GPT-5.6 Luna Guide: Specs, Pricing & When to Use

OpenAI shipped the GPT-5.6 family as three named tiers — Sol (flagship), Terra (balanced), and Luna (fast / cheap) — then cut Luna’s API price 80% on July 30, 2026. That reset every “cheapest LLM” ranking overnight. This guide is the durable reference for operators who need the model card, the gotchas, and a clear when-to-use rule — not another news rewrite.

If you only need the price-cut story, start with OpenAI cut GPT-5.6 Luna 80%. This page is the pillar the rest of the cluster links into.

What is GPT-5.6 Luna?

Luna is the cost-sensitive member of OpenAI’s GPT-5.6 family — roughly the “nano” lane for this generation. Official positioning: high-volume, latency-sensitive, and budget-constrained workloads. Per OpenAI’s model card, it accepts text and image input, returns text, supports reasoning effort levels, and exposes the GPT-5.6 tool surface on the Responses API (including programmatic tool calling).

Family context (GA from July 9, 2026; price cut July 30):

• Sol (gpt-5.6-sol) — flagship quality and cost
• Terra (gpt-5.6-terra) — everyday balance
• Luna (gpt-5.6-luna) — fastest / cheapest GPT-5.6

Model ID and the alias gotcha

Use the explicit id gpt-5.6-luna. The short alias gpt-5.6 routes to Sol, not Luna. Teams who “upgrade to 5.6” without pinning Luna can quietly pay flagship rates — a sticky, expensive mistake. We cover the bill-shock angle in depth in a dedicated alias note; for this guide, the rule is simple: pin the slug you mean.

On CloudAxis, you never type a model id in a dropdown. Auto Mode is fixed to GPT-5.6 Luna so the alias gotcha shows up for API builders, not for operators using the AI cloud computer.

Pricing (official list, July 30, 2026)

Confirm on OpenAI pricing and the Luna model card before you lock a budget. Current official list:

• Input: $0.20 / 1M tokens
• Cached input: $0.02 / 1M
• Output: $1.20 / 1M
• Previous list (pre–Jul 30): $1.00 / $6.00 — the cut is −80%
• Long prompts: OpenAI prices requests with >272K input tokens at 2× input and 1.5× output for the full request

Sibling tiers for orientation: Terra is $2 / $12 (also cut 20% on Jul 30); Sol is unchanged and remains the expensive lane. Model Luna’s spend in our AI cost calculator, or browse the live pricing tracker and cheapest LLM API guide.

Specs at a glance

Numbers from OpenAI’s Luna model card (verify if you need a contract-level quote):

• Context window: ~1,050,000 tokens
• Max output: 128,000 tokens
• Knowledge cutoff: Feb 16, 2026
• Modalities: text + image in → text out
• Endpoints: Chat Completions and Responses (tools are strongest on Responses)
• Features: streaming, function calling, structured outputs; fine-tuning not supported

Tooling note: On /v1/chat/completions, function tools with GPT-5.6 family models require reasoning_effort: "none" (or use the Responses API). Hosted stacks that omit the field can hit a 400 even when thinking is “off,” because OpenAI applies a default effort server-side.

Benchmarks worth knowing (and distrusting)

Headline scores circulate widely — treat them as orientation, not a purchase order:

• Terminal-Bench 2.1: ~84.7% (OpenAI) — within ~3 points of Terra (~87.4%) and Sol (~88.8%) at a fraction of the API price after the Jul 30 cut
• SWE-Bench Pro: ~62.7% — read alongside OpenAI’s own note that a large share of SWE-bench Pro tasks are flawed
• Speed: commonly reported near ~195 tok/s in third-party benches (your latency depends on tools and region)

Independent evaluators have also flagged benchmark gaming risk during GPT-5.6 evaluation. Useful models can still have noisy leaderboards — especially when the same vendor both designs the model and markets the score.

When to use Luna

Pick Luna when token volume × tool rounds is the constraint — not when you need the absolute best single-shot answer.

• Agent loops: browse → extract → write → schedule (many short reasoning steps)
• High-frequency monitors, tagging, classification, and first-pass drafts
• Cost-sensitive Free / starter products where a current-gen OpenAI model must fit a hard credit cap
• Workloads that benefit from ~1M context for stuffing documents — with the recall caveat below

When not to use Luna

Skip Luna (or escalate) when quality or long-range recall is load-bearing:

• Hard reasoning / flagship coding: prefer Sol (or another flagship) when a wrong answer is expensive
• Long-context fidelity: independent testing reports Luna’s recall degrading substantially well before the full 1M window — often starting to summarize instead of cite around ~300k tokens. Sol and Terra hold up better across the window
• Chat Completions + tools without Responses: if your stack cannot set reasoning_effort: none or move to Responses, fix the stack before blaming the model
• You typed gpt-5.6 by accident: that is Sol pricing — not a “Luna upgrade”

ChatGPT access vs API

On ChatGPT Work and Codex surfaces, Free and Go users get Terra; selecting Luna requires Plus or above (per OpenAI’s July 30 price-performance post). The API has no Plus gate — any API account can call gpt-5.6-luna and pay list rates.

That split creates the consumer search intent (“Luna without Plus”) even though Free ChatGPT users may already see Terra. CloudAxis does not expose a Luna dropdown: Auto Mode on every plan is powered by GPT-5.6 Luna, including Free.

How CloudAxis runs Luna

Proof, not a hypothetical: production Auto Mode resolves to OpenAI gpt-5.6-luna (hosted as cloudybot/openai/gpt-5.6-luna) on Free, Growth, Pro, and Max. Plans differ by AI credits, browser minutes, and team capacity — not by a model catalog. See why we removed the picker.

That matters for operators who want a current-generation OpenAI model inside a persistent AI cloud computer — browser, files, schedules — without wiring keys or babysitting a dropdown. Start at app.cloudaxis.ai or compare capacity on pricing.

Related reading in this cluster

This guide is the hub. Sibling posts deepen one angle each:

• gpt-5.6 → Sol alias gotcha — bill-shock post
• Cost calculator · Cheapest LLM API

Frequently asked questions

What is the GPT-5.6 Luna model ID?

Use gpt-5.6-luna. The short alias gpt-5.6 routes to Sol (flagship), not Luna.

How much does GPT-5.6 Luna cost on the API?

As of July 30, 2026: $0.20 per 1M input tokens, $1.20 per 1M output tokens, and $0.02 per 1M cached input — an 80% cut from the previous $1/$6 list.

Does Luna really have a 1M context window?

OpenAI lists about 1.05M context and 128k max output. Independent testing reports usable recall can degrade well before that — often around ~300k tokens — so treat 1M as a capacity limit, not a guarantee of perfect recall.

When should I choose Luna over Terra or Sol?

Choose Luna for high-volume agent loops and cost-sensitive work. Choose Terra for a stronger everyday balance, and Sol when flagship quality matters more than cost. On Terminal-Bench 2.1, Luna (~84.7%) sits close to Terra (~87.4%) at roughly one-tenth Terra’s API price after the July 30 cut — still verify on your own tasks.

Can I use Luna on CloudAxis without ChatGPT Plus?

Yes. CloudAxis Auto Mode is powered by GPT-5.6 Luna on every plan including Free. There is no manual model picker; you get Luna without a ChatGPT Plus subscription.

Does Luna support tools and function calling?

Yes. Tools are supported; on Chat Completions with function tools you typically need reasoning_effort set to none, or use the Responses API where the GPT-5.6 tool surface is designed to run.