llms.li

Small Systems Playbook

For founders, indie hackers, and teams under 20 engineers.

Updated November 3, 2026 shortlist: start with Claude Opus 5.5 or Sonnet 5.5 for agentic tasks, or GPT-6.1 Sol for fast quality. Route high volume to Claude Haiku 5.5, Gemini 3.8 Flash, GPT-6 Luna, DeepSeek V4.1-Flash 1017 or MiniMax M3.1 Flash. Consider Qwen3.8-27B (dense, native vision), Ministral 3 8B or Gemma 4 31B for self-hosted fallback layers. Kimi K3.1 Pro and Mercury 3 are worth a one-day eval, not a migration yet.

Architecture That Actually Ships

  • Use one primary model (GPT-6.1 Sol, Claude Opus 5.5, or Grok 4.7) and one budget model (Claude Haiku 5.5, Gemini 3.8 Flash, GPT-6 Luna, or DeepSeek V4.1-Flash 1017).
  • Keep prompts in versioned files with A/B test metadata.
  • Add retry + timeout + fallback at API boundaries; consider circuit breakers.
  • Use token-counting middleware to prevent surprise bills (especially with extended context models).
  • Cache deterministic prompts; query results aggressively to reduce context passing.

Where Small Teams Fail (2026 Edition)

  • Neglecting to recalibrate evals after pricing cuts and speedups.
  • Over-engineering model routers before product-market fit.
  • Ignoring eval datasets and trusting quick demo prompts.
  • No token budget guardrails; bleeding money on extended-context models unnecessarily.
  • Shipping without refusal/failure UX; users hit model limits and think your product broke.
  • Skipping A/B tests; new faster models tempt you to switch globally instead of validating.

Best Early Recommendation

Start with a premium model for customer-facing outputs and only optimize cost after usage patterns stabilize for 2-4 weeks.