GPT-5.6 API Pricing: Sol, Terra, and Luna Tiers, Real Cost, and How to Call Each One

GPT-5.6 is OpenAI’s July 2026 flagship family, and the important thing for developers is that it is not one price. It ships in three tiers—Sol, Terra, and Luna. On the OpenAI API, Sol costs $5 per million input tokens and $30 per million output tokens, Terra is $2.50 and $15, and Luna is $1 and $6. Each tier has a cached-input discount, all three accept image input as well as text, and the context window is about 1.05M tokens with up to 128K output.

The list price is only the beginning of the real cost. Which tier you actually need, how much you cache, how much output the model generates, and a few pricing rules most guides skip—prompts over 272K tokens, data-residency uplift, and service tiers—decide the bill far more than the headline rate. This guide breaks down each tier’s price, what it is for, how it scores, how to migrate from GPT-5.5, how to call it, and how it compares with Claude, GLM, Gemini, DeepSeek, and Grok.

Quick answer: GPT-5.6 pricing at a glance

- Three tiers (per million tokens, input / output): Sol $5 / $30, Terra $2.50 / $15, Luna $1 / $6.

- Cached input: $0.50 (Sol), $0.25 (Terra), $0.10 (Luna) per million tokens; cache writes cost 1.25x the uncached input rate.

- Positioning: Sol for the hardest reasoning and agentic work, Terra as the balanced default, Luna for high-volume, cost-sensitive tasks.

- Modality: multimodal—GPT-5.6 accepts image input alongside text. Context ~1.05M tokens, max output 128K.

- Two pricing rules to plan for: prompts over 272K input tokens are billed at 2x input and 1.5x output for the whole request, and data-residency endpoints add a 10% uplift.

- Ways to pay less: Batch is 50% off standard, caching cuts repeated input to a tenth, and routing routine work to a lower tier avoids Sol’s $30 output.

- Intelligence: Artificial Analysis places GPT-5.6 Sol near the top of its Intelligence Index (score 59).

GPT-5.6 API pricing, tier by tier

OpenAI prices GPT-5.6 per million tokens, and output is far more expensive than input on every tier. The current published rates are:

- GPT-5.6 Sol: $5.00 input, $0.50 cached input, $30.00 output (cache writes $6.25).

- GPT-5.6 Terra: $2.50 input, $0.25 cached input, $15.00 output.

- GPT-5.6 Luna: $1.00 input, $0.10 cached input, $6.00 output.

Two details matter before you plan a budget. First, the output rate dominates most real bills, especially for reasoning and agentic work where the model generates long traces, tool calls, and revisions before a final answer. Second, there is no separate "Sol Ultra" price line on the pricing page; "Sol Ultra" is a higher-reasoning configuration of Sol (a multi-agent setup), not an independent tier—budget around Sol, Terra, and Luna.

What each tier is actually for

Luna: high-volume and cost-sensitive

Luna is the economy tier at $1 input and $6 output. It is the right default for classification, extraction, routing, summarization, drafting, and the large volume of routine work where a small quality difference does not change the outcome. At six dollars per million output tokens it is cheap enough to run at scale and to sample repeatedly.

Terra: the balanced default

Terra sits in the middle at $2.50 and $15. For most application and coding work it is the sensible starting point: noticeably stronger than Luna on harder tasks, but at half the output price of Sol. Begin here, then move a task up to Sol only when Terra visibly fails it.

Sol: the hardest reasoning and agentic work

Sol is the flagship at $5 and $30, and Artificial Analysis places it near the top of its Intelligence Index. Reserve it for difficult architecture, complex debugging, long-horizon agents, and high-stakes analysis where one avoided failure is worth the premium. Using Sol for routine work is the most common way to overspend on GPT-5.6.

Coming from GPT-5.5? Which tier saves the most

GPT-5.6 Sol is priced identically to GPT-5.5 at $5 input and $30 output, so moving a GPT-5.5 workload straight to Sol changes little on the bill. The saving is in the two new lower tiers. Terra delivers roughly GPT-5.5-class quality on most tasks at half the price, so the simplest migration is to send existing GPT-5.5 traffic to Terra and only promote the tasks that regress back up to Sol.

The practical migration plan is three steps: repoint GPT-5.5 calls to Terra, run your evaluation set, and escalate only the failures to Sol while pushing clearly routine work down to Luna. Most GPT-5.5 workloads come out cheaper this way without a quality loss you can measure.

