GPT-5.6 Pricing After the Cut: Luna vs Terra vs Sol, and Which Tier to Use

O​penAI's GP​T-5.​6 family — GP​T-5.​6 Luna, GP​T-5.​6 Terra, and GP​T-5.​6 Sol — shipped on July 9, 2026, and has been repriced twice since: the two cheaper tiers were cut on July 30, and the flagship followed on August 21. Then September brought two changes that reset the context around all three. On September 22, O​penAI released GPT-6 Sol and GPT-6 Luna at half the GP​T-5.​6 rates. On September 29, it told ChatGPT subscribers that the $200 Pro plan reopens the next day with usage worth half as many API dollars as the old plan bought — a trade the company frames as buying comparable work on cheaper models, with the five-hour usage window retired rather than restored. The consumer side kept moving too: on August 6 GP​T-5.​6 Luna became the default model for free and Go accounts with unlimited text chats while GP​T-5.​6 Sol was updated for Plus and Pro, on August 19 O​penAI and Replit announced that Replit's Free Mode for Core and Pro subscribers runs on GP​T-5.​6 Luna, and in late August O​penAI began rolling out a “Luna Reserve” fallback allowance for select Plus and Pro accounts. This guide lays out the current GP​T-5.​6 rate card, worked cost examples, how the tiers compare to rivals and to O​penAI's own newer tiers, and how to route between them through one endpoint.

Prices are per million tokens (input / output), taken from O​penAI's current rate card and OrcaRouter's pass-through model pages; tiered and cache rates come from those pages, and competitor figures are attributed. The GPT-5.6 Sol rate quoted here is O​penAI's promotional figure — $4 / $20, with the long-context step at $8 / $30, available at least through November 21, 2026 — and not the $5 / $30 this post carried before August 21, 2026, which is still the number on the artwork beside it. The September 22 release of GPT-6 Sol and GPT-6 Luna and their rate card are as O​penAI states them in its own changelog and pricing page. The ChatGPT Pro $200 reopening, the halved API-equivalent allowance, and the retired five-hour window are as O​penAI staff described them on September 29, 2026 — executive-stated, reported the same day by several outlets, and not independently verified by us. The ChatGPT rollout details are as O​penAI stated them on August 6, 2026, the Replit Free Mode details as O​penAI and Replit stated them on August 19, 2026, and the Luna Reserve details as O​penAI documents them in its help center and as O​penAI staff and early users reported them in late August 2026 — vendor- and community-stated, not independently verified. Prices change — verify before building.

TL;DR

GPT-5.6 Luna ($0.20 / $1.20) is the high-volume workhorse; Terra ($2 / $12) is the balanced middle; Sol ($4 / $20, a promotional rate good at least through November 21, 2026) is the flagship for the hardest reasoning and agentic coding. All three share a ~1.05M-token context and 128K max output. Prompt caching and the batch option cut costs further, while very long inputs move you to a higher tier. Since September 22, O​penAI's own GPT-6 Sol ($2 / $10) and GPT-6 Luna ($0.10 / $0.50) sit at half those rates, and on September 29 O​penAI said the ChatGPT Pro $200 plan reopens September 30 with usage worth half as many API dollars as before, the five-hour window retired instead of restored. On the consumer side, O​penAI announced August 6 that GPT-5.6 Luna is becoming the default for Free and Go ChatGPT accounts with unlimited text-only chats, and that an updated GPT-5.6 Sol is rolling out to Plus and Pro; on August 19 Replit launched Free Mode on GPT-5.6 Luna for Core and Pro subscribers; and in late August O​penAI began rolling out a “Luna Reserve” fallback for select Plus and Pro accounts. Use Luna for the easy majority, Terra when you need more, and Sol only where it earns its cost — and route by difficulty through one O​penAI-compatible endpoint to minimize spend.

Key takeaways

• Luna: $0.20 / $1.20 per 1M tokens — fast, cheap, for high-volume, latency-sensitive work.

• Terra: $2 / $12 — the balanced mid-tier for tougher tasks that don't need the flagship.

