GPT-6 Sol and Luna Cut API Prices in Half: What a Micro-SaaS Should Route Where
OpenAI shipped Sol and Luna on September 22 at 50% of GPT-5.6 rates. Here is the routing table for a bootstrapped product that cannot put Astra on every request.
Kirtesh··10 min read·526 words
Image: IndieFounder / Unsplash
GPT-6 Sol is $2/$10 per million tokens. Luna is $0.10/$0.50. Astra stays the flagship. Indie teams should split coding, support, and extraction instead of defaulting every call to the headline model.
OpenAI expanded the GPT-6 family on September 22, 2026 with GPT-6 Sol and GPT-6 Luna. Both sit under Astra and both are priced at half the GPT-5.6 SKUs they replace.
That is the useful news for a solo founder. Astra still wins the benchmark screenshots. Sol and Luna are what a $4k MRR product can actually run all day without turning inference into the largest line on the P&L.
The price sheet that matters
| Model | Role | Input / output per MTok |
|---|---|---|
| GPT-6 Astra | Flagship, gated or expensive | Keep for rare hard jobs |
| GPT-6 Sol | Coding and multi-step work | $2 / $10 |
| GPT-6 Luna | Extract, classify, summarize | $0.10 / $0.50 |
Sol is the mid-tier. Luna is the volume tier. Free and lower ChatGPT seats get Luna. Work and Codex seats get both. Developers call gpt-6-sol and gpt-6-luna.
NOTE
Treat list prices as a starting point. Prompt caching, tool loops, and retries decide the invoice. Log tokens per finished task, not tokens per request.
A routing table for one-person SaaS
- Luna: title generation, tag classification, support ticket triage, changelog summaries, SEO outline drafts, extracting fields from invoices.
- Sol: code review, migration plans, multi-file patches, agent loops that call two or three tools, writing tests that have to compile.
- Astra or Claude Opus 5.5: architecture reviews, security-sensitive diffs, anything you would have paid a senior contractor to stare at for two hours.
If you put Sol on every autocomplete, you will feel clever for a week and angry at the bill on day eight. Autocomplete is a Luna or local-model job.
Worked cost example
Assume a support copilot that handles 2,000 tickets a month. Each ticket averages 1,200 input tokens of history plus 400 output tokens of draft reply.
- All Sol: 2.4M in + 0.8M out ≈ $4.80 + $8.00 = $12.80
- Luna draft + Sol only on 10% escalations: Luna ≈ $0.24 + $0.40, Sol slice ≈ $1.28, total about $1.90
The product did not get worse if your rubric says Luna drafts are edited by a human. You just stopped paying coding-model rates for paragraph cleanup.
Caching and retries
Sol gets cheaper when the system prompt and schema sit in cache. Design the prompt so the static block is large and stable. Do not interpolate the date into the system prompt every request.
Retries are silent spend. Cap tool loops at five. If the agent has not produced a patch, fail to a human. Infinite loops are how a $19 hobby project becomes a $190 surprise.
Why this launch is a price war
Anthropic dropped Opus 5.5 the same day. Open-weight labs keep publishing MIT weights. US labs are answering with cheaper mid-tier SKUs, not only bigger flags. For indie hackers that is good. The bottleneck moves from "can I afford a frontier model" to "can I keep the router honest."
Ship a three-bucket router this week. Measure cache hits. Do not rebuild the homepage around Astra just because the keynote did.
Written by
Kirtesh
Founder
Kirtesh is a software engineer, indie hacker, and tech analyst writing on bootstrapped micro-SaaS, autonomous AI agents, cloud architectures, and the mechanics of building profitable software businesses.