How the tiers actually score

Price only matters next to capability. On OpenAI’s own Terminal-Bench 2.1 numbers, the tiers separate about as you would expect: Sol 88.8%, Sol Ultra 91.9%, Terra 87.4%, and Luna 84.7%. For context in the same table, GPT-5.5 scores 85.6%, Claude Opus 4.8 is 78.9%, Claude Fable 5 is 83.1%, and Gemini 3.1 Pro (preview) is 70.7%. Artificial Analysis separately gives Sol an Intelligence Index of 59, near the top of its ranking.

Read these as vendor-reported and directional. OpenAI has published a limited set of evaluations for GPT-5.6 and not, at the time of writing, traditional coding and reasoning benchmarks such as SWE-bench Verified or GPQA, so use these scores to decide what to test on your own workload rather than as a final verdict.

Each tier’s cached-input read rate is one tenth of its fresh-input rate: $0.50 on Sol, $0.25 on Terra, $0.10 on Luna. Writing to the cache, however, costs 1.25x the uncached input rate (about $6.25 per million on Sol), so caching pays off only when a stable prefix is reused enough times to earn back the write premium—system instructions, a large document, repository rules, or a product taxonomy. Put stable material at the front of the prompt and the changing question at the end, and avoid inserting timestamps or random identifiers ahead of reusable context.

Caching reduces repeated input cost; it does nothing for output. Because output is the expensive side on every tier, the two levers that move a GPT-5.6 bill most are caching a stable prefix and controlling how much the model writes.

Which tier should you pick?

Do not choose a tier by reputation. Choose it by test. Build a small set of representative tasks with clear acceptance criteria, run them on Luna first, and promote only the tasks that fail to Terra, then to Sol. Measure cost per accepted task, not average token price, because a cheaper tier that needs three attempts can cost more than a stronger tier that succeeds once.

One operational habit saves real money: pin the exact tier model ID in each call rather than relying on a bare alias, so a routine request never quietly runs on the most expensive tier. A mixed workload usually resolves to Luna for routine transformation and drafting, Terra for everyday implementation and analysis, and Sol reserved for the difficult minority—most requests never need the flagship at all.

GPT-5.6 vs the field on price

GPT-5.6 does not exist in isolation. Read against the current field (per million tokens, input / output, mid-2026): Claude Opus 4.8 is $5 / $25, Claude Fable 5—Anthropic’s strongest public model—is $10 / $50, GLM 5.2 is $1.40 / $4.40 on Z.AI (third-party hosts can be lower), Grok 4.5 is $2 / $6, Google’s Gemini 3.5 Flash is about $1.50 / $9, and DeepSeek V4-Pro is about $0.44 / $0.87. Confirm each on the vendor’s own page, as these move.

The takeaways are specific. GPT-5.6 Sol matches Opus 4.8 on input but is dearer on output ($30 vs $25), so on output-heavy agentic work Opus is slightly cheaper per token while Sol leads the Intelligence Index. Terra undercuts both flagships and is the value pick within the OpenAI family. Luna competes with cheaper models on price, though Grok 4.5, GLM 5.2, and especially DeepSeek sit lower still. No single tier or vendor wins everything, which is exactly why routing by difficulty beats committing to one.

A worked monthly bill

Put numbers on it. For a mid-heavy month of 50M input and 10M output tokens, at list price before caching and tool use, the model charge works out to roughly:

- GPT-5.6 Luna: about $110.

- GPT-5.6 Terra: about $275.

- GPT-5.6 Sol: about $550.

- For comparison: GLM 5.2 about $114, Grok 4.5 about $160, Claude Opus 4.8 about $500, Claude Fable 5 about $1,000.

The same tasks routed intelligently—Luna and Terra for the bulk, Sol for the hard minority—land far closer to the $110–$275 band than to a flat $550 of Sol-for-everything. That gap, multiplied across a year, is the whole argument for tier-based routing.

Hidden costs to plan for

Several pricing rules do not appear in the headline rate and surprise teams later:

- Long prompts: a request over 272K input tokens is billed at 2x input and 1.5x output for the entire request, so large-context and big-RAG calls can more than double in cost.

- Output volume: output is up to six times the input rate, so uncapped or high-reasoning output is the top overspend risk.