• Sol: $4 / $20 — the flagship for deep reasoning, large-scale coding, and long-horizon agents. That is OpenAI's promotional rate, down from $5 / $30 on August 21, 2026 and available at least through November 21, 2026.

• GP​T-6 Sol and GP​T-6 Luna (released Sept 22, O​penAI-stated): $2 / $10 and $0.10 / $0.50 — Sol is exactly half the GP​T-5.​6 flagship on both sides, and Luna halves the input rate and cuts output further, so the GP​T-5.​6 tiers are no longer O​penAI's cheapest.

• Caching and batch cut costs further; long inputs move you to a pricier long-context tier.

• ChatGPT (O​penAI-stated, Aug 6): Luna becomes the free/Go default with unlimited text chats; Sol gets a reliability update for Plus/Pro.

• Replit (announced Aug 19): Free Mode for Core/Pro runs on GPT-5.6 Luna — up to 30x more usage per Replit, with harder tasks routed up to GPT-5.6 Sol. Replit-stated, not independently verified.

• Luna Reserve (late Aug, O​penAI-documented): a capped fallback allowance that lets select ChatGPT Plus and Pro accounts keep using GPT-5.6 Luna in Codex and ChatGPT Work after regular limits are exhausted — not unlimited, and not available in Business or Enterprise workspaces.

• ChatGPT Pro $200 (exec-stated, Sept 29): the plan reopens to new subscribers on September 30 after a pause that began September 10, with usage worth about half as many API dollars as the old plan and the five-hour usage window permanently retired.

• Biggest savings lever: route by difficulty — don't pay Sol rates for work Luna can do.

The new GP​T-5.​6 base pricing

Here's the current base pricing per million input/output tokens: Luna $0.20 / $1.20 (down from $1 / $6 on July 30, 2026), Terra $2 / $12 (down from $2.50 / $15 the same day), and Sol $4 / $20 (down from $5 / $30 on August 21, 2026 — OpenAI's promotional rate, available at least through November 21, 2026). All three tiers have now been cut; the flagship's turn came three weeks after the other two. And the ladder did not stop there: on September 22, 2026, OpenAI released GPT-6 Sol and GPT-6 Luna at exactly half the GPT-5.6 rates — $2 / $10 and $0.10 / $0.50 — so the GPT-5.6 tiers are no longer OpenAI's cheapest option, and the older GPT-5.6 flagship now costs twice what the newer GPT-6 Sol does.

Tiered, cached, and batch pricing

The base rate is only the starting point. GP​T-5.​6 pricing has three modifiers that materially change your effective cost:

• Long-context tiers. Prices step up for very long inputs — the line sits at 272,000 input tokens. On OrcaRouter's pass-through pages, Luna is $0.20 / $1.20 up to that line and $0.40 / $1.80 beyond it; Terra is $2 / $12 and then $4 / $18; Sol is $4 / $20 at its base tier and $8 / $30 above the line. The tier is chosen by each request's input token count, so a few huge prompts can quietly raise your average price.

• Prompt caching. Reused context is billed at a steep discount — Luna's cache read is roughly $0.02 per million tokens (versus $0.20 fresh), with cache writes around $0.25; Sol's cache read is about $0.40 (versus $4 fresh). For agents and chat with a stable system prompt, caching is often the single biggest saving.

• Batch. Non-urgent jobs run asynchronously at a discount — ideal for bulk classification, evals, and offline generation where latency doesn't matter.

Worked examples: what it actually costs

Concrete numbers make the tiers real. Take 10 million tokens a month with a 70% input / 30% output split (a typical chat/agent mix):

• Luna: about $5 a month (roughly $4.37 with prompt caching).

• Terra: about $50 a month.

• Sol: about $88 a month (roughly $75 with prompt caching) at the promotional $4 / $20 rate.

That's an 18x spread between Luna and Sol for the same volume — which is exactly why matching each request to the cheapest capable tier matters more than any single rate. (These follow the same method as OrcaRouter's on-page cost calculator, which estimates from list price; your real numbers depend on caching and your input/output mix.)

