Snapshot Context
Mid-February snapshot documenting Kimi K2.5 as dominant daily driver, Flow/Veo3 testing for video content, and VPS prep for OpenClaw deployment. Captures the “rate limit reality” across multiple providers. This is when the stack crystallized around practical access over theoretical capability.
The Stack
Primary Models
- Kimi K2.5: Dominant daily driver. Free via OpenCode Zen (generous tier) + API credits. Most cost-effective frontier access right now.
- Opus 4.6: Still in rotation for deep reasoning and complex architectures. See pricing comparison.
Media / Research
- Flow: Google’s generative video workspace (Veo3). Testing for content creation workflows.
- Google AI Studio: Rapid prototyping and prompt engineering before committing API spend.
- Researching the differentiation between Google Labs vs Google AI Studio vs Flow — complex branding overlap. Published the field guide.
Infrastructure Prep
VPS selection underway for OpenClaw walkthrough. Testing providers for:
- UFW + Tailscale compatibility
- Snapshot automation
- $5-10/mo price band
Goal: Disposable, isolated compute for agent experiments. Not buying dedicated hardware.
Key Changes
| Change | Impact | Notes |
|---|---|---|
| Kimi becomes dominant | High | Shift from multi-model to Kimi-primary |
| Flow/Veo3 testing starts | Medium | Video content workflow experiments |
| Rate limit reality documented | High | “Hitting walls everywhere except OpenCode” |
| VPS prep begins | Medium | OpenClaw deployment planning |
The Rate Limit Reality
Status: Hitting walls everywhere except OpenCode.
What Happened
- Google AI Studio (Gemini 3 Flash): Hit the weekly quota in ~2 days of actual work. The “100% Resets in 55 hours” message is… not helpful when you have work to do today.
- Antigravity: Still one of the best offers on paper (1M context, $20/mo), but the Claude Opus integration is effectively unusable due to rate limits. The 5-hour refresh only applies to Gemini models.
- OpenCode Zen (Kimi K2.5): Limits still generous. This is where most of the actual work is happening.
Current Workflow Reality
Google Labs tools when you need them: waiting. OpenCode: actually working.
The irony: Antigravity technically has better specs (1M context vs 256K), but rate limits make it theoretical. Kimi via OpenCode has lower limits on paper but practical access for getting work done.
Verdict: Free tier generosity beats theoretical capability when you need to ship. OpenCode + Kimi K2.5 is the actual daily driver. Google tools are for experiments when the limits reset.
Upcoming (at this time)
- Flow differentiation research: Covering the Google AI suite maze — Labs, Studio, Flow. When to use what.
- OpenClaw VPS guide: Step-by-step from fresh VM to production-hardened agent runtime.
Related
- Current stack →
- Previous snapshot: 4 Feb 2026 →
- Next snapshot: 11 Feb 2026 →
- /value/kimi-access/
- /deploy/openclaw/yolo-safely/
- /compare/claude-vs-openai/pricing/
- /compare/google-labs-vs-ai-studio-vs-flow/
This is a historical snapshot from 2026-02-09. View current stack →