Grok 4.5 Achieves Top Benchmarks, GPT 5.6 Retained, Perplexity Integrates Fast Inference

Key Takeaways

1. Grok 4.5 achieves top scores on multiple benchmarks, including SWE-Atlas-QnA, and is described as Opus-level at lower cost and faster speed.
2. GPT 5.6 Sol remains in ChatGPT subscriptions (Go, Plus, Pro) until a better model is released, ensuring continuity for users.
3. Perplexity integrates Grok 4.5 for fast, cost-effective inference, and plans to use Nvidia Vera CPU to boost coding performance by over 50%.
4. Industry voices (Aravind Srinivas, Yohei, Guillermo Rauch) stress that companies must own their AI infrastructure, learning loops, and agent sandboxes to protect IP and control.
5. Garry Tan and Paul Graham argue that AI creates net jobs and that YC’s SAFE model has transformed startup investing, while founders should focus on shipping rather than over-validation.

1. AI Model Milestones and Product Updates

1. Grok 4.5 excels in multiple benchmarks, including achieving the highest score on SWE-Atlas-QnA, and is described as Opus-level with lower cost and faster inference. Elon Musk also noted improvements in video studio evaluations from 6/33 to 23/33. — via 1 2 3 4
2. GPT 5.6 Sol will remain in all ChatGPT subscription tiers (Go, Plus, Pro) until a better model is introduced, as confirmed by Sam Altman. This provides clarity for subscribers about model continuity. — via 1 2
3. Perplexity integrated Grok 4.5 within hours, citing its top evaluation scores and zero data retention (ZDR) support. A user built a personalized podcast aggregator for just $11, demonstrating practical cost efficiency. — via 1 2

2. AI Infrastructure and Platform Strategy

1. Perplexity plans to use Nvidia's standalone Vera CPU, which showed 1.5x speedup in agent coding tasks, with expected gains exceeding 50%. Detailed benchmarks will be released soon. — via 1
2. Aravind Srinivas argues that restricting distillation while retaining user interaction data for learning is ironic. He believes learning infrastructure must be distributed to every company, allowing each to control its own learning loop, otherwise economic value concentrates with infrastructure owners. — via 1
3. Yohei emphasizes that e2b is the only major open-source agent sandbox under Apache 2.0, and argues that sandboxes are critical AI infrastructure that enterprises should own internally to protect IP. — via 1
4. Guillermo Rauch advocates treating models as a gear in one's own machine, stressing that startups and enterprises must own their data, evaluations, model selection, and software layer—not outsourcing cognition. — via 1

3. Startup AI Economics and Founder Insights

1. Garry Tan argues that AI is currently a net job creator and that the Jevons paradox will drive AI to generate far more new workloads than imagined, citing AI helping theoretical physicists solve long-standing problems. — via 1 2 3
2. Garry Tan also highlights YC’s contribution to the SAFE and founder-friendly ecosystem, noting that YC changed startup investing, even benefiting founders not admitted. — via 1
3. Paul Graham observes that YC’s investor complaints about valuations and batch size have historically been wrong, and that YC funded 208 startups by 2010. He also notes that one startup apologized for only 36% monthly growth due to fundraising focus. — via 1 2 3
4. Jason Fried argues that “validation” has limited value; building and shipping is the real test. He also notes that being an employee is often the smarter choice, and not everyone should start a company. — via 1 2 3
5. Andrew Wilkinson expresses strong impression of the dia browser from @browsercompany, calling it a far better product than OpenAI’s canceled Atlas project. — via 1