What each tier is actually for

GPT-5.6 Luna — the volume workhorse

Luna is the fast, cost-efficient tier, tuned for high-volume, latency-sensitive workloads: chat, classification, extraction, routing, and lightweight agentic tasks, with a p50 time-to-first-token around 1.65 seconds. After the cut, at $0.20 / $1.20 it's priced to compete directly with cheap open models — the default for the easy majority of calls. It is also the tier O​penAI is steering consumer ChatGPT toward: per its August 6 announcement, GPT-5.6 Luna is becoming the default model for Free and Go accounts, with text-only chats going unlimited and a new Think button for higher reasoning on harder questions arriving the following week (file, image, and voice limits remain). O​penAI says Luna makes 62% fewer factual errors than the GPT-5.5 Instant it replaces — a vendor-reported figure, but a clear signal that the volume tier is where the company is pointing most of its traffic. It is also no longer the cheapest thing O​penAI sells: the GPT-6 Luna released on September 22 halves its input rate at $0.10 / $0.50.

On August 19, 2026, Replit made GPT-5.6 Luna the engine of its new Free Mode — the default experience for Core and Pro subscribers. Everyday chat, ideation, and simple coding tasks run on Luna without consuming the user's paid AI-credit budget; Replit says Core users can now create up to 30x more per month, including up to 30 hours of chat. When a task needs more reasoning, Replit's Agent routes the request to a larger model for that portion — its Power/Max modes or GPT-5.6 Sol — then falls back to Luna. These are Replit- and O​penAI-stated details, not independently verified — but the direction matches the price cut: an 80%-cheaper Luna is what makes an always-on, included-by-default tier viable.

Another consumer-side wrinkle arrived in late August 2026: Luna Reserve. As O​penAI documents in its help center, select ChatGPT Plus and Pro accounts can get a limited pool of GPT-5.6 Luna usage in Codex and ChatGPT Work as a fallback once their regular usage allowance is exhausted — a notice or moon indicator appears at the cap, and you can keep working on Luna while the reserve lasts. It is not an unlimited extension: the reserve is separate from the regular allowance and itself capped, it only ever serves GPT-5.6 Luna (never Terra or Sol), it is not available in Business or Enterprise workspaces, and hitting a regular limit does not guarantee access. O​penAI staff confirmed the reserve’s existence on August 27, and early user reports describe it rolling out to selected accounts — vendor- and community-stated, not independently verified. Some early community reports also note the reserve runs the browser-light side of GPT-5.6 Luna, so live web-inspection tasks may not work while on reserve.

GPT-5.6 Terra — the balanced middle

Terra sits between volume and flagship: more capable than Luna for tougher reasoning and coding, but far cheaper than Sol. At $2 / $12 it's a sensible default when Luna isn't quite enough but you don't need the flagship — mid-complexity extraction, drafting, and multi-step tasks that still run at scale. Note where it now sits in the wider catalogue: OpenAI's newer GPT-6 Luna is priced below it at $0.10 / $0.50, so Terra's position is the middle of the GPT-5.6 family rather than the middle of OpenAI's whole line.

GPT-5.6 Sol — the flagship

Sol is built for the hardest work: deep multi-step reasoning, large-scale software engineering, and long-horizon agentic workflows, staying coherent across a ~1.05M-token context and up to 128K output. At $4 / $20 base — $8 / $30 once a request passes 272K input tokens — it is cheaper than it has ever been, on a promotional rate O​penAI says is good at least through November 21, 2026. It is also no longer the most expensive model in the catalogue: GPT-6 Astra sits well above it, and the newer GPT-6 Sol sits below it at $2 / $10. Reserve it for tasks that genuinely need it, like complex multi-file coding or long agent runs. On August 6, O​penAI also updated GPT-5.6 Sol in ChatGPT for Plus and Pro users: the company says the chat version is more reliable with facts — 68% fewer factual errors, per O​penAI — and gives more focused answers, with a new slider to control how much reasoning effort it applies. The update is chat-only (the Sol behind Work and Codex is unchanged), and the API tier stays at its promotional $4 / $20.

