Welcome back! Today’s playbook: The AI pricing market just hit a 100× spread — and most teams are still paying flagship prices for tasks that don't need it. Three steps to fix it, one trap to avoid.
Stop Defaulting to Flagship Models
GPT-5 Nano costs $0.05/M input tokens. Claude Opus 4.6 costs $5/M. Same production-grade quality bar. 100× price difference. If you're still defaulting everything to your flagship, that's not a technical decision — it's a habit that's costing you.
Haiku 4.5, GPT-5 Nano, Gemini 2.5 Flash-Lite aren't the flaky cheap models they were 18 months ago. They handle classification, extraction, summarization, and RAG reliably now. The gap that justified "use the best, always" is mostly gone.
STEP 01 — Audit before you touch anything.
Pull your last 30 days of calls. Look at what they're actually doing. Most production systems have classification, extraction, RAG, and multi-step reasoning all mixed together — and they're not the same problem. You probably don't need Opus to label intents.
Three tiers:
Fast / Cheap — GPT-5 Nano · Gemini 2.5 Flash-Lite · Haiku 4.5 Classification, extraction, structured output. Route here by default.
Balanced — GPT-5 Mini · Gemini 2.5 Flash · Sonnet 4.6 RAG Q&A, summarization, moderate context. Sonnet 4.6 at $3/M is closer to Opus than you'd expect.
Heavy — GPT-5.4 Pro · Gemini Pro · Opus 4.6 Multi-step agents, complex codegen, high-stakes tasks. Reserve it.
Most teams find 60–70% of volume fits in the first two tiers. At 1M calls/day that's roughly $5,000/day at Opus pricing down to ~$1,600/day — $3,400 back, daily, same outputs.
STEP 02 — Route up front, not after failure.
Escalating after a bad response means you pay for both calls. Use pre-call signals instead:
→ Input token count → Presence of tool calls → Multi-turn context depth
Want off-the-shelf? Microsoft's Model Router in Azure AI Foundry routes across GPT-5, Claude (Haiku 4.5, Sonnet 4.6, Opus 4.6), DeepSeek, Llama, and Grok behind one API. You pick Quality, Cost, or Balanced mode. It handles the rest.
STEP 03 — Tag every trace with model_tier
Watch tier distribution and quality delta. Without this, routing drifts — someone defaults to flagship on a new use case and six months later you're wondering why the bill went up.
WATCH OUT — Context window ≠ max output
Per Million Tokens(M)
Claude Haiku 4.5: 200K context · 64K max output · $1/$5 per M GPT-5 Nano: 400K context · 128K max output · $0.05/$0.40 per M
These are different numbers. Cheaper models truncate silently — most SDKs won't throw, they'll just answer on incomplete context. Add truncation logic on your side before the call.
What did you think of today's email?
Your feedback helps me create better emails for you! comment down 👇
Loved It 😊
It was ok 🙂
Could be better 🤔
Until next time - Teja Derangula,
The gap between thinking and building has shrunk — take advantage.