Lab
Now
Lab setup last observed August 11, with September 27 model-selection links. Personal usage has not been revalidated.
September 27 editorial update: the setup below was last observed August 11; personal usage has not been revalidated. For current model decisions, use the Opus 5.5 review and GPT-6 Sol/Luna guide. The table records a setup, not a current frontier ranking.
Last observed stack — August 11
| Role | Tool | Status | Notes |
|---|---|---|---|
| Primary coding | GPT-5.6 Luna | 🟢 Active | First route for bounded routine Codex work; move to Terra when retries erase the advantage and Sol for hard or consequential work. |
| Cheap API overflow | DeepSeek V4 Flash 0731 | 🟢 Active | Direct API and MIT-weight value candidate. Keep the OpenCode free promotion separate until its backend revision is confirmed. |
| Supported-tool overflow | GLM-5.2 / Z.AI | 🟢 Active | 1M-context supported-tool coding lane. Value case is cost, context, and tool fit, not “best model” hype. |
| Premium review | Claude Code / Claude Max | 🟡 Limited | Viable again after Anthropic doubled Claude Code five-hour limits for paid tiers. Still expensive and managed by limits. |
| API overflow | Kimi K2.7 Code | 🟡 Limited | Useful for long-context or API-heavy coding when provider routing is acceptable. HighSpeed is the latency-priced option. |
| Watch list | Gemini paid apps | 🟡 Watch | Large context is useful, but compute-based usage is too volatile for the primary coding lane this week. |
| Infrastructure | OpenClaw on VPS | 🟢 Active | Disposable VM strategy. UFW + Tailscale isolation. |
Legend: 🟢 Active | 🟡 Limited/Costly | 🔴 Blocked/Deprecated
Recently Changed
- 2026-08-11: Added a safe agents-running-agents probe for authorized OpenCode-to-DeepSeek calls. Connectivity is validated; site-owned CAR remains not-run.
- 2026-08-01: Moved the bounded-work default from GPT-5.5 to GPT-5.6 Luna and added DeepSeek V4 Flash 0731 as the direct-API value candidate. Site-owned CAR remains not-run.
- 2026-06-20: Refreshed overflow notes around GLM-5.2 and Kimi K2.7 Code. GLM-5.1 and K2.6 are now prior-release context on current-facing pages.
- 2026-05-21: Switched current coding recommendation to OpenAI first, GLM overflow, Claude/Kimi second lanes, Google watch-only.
- 2026-02-11: Kimi quota weirdness—monthly became weekly, actually better for workflow
- 2026-02-09: Flow/Veo3 testing for video content; rate limits hit fast
- 2026-02-04: OpenCode Zen limits reduced; Kiro CLI removed Opus from free tier
- 2026-02-03: Kilo Code’s free Kimi option appears gone
- 2026-02-01: Rate limit reality: OpenCode still working when others are gated
Active Investigations
Active investigations stay private until promoted to a public lab note or post.
Watching/Testing
- Gemini paid app compute meter — Useful context window, but effective limits are less predictable for hard coding prompts.
- Kimi K2.7 Code provider routing — Good API overflow candidate when setup friction is acceptable.
- OpenClaw enterprise features — Waiting for skills documentation to mature
Blockers
- Google paid apps are not a primary coding lane until compute-meter behavior stabilizes.
Personal setup last observed August 11. View historical snapshots →
Related
- /value/free-stack/ — Free tier guide
- /value/smart-spend/ — Pricing strategy
- /posts/gpt-5-6-luna-price-cut/ — Historical July OpenAI value analysis
- /posts/deepseek-v4-flash-0731-value-frontier/ — Direct-API value frontier
- /posts/ai-coding-subscription-limits-may-2026/ — Historical May coding-limit analysis
- /deploy/openclaw/yolo-safely/ — Current deployment setup