launch expected mid-July.
xAI has released Grok 4.5, and the headline isn’t the benchmark scores — it’s the price tag. The model costs $2 per million input tokens and $6 per million output tokens (tokens are the small chunks of text AI models read and write; a million tokens is roughly 750,000 words). Compare that with Anthropic’s Fable 5 at $10 and $50, or GPT-5.5 and GPT-5.6 at $5 and $30, and you can see the strategy: don’t beat the leaders, undercut them.
On performance, the picture is mixed. On Terminal Bench 2.1, which tests complex command-line tasks, Grok 4.5 scores 83.3 percent — essentially tied with GPT-5.5 (83.4) and one point behind Fable 5 (84.3). On harder software-engineering tests the gaps widen: 53 percent on DeepSWE 1.1 versus Fable 5’s 70, and 64.7 percent on SWE Bench Pro versus Fable 5’s 80.4. xAI also claims Grok 4.5 is unusually efficient, using 4.2 times fewer tokens than Anthropic’s Opus 4.8 on comparable tasks — so each job costs less on top of the lower rates.
What’s behind this? A price war, plainly. Chinese labs like Zhipu and DeepSeek proved you can win customers by being “close enough” on quality and dramatically cheaper — a Hacker News discussion this week about GLM 5.2 and a coming “AI margin collapse” made the same argument. Now a US lab is running the same play. xAI trained the model on tens of thousands of Nvidia GB300 GPUs, alongside the code editor Cursor, which SpaceX acquired in June for $60 billion in stock. Worth keeping expectations grounded: efficiency claims come from xAI’s own testing, and independent verification takes time.
What this means for you: If you just chat with an AI assistant, nothing changes today — this is mostly an API story, and Grok 4.5 isn’t available in the EU yet (xAI is targeting mid-July). But if you build with AI or pay per token, the math is worth a look: for many everyday tasks, a model that’s 90 percent as good at 20 percent of the price is the rational choice. The bigger takeaway is the trend. Frontier-level AI is getting cheaper fast, and the labs charging premium prices now have to prove they’re worth the difference.
Sources
Source: https://x.ai/news/grok-4-5
MiniMax Reportedly Plans a 2.7 Trillion Parameter Open-Source Model
The Information reports the Chinese lab is preparing M3 Pro, which would be the largest Chinese AI model yet — and freely downloadable, possibly as early as Q3.
