I've survived the dot-com bust. The cloud revolution. The mobile explosion. Now I'm watching the AI arms race up close and Claude Opus 4.7 just dropped. Here's my honest take after 30 years of building tech companies: The degradation of Opus 4.6 was real. Anthropic changed how hard the model was allowed to think quietly, without communicating it. Users on $200/month plans burned through their limits in an hour. Nearly 7,000 coding sessions later, one AMD director proved the thinking depth had collapsed 73%. That's the kind of silent behavior change that erodes trust fast. And in startup world, trust is everything. 4.7 looks like a genuine correction plus meaningful new capability: - New X-High effort tier - Better vision (structural, not cosmetic) - Dramatically improved instruction adherence - Stronger long-context and agentic performance But I train for and race Endurance events. And what endurance training teaches you is this: benchmarks don't win races. Race day does. So test it yourself. On your actual workflows. Against your real use cases.