OpenAI turned "which GPT should I use" into a three-way choice on July 9, 2026, when it previewed the GPT-5.6 series as three named tiers: Sol, Terra, and Luna. The number is the generation; the names are durable capability tiers that OpenAI says will persist across future releases. For a marketing team that is a new kind of decision. It is no longer "use the best model", it is "which tier is the right price for this job", and the honest answer is different for a weekly performance report, a quarter-end attribution deep dive, and a classifier that tags ten thousand ad creatives overnight.
This guide is the marketing router. It covers what each tier is built for in OpenAI's own words, the verified API pricing as of August 28, 2026, a cost-per-task table using realistic marketing workloads, and a decision tree that puts each recurring marketing job on the cheapest tier that will actually do it well. Where a fact could only be confirmed from secondary reporting, the text says so.
Key Takeaways
- GPT-5.6 is a family, not a model. Sol is the flagship for complex, long-running professional and agentic work; Terra is the balanced everyday model; Luna is the fastest and cheapest, meant for classification, routing, and high-volume tasks.
- All three share the same 1,050,000-token context window, 128,000-token maximum output, and the same feature list: structured outputs, function calling, file search, image input, web search, and prompt caching, with computer use as a tool. The tiers differ in reasoning depth, speed, and price, not in features.
- API list prices as of August 28, 2026: Sol $4 per million input tokens and $20 per million output; Terra $2 and $12; Luna $0.20 and $1.20. Terra at roughly half of Sol's price, with performance OpenAI positions as competitive with the previous flagship, is the story for most marketing budgets.
- OpenAI has cut these prices twice since the July launch, so a price in a slide deck goes stale fast. Cached input is a tenth of the base rate on every tier, which rewards workflows that keep brand and data context stable.
- Default to Terra for reporting, analysis, and content; reserve Sol for multi-step agentic work and long-horizon synthesis; route bulk classification and tagging to Luna. No tier fixes a model that cannot see current campaign data.
What Are GPT-5.6 Sol, Terra, and Luna?
OpenAI's model documentation describes the split plainly: the number identifies a model's generation, while Sol, Terra, and Luna identify durable capability tiers. The API model IDs are gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna; the bare alias gpt-5.6 routes to Sol. The series was previewed on July 9, 2026 for the API and Codex first, with a select group of partners, and broader access followed over the summer.
Sol is the flagship. OpenAI positions it for complex professional work and for frontier reasoning and long-horizon agentic work: the tasks where a model has to run for many steps, hold a large amount of context, and recover from its own mistakes.
Terra is the balanced everyday model. OpenAI's own framing is GPT-5.5-competitive performance at two times lower cost, aimed at everyday interactive and agentic work with production-quality output, priced for sustainable deployment at scale.
Luna is the fastest and most affordable member of the family, positioned for smaller, faster tasks: classification, routing, and high-volume work where latency and unit cost matter more than depth.
What the tiers share matters as much as what separates them. Per OpenAI's model reference, all three carry a 1,050,000-token context window (922,000 tokens of input), a 128,000-token maximum output, a February 16, 2026 knowledge cutoff, and the same feature list: streaming, structured outputs, function calling, file search, image input, web search, and prompt caching, with computer use available as a tool through the Responses API and adjustable reasoning effort from none up to max. A workflow built on one tier moves to another by changing the model ID.
The Three Tiers Compared
| Dimension | GPT-5.6 Sol | GPT-5.6 Terra | GPT-5.6 Luna |
|---|---|---|---|
| OpenAI's positioning | Flagship for complex professional work; frontier reasoning and long-horizon agentic work | Balanced everyday model; GPT-5.5-competitive performance at 2x lower cost | Fastest, most affordable; classification, routing, high-volume tasks |
| API input / output per 1M tokens | $4.00 / $20.00 | $2.00 / $12.00 | $0.20 / $1.20 |
| Cached input per 1M tokens | $0.40 | $0.20 | $0.02 |
| Context / max output | 1,050,000 / 128,000 | 1,050,000 / 128,000 | 1,050,000 / 128,000 |
| Tools | Structured outputs, function calling, file search, image input, web search, prompt caching, computer use (all tiers) | ||
| Independent index score (Artificial Analysis Intelligence Index, max reasoning) | 59 | 55 | 51 |
| Cost per task on that index (Artificial Analysis) | $1.04 | $0.55 | $0.21 |
Two readings of that table. First, the capability gap between tiers is real but narrow on general intelligence: four index points separate Sol from Terra and another four separate Terra from Luna, while the cost per task roughly halves at each step down. Second, the independent picture comes from Artificial Analysis, whose Coding Agent Index also puts Sol ahead of Terra and Luna, and whose Intelligence Index has Claude Fable 5 edging the family at 60. OpenAI's own launch framing for Sol is long-running professional and agentic work, web research, and computer use: that is the work Sol is priced for, and the cross-vendor context for the comparison at the end of this guide.