The reasoning-token caveat

GP​T-5.​6 are reasoning models, so effective output cost can exceed a naive estimate: when reasoning is on, internal reasoning tokens are billed as output. On hard prompts with high reasoning effort, that can dominate your bill. Tune reasoning effort to the task (low or off for simple calls), cap output tokens where you can, and measure actual usage. Cheaper per-token rates help; generating fewer tokens helps more.

How each tier compares to rivals

The cut repositioned GP​T-5.​6 against the field. Per reporting, Luna's $0.20 / $1.20 now undercuts Claude Haiku 4.5 by roughly 5x on input and 4x on output, and Terra's $2 / $12 falls below Claude Sonnet 5's standard pricing (reported at $3 / $15). Against open weights, Luna sits near the floor set by Deep​Seek's V4 line and Zhi​pu's GLM-5.2 (about $1.20 / $4.10), with Qw​en and Gem​ini Flash tiers nearby. The upshot: for cost-sensitive work, Luna is now competitive with the cheapest capable models rather than a premium alternative to them — while Sol remains a genuine premium tier for capability you can't get cheaply. The sharper comparison this month is inside OpenAI's own line: GPT-6 Sol at $2 / $10 and GPT-6 Luna at $0.10 / $0.50 — both released September 22 — now sit below every GP​T-5.​6 tier, halving the entry-level Luna and halving the GP​T-5.​6 flagship.

The real savings lever: route by difficulty

The cheapest bill isn't a single tier — it's matching each request to the least expensive model that can do it. In practice: send easy, high-volume calls to Luna (or a cheap open model), step up to Terra for harder tasks, and use Sol only for the genuinely difficult minority. Combined with caching and batch, this routinely cuts costs far more than any single price change. The catch is operational: you don't want to re-integrate three O​penAI tiers plus open-model alternatives separately. And there is now a further step above Sol inside O​penAI's own catalogue, at a lower price than Sol: difficulty routing has one more rung on it than it did in July.

Access all three (and cheaper rivals) through one endpoint

This is where a vendor-neutral router helps. OrcaRouter exposes GPT-5.6 Luna, Terra, and Sol — at the same current prices, 0% markup — through one O​penAI-compatible endpoint, alongside O​penAI's own newer GPT-6 Sol and GPT-6 Luna at their half-price rates and cheaper open models like DeepSeek V4 Pro, GLM-5.2, and Qw​en. Switching tiers (or A/B testing Luna against an open model) is a config change, not a re-integration, and you can route each request to whatever is cheapest. Luna Reserve is the consumer-side version of that same idea — a built-in fallback that keeps select users on GPT-5.6 Luna after their regular allowance is spent — and the API-side equivalent is automatic failover, which OrcaRouter handles when a provider rate-limits or errors on a request. That matters more this month than last: when a vendor re-cuts a flagship's price, a pass-through router shows the new number the same day, and when a subscription's allowance is re-based, routing by difficulty is what decides whether the change costs you anything.

The post-cut Luna rate card is visible on OrcaRouter's own model page at list price — $0.20 / $1.20 per million tokens — because the router passes provider prices through at 0% markup, with an on-page cost calculator for a typical monthly bill. The same pages carry Sol at its promotional $4 / $20. There's also a free tier and an Offers page for additional savings.

FAQ

What are the GP​T-5.​6 prices after the cut?

Luna $0.20 / $1.20, Terra $2 / $12, and Sol $4 / $20 per million input/output tokens. Luna and Terra were cut on July 30, 2026; Sol followed on August 21 at a promotional rate OpenAI says is available at least through November 21, 2026. Above the 272K-input-token line, Sol is $8 / $30.

Which GP​T-5.​6 tier should I use?

Luna for high-volume, latency-sensitive work; Terra for tougher tasks that don't need the flagship; Sol for the hardest reasoning and agentic coding. Route by difficulty to minimize cost.

Is GPT-5.6 Luna free in ChatGPT?