- Service tier: Priority processing is roughly double the standard rate (about $10 input / $60 output on Sol); Flex and Batch are cheaper. Choose the tier deliberately with the service_tier setting.

- Data residency: regional-processing endpoints add a 10% uplift for models released on or after March 5, 2026, which includes GPT-5.6.

- Agent loops and retries: instructions, tools, and history are re-sent every turn, and a failed long request wastes the input you already paid for.

- Tier drift: leaving everything on Sol "to be safe" is the single most expensive habit.

Ways to pay less

Three levers cut a GPT-5.6 bill without changing what you build. Batch: asynchronous requests through the Batch tier are 50% off standard pricing, ideal for high-volume Luna work that is not latency-sensitive. Caching: reuse a stable prefix enough times to beat the 1.25x write premium and repeated input drops to a tenth. Routing: send routine work to Luna or Terra and escalate only the hard minority to Sol, and keep prompts under the 272K threshold when you can.

How to call GPT-5.6

GPT-5.6 is available through the OpenAI API and through OpenAI-compatible gateways, so the integration stays stable while you change the model or tier behind it. The following is a standard OpenAI-compatible pattern; confirm the current model ID before production use, and pin the exact tier rather than a bare alias.

from openai import OpenAI
client = OpenAI(base_url="https://www.orcarouter.ai/v1", api_key="YOUR_API_KEY")
response = client.chat.completions.create(
    model="openai/gpt-5.6-terra",
    messages=[{"role": "user", "content": "Refactor this module and explain the risks."}],
)
print(response.choices[0].message.content)

Switching tiers is a one-line change to the model field, which is what makes difficulty-based routing practical: send routine calls to Luna or Terra and escalate only the hard ones to Sol.

Testing GPT-5.6 through OrcaRouter

A gateway makes GPT-5.6 easier to adopt for two reasons. First, the same OpenAI-compatible client can call all three tiers and comparison models such as Claude Opus 4.8, GLM 5.2, and Grok 4.5, so you can run one evaluation set across everything without rewriting integration code and measure real cost per accepted task—tool fees, retries, and the 272K surcharge included. Second, a gateway can fail over between providers and enforce a budget cap, so a temporary capacity problem or a stray Sol call does not blow the month. Confirm the live GPT-5.6 route and current tier IDs before you build a budget on them.

Next step: Register for OrcaRouter and run your own GPT-5.6 tier comparison across one API endpoint.

FAQ

What is the difference between GPT-5.6 Sol, Terra, and Luna?

They are three tiers of the same family at different prices and capability levels. Sol is the flagship for the hardest work ($5/$30), Terra is the balanced default ($2.50/$15), and Luna is the economy tier for high-volume tasks ($1/$6). Pick the lowest one that passes your tests.

Is there a GPT-5.6 Sol Ultra pricing tier?

No. "Sol Ultra" is a higher-reasoning, multi-agent configuration of Sol, not a separate line on the pricing page. It uses more tokens and costs more per task, but you budget it as Sol at higher effort.

Why did my GPT-5.6 bill exceed a simple estimate?

Usually output tokens, the 272K long-prompt surcharge (2x input, 1.5x output), Priority service tier, data-residency uplift, or retries. Log the full request chain and measure cost per accepted task rather than per prompt.

Can I fine-tune GPT-5.6, and is there a free tier?

No on both. GPT-5.6 does not support fine-tuning, and the API has no free tier—billing starts on the first token. Free access only exists inside certain ChatGPT plans, which is a separate product from the API.

Which ChatGPT plan can use which tier?

It varies by product. In ChatGPT Work and Codex, Free and Go can use Terra while Plus and above can choose among Sol, Terra, and Luna; the highest ultra setting is limited to Pro and Enterprise. Treat the API and ChatGPT plans as separate—an API key does not inherit a subscription’s access.

The bottom line

GPT-5.6’s pricing story is the tiering. Sol at $5/$30 leads the intelligence rankings, Terra at $2.50/$15 is the value default and the natural home for former GPT-5.5 traffic, and Luna at $1/$6 handles volume. Caching cuts repeated input to a tenth (after a 1.25x write premium), Batch is half price, but output is the expensive side on every tier and long prompts over 272K carry a surcharge.

Treat the three tiers as a routing problem, not a single purchase: default low, escalate on failure, cache stable context, use Batch where latency allows, and keep a comparison model reachable through the same endpoint. That produces a far lower cost per accepted task than paying flagship rates for routine work.