GPT-5.6 Pricing, as of August 28, 2026
Pricing is where the series has moved most. Per Artificial Analysis, the July 9 launch prices were $5 and $30 for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna, per million input and output tokens. OpenAI has since cut Terra and Luna, then Sol, and the live pricing page now lists Sol at $4 and $20, Terra at $2 and $12, and Luna at $0.20 and $1.20. Luna's cut is the dramatic one: roughly 80 percent off its launch price, which changes what "high-volume" can mean for a marketing team.
Four details on that page decide real bills:
- Long-context surcharge. Requests past roughly the 200,000-token mark bill at a higher rate: Sol $8 and $30, Terra $4 and $18, Luna $0.40 and $1.80. Stuffing a full year of campaign exports into one prompt costs double per token.
- Cached input. Repeated context bills at a tenth of the base input price on every tier: $0.40, $0.20, and $0.02 per million tokens for Sol, Terra, and Luna. Brand guidelines, taxonomy rules, and reporting templates that stay stable across calls should sit in the cached prefix.
- Batch and Fast modes. The batch API runs at 50 percent off standard rates for work that can wait; Fast mode costs twice standard on Sol, Terra, and Luna for latency-sensitive use.
- Volatility. The pricing page itself states that promotional pricing is available at least through November 21, 2026, and the tiers have been cut twice since July, so any number in this article can be superseded. Treat the pricing page as the source of truth at the moment a budget is signed off.
On the subscription side, ChatGPT plan access to the tiers is documented on OpenAI's help center, which was not reachable for verification while this guide was written; secondary reporting describes Sol on the paid plans and Luna on the free tier, with ChatGPT increasingly routing between tiers on its own. For a marketing team the API pricing is the durable number anyway, because that is where recurring workflows run.
Which Tier for Which Marketing Job: The Decision Tree
Start from the task, not the model. Three questions sort almost every recurring marketing workload.
1. Does the task run for many steps and have to recover from mistakes? Multi-step agentic work such as an agent that pulls campaign data, reconciles it against the budget plan, drafts the variance explanation, and files a ticket when something is off is Sol territory. OpenAI's Sol framing is explicitly long-horizon agentic work, and on Artificial Analysis's Coding Agent Index Sol leads the family. If the answer is no, drop down a tier.
2. Does a human read the output directly, and would a mediocre answer cost real money? Board-level narrative, a quarter-end attribution readout, a competitive deep dive that spans dozens of documents: Terra is the default, and Sol is the upgrade only when the synthesis is genuinely long-horizon. For everyday analysis, report drafting, and content work, Terra's positioning (production-quality output at roughly half the cost) is exactly the point.
3. Is the job the same small decision, made thousands of times? Tagging creatives by theme, classifying search terms by intent, mapping campaign names to a taxonomy, routing inbound leads, flagging anomalies in a daily feed: Luna. At $0.20 per million input tokens, the unit economics change the design, and OpenAI names classification and routing as Luna's job.
| Marketing job | Recommended tier | Why |
|---|---|---|
| Weekly cross-channel performance report (draft from a data pull) | Terra | Structured input, bounded output, read by a human who can spot errors; Sol adds cost, not accuracy |
| Quarter-end attribution and budget narrative across dozens of sources | Terra, Sol if multi-document | Long-horizon synthesis is where OpenAI says Sol earns its price |
| Agent that monitors pacing, investigates anomalies, and acts | Sol | Many steps, tool calls, and self-correction; OpenAI's stated Sol use case |
| Ad copy and landing-page variants | Terra | Everyday generation; the reviewer, not the model, is the quality gate |
| Creative tagging, search-term intent classification, taxonomy mapping | Luna | Thousands of small identical decisions; OpenAI's stated Luna use case |
| Lead routing and inbox triage | Luna | Latency and volume dominate; escalate the ambiguous 5 percent to Terra |
| Competitive research across the open web | Sol | Multi-hop web research is the long-horizon agentic work OpenAI positions Sol for |
What a Real Marketing Workload Costs on Each Tier
List prices only become decisions once they are multiplied by a real workload. Take a common one: a weekly performance report for 40 campaigns, drafted from an exported data table. Assume 30,000 input tokens (the data plus a stable 10,000-token instruction and template prefix that prompt caching serves at the cached rate after the first run) and 4,000 output tokens, once a week, for a year.
