# GPT-5.6 Sol, Terra and Luna: What OpenAI Launched

> OpenAI launched GPT-5.6 on June 26 as three tiers: Sol, Terra and Luna. Here are the confirmed prices, the benchmarks, and the government access catch.

Source: https://capitalandcompute.net/blog/gpt-5-6-sol-terra-luna-launch/
Published: 2026-06-27
Updated: 2026-09-23

After weeks of leaks and a public spat with the federal government over timing, the GPT-5.6 launch finally arrived on June 26, 2026, and it was not one model. OpenAI [previewed GPT-5.6 as a three-tier suite](https://openai.com/index/previewing-gpt-5-6-sol/): Sol, the flagship; Terra, a mid tier; and Luna, a budget option. The whole suite went out in a US-government-gated limited preview to roughly 20 partners, the rollout the administration had asked for.

Key takeaways

* GPT-5.6 launched June 26, 2026 as three tiers: Sol (flagship), Terra (mid) and Luna (budget), not a single model.
* Launch pricing per million tokens: Sol $5 / $30, Terra $2.50 / $15, Luna $1 / $6, with cached input 90 percent off on each tier. Terra and Luna were later cut to $2 / $12 and $0.20 / $1.20 on July 30, 2026.
* At launch Sol held GPT-5.5 pricing exactly. It fell to $4 / $20 on August 21, 2026, and GPT-6 Sol replaced the tier entirely at $2 / $10 on September 22.
* Sol leads Claude Mythos 5 on agentic coding: 88.8 percent to 88 percent on Terminal-Bench 2.1 in max mode, 91.91 percent in ultra mode.
* The catch: access is limited to about 20 government-approved partners. OpenAI called the gated access process unsustainable as a long-term default.

## The three tiers and what they cost

The headline change is the split. Where GPT-5.5 was a single flagship rate, GPT-5.6 ships as a ladder. Sol is the heavy model for long-horizon coding, security and agentic work. Terra is positioned at GPT-5.5-level performance for half the cost, aimed at high-volume production. Luna is the fast, cheap option for routine tasks.

| Tier  | Launch input / output | Current input / output | Current cached input | Positioned for                |
| ----- | --------------------- | ---------------------- | -------------------- | ----------------------------- |
| Sol   | $5.00 / $30.00        | $4.00 / $20.00 (promo) | $0.40                | Long-horizon coding, security |
| Terra | $2.50 / $15.00        | $2.00 / $12.00         | $0.20                | High-volume production work   |
| Luna  | $1.00 / $6.00         | $0.20 / $1.20          | $0.02                | Fast, routine, low-cost tasks |

Update, August 25, 2026: Sol was repriced too

On July 30, 2026 OpenAI [cut Terra by 20 percent and Luna by 80 percent](/blog/did-openai-cut-ai-prices-gpt-5-6/), leaving Sol untouched. On August 21, 2026 the flagship followed: Sol fell to $4 / $20 (cached input $0.40), a promotional rate OpenAI says runs at least through November 21, 2026, with $5 / $30 as the published fallback. OpenAI settled that on September 22, 2026 by releasing GPT-6 Sol at $2 / $10 rather than letting the rate revert. The current-rate column above reflects that. Everything below describes the launch ladder, which is what this post is about.

Update, July 31, 2026: Terra and Luna were repriced

On July 30, 2026 OpenAI [cut Terra by 20 percent and Luna by 80 percent](/blog/did-openai-cut-ai-prices-gpt-5-6/), leaving Sol untouched at that date. The same announcement replaced Priority Processing with Fast mode, billed at 2x standard.

The most quietly important number is Sol’s. At $5 / $30 it is priced at exactly GPT-5.5’s standard rate, the “hold the line” outcome that the [GPT-5.6 pricing breakdown](/blog/gpt-5-6-pricing-cost-per-task/) had flagged as the base case. OpenAI did not charge more for the flagship. It added a cheaper tier instead. Against Anthropic, that lands hard: Sol is half the per-token price of [Claude Fable 5 at $10 / $50](/blog/gpt-5-6-vs-claude-opus-4-8-vs-fable-5/).

