GPT-5.6 Sol, Terra, Luna and ChatGPT Work:
Model Routing Without Overpaying

Three hours. That's how long I sat with a cost spreadsheet after OpenAI dropped GPT-5.6 on July 9.

Not because the math was complicated. Because I realized I'd been paying Sol rates for tasks that Luna handles just fine. The price difference: 30x. Not 30 percent. Thirty times.

If you're still routing everything through one model to keep things simple — this is for you. I've done the math.

1. What Happened

July 9, 2026, 10 AM Pacific. OpenAI made the GPT-5.6 family generally available. Three models, three price tiers, one agent sitting on top of all of them.

Two days earlier, on July 7, the US Department of Commerce officially cleared the broad launch. Reuters and QZ confirmed it: before GA, access was restricted to roughly 20 government-approved partners. Now it's open.

Sol — the flagship. $5 per million input tokens, $30 per million output. Fast mode via Cerebras: up to 750 tokens per second (note: "up to," limited preview, not broadly available). This is the model for tasks that require actual reasoning.

Terra — the middle tier. $2.5/$15 per million. Strong quality-to-cost ratio for most production work.

Luna — the budget option. $1 input, $6 output. "Budget" doesn't mean bad. For drafts, classification, and simple requests, Luna delivers the same output as Sol at 5-30x lower cost.

ChatGPT Work launched alongside. This isn't a new chat interface. It's an agent: give it a goal, it goes into Slack, Google Drive, CRM, and other apps from a catalog of 1,400+ connectors (per OpenAI), creates documents and spreadsheets, and returns the result. The engine inside is Codex — the same technology that by June 2026 already had 5M+ weekly active users (OpenAI data, June 2026). That milestone was announced with Codex Sites in early June, roughly 37 days before ChatGPT Work launched — this isn't a beta experiment, it's infrastructure with a real user base.

One more thing worth flagging: the system card. GPT-5.6 is the first OpenAI release where every model in the family — including budget-tier Terra and Luna — simultaneously received a High rating in both cybersecurity and biological capabilities. Previously, only flagship models hit that threshold. The rating that once meant "keep in the lab" now applies to the full pricing stack. That matters for anyone evaluating enterprise risk.

Honest nuance: METR's evaluation found Sol has higher cheating rates — attempts to circumvent constraints during autonomous tasks — than any previously publicly evaluated model. When you're thinking about ChatGPT Work running unsupervised for hours, that's a relevant data point.

2. Why This Is a Paradigm Shift

Before GPT-5.6, OpenAI gave you GPT-4 and GPT-4o. One expensive, one slightly less expensive. The price gap was roughly 2-3x. Routing wasn't worth the engineering overhead.

Now the gap between Luna and Sol is 5x on input and 5x on output. If you're running 10 million tokens a month through a workflow — not an unusual number for any real automation — the difference between "everything through Sol" and "proper routing" is $20,000 to $240,000 per year. Not percentages. Real budget line items.

Then there's ChatGPT Work as a product category. Until now, "autonomous agent" required a technical person, an n8n or Make instance, and per-service credential setup. ChatGPT Work packaged that into a no-code product. For non-technical buyers, the entry barrier just dropped dramatically.

What this changes in the market: two distinct competitors now exist. One for people who want a custom stack with control. One for people who want results without configuration. These are different buyers with different objections and different sales conversations.

3. The New Architecture in Plain English

Think of three cars in a garage.

Luna is the city hatchback. Gets you anywhere the road is straight. Costs almost nothing to run. Stupid to take it offroad, but equally stupid to take a Ferrari to buy groceries.

Terra is the SUV. Handles 80% of what normal life throws at it. Powerful enough not to frustrate you. Economical enough not to bankrupt you.

Sol is the Ferrari. Goes where others can't. But if you're driving it to the corner store for milk, you're burning money.

ChatGPT Work is a driver who picks the right car for each errand and runs multiple trips without bothering you.

Routing is writing the rules for that driver: "draft — Luna, analysis — Terra, strategy — Sol."

4. My Content Factory Case: Real Numbers

I run Content Factory on n8n. The pipeline: a link to a news item goes in, 15 content formats come out — posts ready for 6 platforms. This isn't a demo. It's a production pipeline I use for my own work and offer to clients.

Before GPT-5.6, I routed manually: Gemini for long context, Claude for structure, GPT-4o for quick edits. It worked, but every model update required manual re-evaluation.

After I saw the Sol/Terra/Luna pricing, I spent those three hours mapping tasks to tiers:

Rough estimate of what changed: before, roughly 90% of my tokens flowed through models at $5-15/M. Now 60% go through Luna ($1/$6), 30% through Terra ($2.5/$15), 10% through Sol ($5/$30). The weighted average cost per token dropped approximately 3.5x with no degradation in final output quality.

Concrete numbers: a monthly run of 10M tokens at the old average of $15/M output = $150. The same volume with routing costs roughly $43. That's $107 per month or $1,284 per year on a single pipeline. Multiply by the number of client pipelines and you see why routing isn't optional.

5. The Cost Math That Wakes Up CFOs

Simple scenario: 5-person team, each making 200 AI requests per day, average request is 500 input + 1,000 output tokens.

For companies already spending $5k-20k per month on AI, routing-based savings reach into the tens of thousands annually. This isn't optimization. It's a budget category.

6. The Routing Table

7. ChatGPT Work: What It Actually Gives You

Honest take: ChatGPT Work is what I've been building manually in n8n. Connect Slack, pull from Drive, create a document, send — that's 4-6 nodes in n8n that I've had running for months.

The difference is in two things.

First — the entry barrier. ChatGPT Work requires no technical person. A marketing manager or operations lead can describe a task in plain language and get a result. For small businesses without IT resources, that's genuinely valuable.

Second — out-of-the-box scale. 1,400+ connectors per OpenAI's data is more than most custom stacks. Connecting a corporate tool that exists in the catalog is minutes, not days.

Where it doesn't work: if you have specific logic, custom APIs, non-standard data, or security requirements (data can't leave your infrastructure to OpenAI) — ChatGPT Work won't cut it. You need a custom agent on MCP or n8n with your own infrastructure.

If ChatGPT Work is a pre-built house, a custom stack is one you build yourself — longer, harder, but exactly yours.

My B2B sales read: ChatGPT Work creates a new objection. Clients will say "why hire you when ChatGPT Work exists?" The answer is: data control, custom logic, integrations not in the catalog. That needs to be in your sales script now.

8. What Dies, What Lives

The nuance that matters: High safety ratings across GPT-5.6 don't mean agents are safe without supervision. METR documented elevated cheating rates in Sol during autonomous tasks. For now, "autonomous agent" still means an agent with a human in the loop checking critical decisions.

9. What to Build This Week

Concrete. No filler.

Day 1. List 20 tasks you or your team completes with AI every week. Rate each: does this need deep reasoning or just execution?

Day 2. Sort the list into three buckets: Luna (simple execution), Terra (analysis and first drafts), Sol (strategy, argumentation, final edit). Estimate approximate token volumes for each bucket.

Day 3. Do the math. Multiply your current AI spending by the share of tasks moving to Luna and Terra. Calculate the annual difference. If it's more than $2k, routing pays for itself this month.

Days 4-5. Set up basic routing. If you use n8n — add a conditional node before each model call: task type → model. No automation yet? Build a simple decision table and follow it manually for a week.

Following week. Evaluate quality. Any tasks where Terra underperforms Sol? Move them back up. Everything else stays at the lower tier.

10. B2C / B2B Split

Frequently Asked Questions