O​penAI announced on August 6 that GPT-5.6 Luna is becoming the default model for free and Go ChatGPT accounts, with unlimited text-only chats; limits remain on file uploads, images, and voice tools, and a Think button for higher reasoning is rolling out separately. These are O​penAI-stated plans, not independently verified.

Why is Replit using GPT-5.6 Luna for Free Mode?

Replit's Free Mode, announced August 19, 2026, runs everyday chat, ideation, and simple coding on GPT-5.6 Luna by default for Core and Pro subscribers, routing harder tasks up to a larger model — its Power/Max modes or GPT-5.6 Sol — as needed. Replit credits the July 30 Luna API price cut, roughly 80%, with making the economics work. These are vendor-stated plans, not independently verified.

What is Luna Reserve?

A fallback allowance O​penAI documents for Codex and ChatGPT Work: select ChatGPT Plus and Pro accounts get a limited pool of GPT-5.6 Luna usage once their regular usage limit is exhausted, so they keep working instead of being cut off. It only ever serves GPT-5.6 Luna — never Terra or Sol — has its own cap, is separate from regular usage, and is not available in Business or Enterprise workspaces. The existence is O​penAI-stated (staff confirmed it on August 27); the rollout to selected accounts is community-reported, not independently verified.

How much does GP​T-5.​6 cost per month?

At 10M tokens/month (70% input): roughly $5 on Luna, $50 on Terra, and $88 on Sol — an 18x spread, before caching. Your actual cost depends on caching and your input/output mix.

Do all three tiers have the same context window?

Yes — roughly a 1.05M-token context and up to 128K output tokens, with vision, tools, JSON, and reasoning support.

How does prompt caching affect cost?

Caching sharply reduces repeated-context cost — Luna's cache read is around $0.02 per million tokens versus $0.20 fresh — so reuse cached prompts wherever possible. Long inputs, conversely, can move you to a pricier long-context tier.

How do the tiers compare to Anthropic or open models?

Per reporting, Luna undercuts Claude Haiku 4.5 (~5x input / 4x output) and Terra falls below Claude Sonnet 5 ($3 / $15). Luna is also near the cheap open-model floor (Deep​Seek, GLM-5.2 ~$1.20/$4.10). Inside OpenAI's own line, the GPT-6 tiers released September 22 now undercut all three GPT-5.6 tiers.

Where can I use all three tiers together?

Through OrcaRouter's single O​penAI-compatible endpoint at 0% markup, alongside O​penAI's newer GPT-6 tiers and cheaper open models for difficulty-based routing.

Is the ChatGPT Pro $200 plan available again?

O​penAI paused new sign-ups and upgrades to the $200 Pro plan on September 10, 2026, and on September 29 an O​penAI staff member said it reopens the next day. The reopened plan is stated to be worth about half as many API dollars as the old one, with the five-hour usage window retired rather than restored; existing $200 subscriptions were never affected by the pause. Executive-stated and reported the same day, not independently verified.

Bottom line

GPT-5.6 Luna ($0.20 / $1.20) and Terra ($2 / $12) remain compelling for high-volume and mid-tier work, and Sol — at a promotional $4 / $20 rather than $5 / $30 — is the premium flagship of the family, though no longer the priciest or the most expensive per token model O​penAI sells. Two September moves reset the context around all three: GPT-6 Sol and GPT-6 Luna landed on September 22 at half the GPT-5.6 rates, and on September 29 O​penAI said the ChatGPT Pro $200 plan reopens September 30 with usage worth half as many API dollars as before, on the argument that cheaper tokens mean a smaller allowance still does comparable work. Both moves point the same direction — the vendor keeps cutting the price of capability, and the buyer who routes by difficulty keeps the difference. For API builders the practical move has not changed: send easy traffic to Luna, step up only when a task needs it, lean on caching and batch, and control reasoning tokens rather than defaulting to one tier. Doing that is easiest through a single 0%-markup endpoint like OrcaRouter, where the GPT-5.6 tiers, O​penAI's newer GPT-6 tiers, and the cheaper open models you'll want to compare them against all sit behind one API.