July 9, 2026 — Three major AI releases landed in the last 24 hours, and one of them is redefining the cost-performance frontier.
SpaceXAI officially launched Grok 4.5 to the public today after a week of beta testing. Elon Musk's claim? It's an "Opus-class" model — but faster, cheaper, and built entirely with Cursor.
Here's what the benchmarks show:
- $2/M input tokens — less than half the cost of comparable frontier models
- Half the tokens per task vs. GPT-4.8 and Claude Opus 4.6
- Higher tokens per second throughput
- Outperforms OpenAI 4.8 on software engineering benchmarks
- Advanced legal & finance reasoning — a first for the Grok lineage
This is SpaceXAI's first public model since going public, and the Cursor-native training pipeline is a notable differentiator. The model was trained and aligned using agentic coding workflows rather than traditional supervised fine-tuning, which SpaceXAI says is why it's so token-efficient.
Also dropping this week: NVIDIA nvDock & Microsoft HARC
NVIDIA nvDock (July 8) — a major new model release from NVIDIA, available via their NIM API. While details are still emerging, it signals NVIDIA's continued push beyond infrastructure into full-stack AI model offerings.
Microsoft HARC-Qwen2.5-7B-Instruct — Microsoft published safety-alignment adapters (LoRA) for Qwen2.5-7B, backed by a July 2026 arXiv paper. This gives teams reproducible artifacts for testing refusal robustness and safety alignment on open-weight models.
The big picture
Grok 4.5's pricing ($2/M in) is a shot across the bow at OpenAI and Anthropic. At half the cost and half the token usage per task, the effective cost per completed task is roughly 75% cheaper than competing models. If the benchmarks hold up in production, this could trigger another pricing war — just days after the industry saw API prices collapse across the board.
The combination of Grok 4.5's public launch, NVIDIA's nvDock, and Microsoft's safety alignment work makes this one of the busiest 48-hour windows of July. And it's only the 9th.
What are you testing first? Grok 4.5's coding chops, or diving into the safety alignment research? Drop your thoughts below.
Top comments (0)