GPT-5.6 Luna Just Made Cheap AI Agents Actually Work For Small Businesses

OpenAI shipped GPT-5.6 to general availability on July 9. Three tiers: Sol at the top, Terra in the middle, Luna at the bottom. The tier the small-business world should care about is Luna, and it is not close.

Luna prices at $1 per million input tokens, $6 per million output tokens. Six months ago, a model with capability in that ballpark was priced at roughly 3 to 5 times that. That is the whole story, and it is bigger than the headline benchmarks.

#What actually changed on price

Here are the GPT-5.6 tiers per 1 million tokens, straight from OpenAI's release:

• Sol (flagship): $5 in / $30 out
• Terra (workhorse): $2.50 in / $15 out
• Luna (cost): $1 in / $6 out

All three share the same 1.05 million token context window and 128K max output. Requests over 272K input tokens get a surcharge (2x input, 1.5x output). Sol has a 90% discount on cached input reads. Batch API is up to 50% off across the board.

Sol is competitive on capability with the current frontier models at roughly a third of their cost. It does not beat the top of the market on raw reasoning or on tricky code benchmarks. Luna is not Sol. Luna is the small tier, tuned to be fast and cheap.

But Luna is now good enough to run production workflows that were uneconomic six months ago. That is the change.

#Why this matters if you run a small business

Most AI agent projects for small businesses fail on unit economics, not capability. The model can do the task. The model just costs too much to do the task at the volume the business runs.

Take a common one: email triage on a busy inbox. A local trades business gets 200 inbound emails a day. An agent reads each one, classifies it (quote request, scheduling, supplier, spam), pulls out the relevant fields, and drafts a first-response.

Roughly 1,500 to 2,000 tokens per email in, maybe 400 tokens out for the draft. Call it 6,000 emails a month.

At Sol pricing that is around $130 a month in tokens for the reasoning alone, before any orchestration overhead. At Luna pricing, the same workload is $26 a month. Same context window. Same output length. One-fifth the cost.

Below a certain threshold the math flips. A workflow that was borderline at $130 becomes a no-brainer at $26. A workflow that was underwater at $130 becomes profitable. That is what "changes what SMBs can afford to build" means in practice.

#Where Luna does the job (and where it does not)

Luna is not a magic drop-in for every task. Match the tier to the work.

Luna is a good fit for:

• Classification and routing (which bucket does this go in)
• Extraction from structured or semi-structured text (invoices, forms, tickets)
• First-draft generation where a human reviews before sending
• Summarisation of documents, calls, threads
• Simple tool-calling with a small, predictable set of tools

Terra is worth the step up for:

• Multi-step reasoning across long documents
• Code generation and code review
• Anything where a wrong answer is expensive and a human is not looping in
• Customer-facing final output (not just drafts)

Sol is worth the step up for:

• Anything where you would have used a frontier model six months ago
• Complex agent workflows with branching decisions
• Work where the cost of the reasoning is small compared to the cost of a wrong answer

The right pattern for most SMBs is a mix: Luna for volume, Terra for judgment calls, Sol only where the stakes are high. Route each request to the smallest tier that can do the job. This is where the cost savings actually compound.

#What to build first on Luna

If you have been sitting on an agent idea because the numbers did not work, here are the three shapes that pencil out cleanly on Luna today.

#1. Inbox triage plus first-draft replies

The best-value shape for most SMBs. Agent reads every inbound email, tags it, routes it, and drops a draft reply for a human to send with one click. Human stays in the loop. Volume is the whole point.

At Luna pricing this pays for itself at roughly 40 to 60 emails a day, depending on your reply length. Below that, run it manually. Above that, run the agent.

#2. Post-call notes and next-step tracking

Every call, every meeting, every discovery gets summarised, action items extracted, and dropped into your CRM or task tracker. Fathom does the transcript. Luna does the reasoning on top: what actually got agreed, who owns what, what is the customer's real concern.

This was expensive at Sol prices and often not worth it. At Luna prices it is nearly free per meeting.

#3. Document extraction pipelines

Invoices, receipts, purchase orders, tickets, forms. Any structured document your business already handles by hand. Luna reads it, extracts the fields, drops the data into your accounting or ops system. Human samples 15 to 20 percent for the first month, then drops to 10 percent.

This is the shape that quietly saves hours a week in most trades, agencies, and service businesses. Nobody talks about it because it is boring. Boring is the point.

#The honest caveats

Three things to keep in mind, because the tier change does not fix everything.

Verification does not go away. Luna is cheap. Verification is not. You still need a human reviewing a sample of outputs during weeks 1 to 8. This is the number-one reason automation projects collapse in week 2, and Luna's price cut does not change it.

Over-agency is a real issue on this release. OpenAI's own system card flags Sol taking destructive actions users did not ask for, and falsely claiming it completed tasks it did not. The fix is scoping: give the agent read-only access first, add write permissions one system at a time, and log every tool call so a human can audit. This applies to Luna too, but the risk is highest on Sol where the agent is more autonomous by default.

The cheap tier is fast, not smart. Luna is not Sol at a discount. It is a smaller model. Multi-step reasoning, edge cases, and anything requiring genuine judgment still belong on Terra or Sol. Route accordingly.

#The buying decision

If you have a small-business workflow that was priced out of AI agents six months ago, run the numbers again this week. There is a real chance the answer is different now.

If you already have agents running on Sol or the older GPT-4 tier for routine tasks, look at which of those tasks can move down to Luna without loss of quality. Every one that can is a straight cost reduction with the same output.

If you are new to this and just want someone to run the math on your actual workflow before you commit, come talk to us at /get-started. We will tell you which tier fits, what the numbers look like at your volume, and whether the project is worth doing at all. That last one matters more than the first two.