Grok 4.5 released - GLM-5.2 shows up in xAI's own charts, 2.6 pts behind on SWE Bench Pro
Grok model activity is surfacing — worth tracking for capability changes, ecosystem impact, and availability.
xAI has released Grok 4.5, priced at $2/M input and $6/M output, with closed weights and a mid-July EU release.…
https://x.ai/news/grok-4-5
Quick numbers from the launch page:
- $2/M input, $6/M output. Closed weights. No EU until mid-July.
- SWE Bench Pro: Fable 80.4% > Opus 4.8 69.2% > Grok 4.5 64.7% > GLM-5.2 62.1% > GPT 5.5 58.6%
- Token efficiency is their headline: ~16k output tokens per SWE Bench Pro task vs Opus's 67k (4.2x fewer), served at 80 TPS
- Trained on "tens of thousands of GB300s", alongside Cursor
Two things worth noting for this sub:
- An MIT-licensed model you can self-host is 2.6 points behind their brand-new frontier model on SWE Bench Pro - and ahead of GPT 5.5. That's their own marketing chart.
- Harness fine print: Grok scores 62% on DeepSWE 1.0 "within each provider's harness" but 53% on DeepSWE 1.1 run independently by DataCurve. Fable went up on the independent run (66-> 70). Draw your own conclusions.
Numbers from the linked page