Three observations from this week that are interesting when viewed together.
GPT Image Generation Is Degrading
GPT’s image generation model has been showing signs of degradation recently. No matter how many constraints you set, subsequent rounds of revision always make things worse. For instance, the aspect ratio of CJK characters in generated text is nearly impossible to fix. OpenAI’s stealth quality reduction is starting to rival Google Gemini.
Claude Quietly Shrinks Rate Limits
Claude looked generous last week — “5-hour rate limit lifted, weekly quota +50% through July.” In practice, that lasted about three days. Limits are quietly tightening again.
Sonnet Being Late Is Actually Good
Based on the usual cycle, Sonnet should update 2-3 weeks after an Opus release. Sonnet 4.7 or 4.8 has been conspicuously absent.
This is actually good news. Quietly building something substantial is better. Big announcements and marketing pushes usually precede disappointment.
Three Takeaways
- Whoever sits on top starts cutting corners. Model providers make no long-term commitments. Annual subscriptions are the worst bet you can make.
- How do you know when Anthropic is overpromising? When they’re loudly promoting. The gap between marketing and actual experience is universal across all model providers.
- For Anthropic, no news is good news. Sonnet being late, the team staying quiet — that suggests they’re actually heads-down working.
This cycle repeats at every company. The leader harvests, the challenger sprints, and users get squeezed in between. The only defense is never locking into a single provider. Keep your tools swappable.