| Workload (per run) | Sol | Terra | Luna |
|---|---|---|---|
| 20,000 fresh input tokens | $0.080 | $0.040 | $0.004 |
| 10,000 cached prefix tokens | $0.004 | $0.002 | $0.0002 |
| 4,000 output tokens | $0.080 | $0.048 | $0.0048 |
| Per report | $0.164 | $0.090 | $0.009 |
| 52 weekly reports | $8.53 | $4.68 | $0.47 |
The arithmetic makes two points that survive any future price change. The model bill for a single reporting workflow is small on every tier; the expensive part is the analyst hour spent checking a wrong report, which is an argument for Terra over Luna on anything a human acts on. And the tiers only diverge into real money at scale: the same three prices applied to 50,000 creative-tagging calls a month put Sol at roughly eighteen times Luna's bill for a job Luna is designed to do. Match the tier to the job, then spend the savings on the data feeding it.
How the Tiers Sit Against Claude and Gemini
Cross-vendor pricing, from each company's official page on August 28, 2026: Anthropic lists Claude Fable 5 at $10 and $50 per million input and output tokens and Claude Opus 5 at $5 and $25; Google lists Gemini 3.5 Flash at $1.50 and $9 standard. On those numbers Sol undercuts both Claude tiers on list price while Artificial Analysis scores Fable 5 one index point above it, Terra sits between Gemini 3.5 Flash and Opus 5, and Luna has no direct price peer among the three. The practical reading for a marketing stack that already uses more than one vendor is that GPT-5.6 Terra is the cost-performance anchor for everyday analysis, Sol competes with Fable 5 and Opus 5 on the long-horizon work, and the choice within each band comes down to the tooling and data connections each vendor's ecosystem gives the team, which is the subject of our full Claude vs ChatGPT vs Gemini vs DeepSeek comparison and the Claude Fable 5 explainer.
What No Tier Fixes
Every tier in this guide reasons over what it is given. A weekly report drafted by Sol from a stale export is an expensive way to be confidently wrong; the same report drafted by Terra from live, reconciled campaign data is the one the team can act on. The teams getting durable value from the GPT-5.6 tiers spent their effort on the data layer first: connecting the model to ad platforms, analytics, and the CRM through the Model Context Protocol, keeping the taxonomy and metric definitions in the cached prefix, and letting the tier decision be a per-workflow cost choice rather than a leap of faith. Our guide to MCP servers for marketing data covers that layer.
Frequently Asked Questions
What is the difference between GPT-5.6 Sol, Terra, and Luna?
They are three capability tiers of the same generation. Sol is the flagship for complex professional and long-horizon agentic work; Terra is the balanced everyday model that OpenAI positions as GPT-5.5-competitive at half the cost; Luna is the fastest and cheapest, built for classification, routing, and high-volume tasks. All three share the same context window, output limit, and tool support.
How much does GPT-5.6 cost?
As of August 28, 2026, OpenAI's pricing page lists Sol at $4 per million input tokens and $20 per million output, Terra at $2 and $12, and Luna at $0.20 and $1.20, with cached input at a tenth of the base rate, a surcharge past roughly 200,000 tokens of context, batch at 50 percent off, and Fast mode at twice standard. The page states promotional pricing is available at least through November 21, 2026, and prices have been cut twice since the July launch; check the page before committing a budget.
Is Terra good enough for marketing analytics?
For reporting, everyday analysis, and content it is the sensible default: OpenAI positions it as production-quality at roughly half of Sol's price, and independent testing puts it four index points behind Sol at about half the cost per task. Move a workflow to Sol when it is genuinely multi-step and agentic or spans dozens of documents.
When should a marketing team use Luna?
Whenever the job is the same small decision made thousands of times: tagging creatives, classifying search terms, mapping campaign names to a taxonomy, routing leads. OpenAI names classification and routing as Luna's purpose, and at $0.20 per million input tokens the volume economics change what is worth automating.
Can any of the tiers read my marketing data directly?
Not on their own. The tiers support function calling, file search, and web search, but live ad-platform, analytics, and CRM data reach the model through connectors such as MCP servers. The tier decides how well the model reasons; the data connection decides whether it reasons about something true.
