Sonoma Sky & Dusk Alpha: 2M Context, Real-World Cline Data
Sonoma Sky and Dusk Alpha models launched with 2M context windows. Real Cline data shows 84-87% success rates — behind Claude 4 Sonnet (96%). Fast but less reliable for production work.
TL;DR
- Sonoma Sky and Dusk Alpha models launched with 2M token context windows and free access via Vercel AI Gateway and OpenRouter
- Real Cline data shows 84-87% success rates on diff edits — behind Claude 4 Sonnet (96%), GPT-5 (92%), and Gemini 2.5 Pro (90%)
- Fast inference but reliability gaps make them better for experimentation than production coding work
What Dropped
Two new AI models appeared this week on Vercel AI Gateway and OpenRouter: Sonoma Sky Alpha (positioned for reasoning) and Sonoma Dusk Alpha (optimized for speed). Both offer 2 million token context windows and free alpha access. Their mysterious origins and impressive specs generated immediate buzz in the AI community.
The Dev Angle
We analyzed thousands of diff edit operations in Cline over the past two weeks to measure real-world performance. The results are sobering: Sky achieved 84% success rate and Dusk 87%, trailing established models significantly. Claude 4 Sonnet leads at 96%, followed by GPT-5 (92%) and Gemini 2.5 Pro (90%).
The 2 million token context window is genuinely impressive on paper — useful for large codebases and complex multi-file edits. But in practice, accuracy matters more than context size for production work. Community reports from our Discord highlight the tradeoff: users appreciate the speed, but some encounter hallucinations and tool-calling failures that established models handle reliably.
Should You Care?
If you're experimenting with new models or working on non-critical tasks, Sonoma models are worth a quick test — they're free and fast. If you're shipping production code, stick with Claude or GPT-5 as your primary choice. The performance gap is real enough that reliability should win.
The bigger picture: these models show where the industry is heading (massive context, free access, rapid iteration), but they're not ready to replace your current workflow. Use them to explore what's possible, not to replace what works.
Both models are available now through Vercel AI Gateway and OpenRouter.
Source: Cline
