I'm an AI Engineer - Opus 4.7 is CRAZY Good

💼 Business owner or operator with a team? We build AI automation systems that cut costs and scale ops — done for you: https://ryanandmattdatascience.com/opus-47-consultant 🚀 Want to make money with AI skills? Join our free community — real projects, real client strategies, and the exact stack we use: https://ryanandmattdatascience.com/opus-47-skool Claude just dropped a new Opus model — and Ryan Matt breaks it down from a machine learning engineer's perspective. As a former ML engineer from Capital One who's built hundreds of agentic frameworks, Matt cuts through the marketing fluff and gives you a real take on what actually got better, what quietly got worse, and what it means for your Claude Code workflows right now. From agentic coding benchmarks to vision improvements, instruction following, the new "extra high" thinking level, and tokenizer changes — this is the honest breakdown you won't get anywhere else. TIMESTAMPS 00:00 – Intro: New Claude Opus model dropped 00:42 – Diving into the Anthropic model paper 00:56 – Biggest improvement: software engineering 01:24 – Vision improvements & higher quality image support 02:00 – Benchmarks overview (Mythos preview take) 02:28 – Agentic coding benchmarks (SWE, verified, terminal) 03:18 – What got WORSE: agentic search 04:05 – Scaled tool use & agentic computer use 04:33 – Graduate-level reasoning (all models basically tied) 04:56 – Visual reasoning: big out-of-the-box jump 05:45 – Cost: $25 per million output tokens 06:29 – Early tester highlights (skipping the marketing fluff) 06:40 – Instruction following: finally fixed? 07:19 – Multimodal support & real-world task performance 07:38 – File system-based memory improvements 08:15 – Office tasks: GPT 5.4 deserves more credit 09:00 – Vision benchmarks deep dive 09:35 – Document reasoning: massive jump vs. other models 10:20 – Long context reasoning 10:33 – Coding benchmarks across languages 11:03 – Safety & alignment (Ryan's honest take) 11:13 – New "extra high" thinking level 11:42 – Default raised to extra high for all plans 11:51 – Better image handling in the API 12:15 – Claude Code: new /ultra review command 12:51 – Tokenizer updates & token planning changes 13:23 – Token usage across thinking levels (high vs. max) 14:23 – Outro: paper link, Skool group deep dive coming OTHER SOCIALS: 🌐 Website & Blog: https://ryanandmattdatascience.com/ Ryan’s LinkedIn: https://www.linkedin.com/in/ryan-p-nolan/ Matt’s LinkedIn: https://www.linkedin.com/in/matt-payne-ceo/ Twitter/X: https://x.com/RyanMattDS *This is an affiliate program. We receive a small portion of the final sale at no extra cost to you.