Grok 4.5 Challenges GPT-5.6: AI Model Race Heats Up with Competitive Benchmarks and Cost Efficiency

Key Takeaways

1. Grok 4.5 achieves top scores on WANDR at half the cost of Opus 4.8, and ranks high on SWE benchmarks, challenging GPT-5.6's "best model" claim.
2. Sam Altman asserts GPT-5.6 Sol is the best, backed by doctor preference data; Elon Musk counters with Grok 4.5's superior cost-performance.
3. OpenAI fixes Codex/Work issues, announces Build Week starting July 13, and reveals 30% of Sol/Fable costs come from Fable.
4. Amjad Masad reports Claude and Replit agents autonomously discussing bugs, signaling AI-to-AI debugging future.
5. @levelsio advocates "VibeOps" (editing production code directly) and warns of data loss risks when running local coding agents.
6. Aravind Srinivas predicts 12-month path to Opus-4.8-level local models, and emphasizes AI competition shifting from model size to system cost.

1. AI Model Race: Grok 4.5 vs GPT-5.6

1. Elon Musk announces Grok 4.5 achieves leading scores on WANDR at half the cost of Opus 4.8, constructs 4–12 dimension hyper-contraction counterexamples on xAI V9 (1.5T params), ranks second on APEX-SWE (Pass@1 51.2%, +30.2pp in one year), and ties with Codex GPT-5.6 on SWE-Atlas-QnA. The model is available via Grok Build CLI and Perplexity Computer at low or no cost. — via 1 2 3 4 5 6
2. Sam Altman claims GPT-5.6 Sol is currently the best model, citing Musk's renewed obsession as evidence, and notes doctors find GPT-5.6's responses have fewer defects than human-written ones. — via 1 2
3. Greg Isenberg confirms Grok 4.5 is the best model on Hermes/OpenClaw, cheaper than Opus 4.8 by over 60%, fast at ~$2.49 per task. Aravind Srinivas adds Grok 4.5 achieves top WANDR score at half Opus cost and is enabled for Perplexity Enterprise. — via 1 2 3
4. Musk mocks Altman as "taking the scam to new heights," while Altman jibes that Musk's renewed obsession confirms GPT-5.6's superiority. — via 1 2
5. Aravind predicts within 6 months a model with Fable 5 quality at 3–4x lower cost, and within 12 months a locally runnable Opus 4.8-level model (>50% probability), asserting the AI race is shifting from scaling up to systems that are cheaper and smarter. — via 1 2

2. OpenAI Product Updates and Build Week

1. Sam Altman announces fixes to ChatGPT Work and Codex: reset usage limits, correct overly high defaults, resolve desktop app reorganization inconveniences, and clarify that Codex is not going away. Larger improvements are planned next week. — via 1
2. OpenAI Build Week begins July 13, open for registration, encouraging projects built with Codex. — via 1
3. Altman reveals that when using the Sol and Fable models, 30% of costs come from Fable. — via 1

3. Developer Insights and Tools

1. Amjad Masad reports that Claude and Replit agents autonomously connected and began discussing product bugs, with users unable to keep up—signaling an AI-to-AI future. He also built a Notion-like app for his team using Replit at under $100, and advocates shifting AI from "token-maxxing" to "outcome-maxxing." — via 1 2 3
2. @levelsio introduces "VibeOps": editing production code directly via SSH and Claude Code without git, reporting only two brief outages in 12 months. He warns against running coding agents locally without backups. — via 1 2
3. Paul Graham recommends legal tool Legora, which helps lawyers handle complex cases. Garry Tan highlights Unsloth for fine-tuning Llama models, Insforge serving over 40k projects, and AI design workflows changing how we build. — via 1 2 3 4 5
4. Paul Graham tells a 14-year-old that those who persist in reading books will have a massive advantage. — via 1