__GPT-5.6 output token rates by tier__
| Tool           | Cost per task | Multiple of baseline |
| -------------- | ------------- | -------------------- |
| GPT-5.6 Luna   | $6.00         | \-                   |
| GPT-5.6 Terra  | $15.00        | \-                   |
| GPT-5.6 Sol    | $30.00        | \-                   |
| Claude Fable 5 | $50.00        | \-                   |

Launch-day output token rates per million tokens across the GPT-5.6 tiers, with Claude Fable 5 for context. Sol holds GPT-5.5's $30, while Terra and Luna step down. The flagship did not get more expensive; OpenAI added cheaper rungs below it. Terra and Luna were cut again on July 30, 2026, to $12 and $1.20 of output.Source: [OpenAI, Previewing GPT-5.6 Sol (June 2026); Anthropic pricing](https://openai.com/index/previewing-gpt-5-6-sol/)

## Two reasoning modes, and a Cerebras speed claim

Every tier ships with two reasoning modes. Max mode gives the model the most time to reason down a single deep chain, trading latency for accuracy. Ultra mode splits the work across parallel subagents that coordinate on one task. The benchmark gap between them is real: in OpenAI’s reported Terminal-Bench 2.1 numbers, Sol scores 88.76 percent in max mode and 91.91 percent in ultra, against GPT-5.5 at 83.4 percent.

88.8%

Sol (max) on Terminal-Bench 2.1

vs Claude Mythos 5 at 88.0%

91.9%

Sol (ultra) on Terminal-Bench 2.1

parallel subagents mode

750 tok/s

Sol on Cerebras, from July

about 11x GPT-5.5 output speed

OpenAI also said Sol will run on Cerebras wafer-scale hardware at up to 750 tokens per second starting in July. That sounds like a model breakthrough but is mostly a chip story: the same hardware has run smaller models far faster, and Sol is the slowest of them because it is the largest. The full breakdown is in [GPT-5.6 Sol on Cerebras: 750 tokens per second](/blog/gpt-5-6-sol-coding-leaderboard-cost/). Speed changes what the model is usable for, not what it costs per token.

## The catch: you probably cannot use it yet

The suite launched into a limited preview open only to about 20 partner organizations, through the API and Codex, at the direction of the US government. The arrangement followed a Trump-administration request to stagger the release on national security grounds, built on a [June 2, 2026 executive order](/blog/gpt-5-6-delayed-government-ai-release-gate/) that invites federal vetting of frontier models before release.

OpenAI does not like the arrangement, and said so

In its launch note OpenAI stated it does not believe this kind of government access process should become the long-term default, arguing it keeps the best tools from users, developers, enterprises and cyber defenders. The same gate had already led Anthropic to disable Fable 5 and Mythos 5\. The policy mechanics, and why this favors incumbents, are in the [government-as-release-gate breakdown](/blog/gpt-5-6-delayed-government-ai-release-gate/).

That preview-only status is the asterisk on every other number here. The pricing is published and the benchmarks are reported, but for most teams GPT-5.6 is not yet something you can run. OpenAI said wider access is planned in the weeks after launch. The leak cycle that preceded all this, and how the prediction markets misread the timing, is covered in [the GPT-5.6 preview leak post](/blog/gpt-5-6-preview-leak-release-date/).

## What to do with this now

If you are budgeting, the practical takeaway is that the GPT-5.6 flagship launched without raising the headline rate, and has since fallen under it: the promotional $4 / $20 Sol rate is below GPT-5.5’s $5 / $30 through at least November 21, 2026, and the Terra and Luna tiers open cheaper routing for high-volume work. Drop the confirmed rates into the [cost-per-task calculator](/ai-coding-cost-calculator/) to see what they mean for your own token counts, compare tiers head to head in the [model comparison tool](/ai-model-comparison/), and watch the [AI model release tracker](/ai-models/) for the moment broader access opens.

## Frequently asked questions

What is GPT-5.6 Sol, Terra and Luna?

GPT-5.6 is the OpenAI model suite launched June 26, 2026 in three tiers: Sol is the flagship for hard coding, security and agentic work; Terra is a mid tier at GPT-5.5-level performance for half the cost; and Luna is a fast, low-cost tier for routine tasks. All three offer max and ultra reasoning modes.

How much does GPT-5.6 cost?

Per million tokens, as of August 25, 2026: Sol is $4 input and $20 output (a promotional rate that OpenAI settled on September 22, 2026 by shipping GPT-6 Sol at $2 / $10 instead of letting it revert), Terra is $2 and $12, and Luna is $0.20 and $1.20\. Cached input bills 90 percent off on each tier ($0.40, $0.20 and $0.02). Sol launched at $5 / $30, matching GPT-5.5; Terra and Luna launched higher, at $2.50 / $15 and $1 / $6, and were cut 20 percent and 80 percent on July 30, 2026, with Sol following on August 21, 2026.

Is GPT-5.6 better than Claude Mythos 5?

On the agentic-coding benchmark Terminal-Bench 2.1, OpenAI reports Sol ahead: 88.76 percent in max mode and 91.91 percent in ultra mode, against Claude Mythos 5 at 88 percent. OpenAI also reports matching Mythos 5 on cybersecurity benchmarks with roughly one-third fewer output tokens. Independent comparisons are still limited.

Can I use GPT-5.6 right now?

Only if you are one of about 20 government-approved partners. GPT-5.6 launched in a limited preview via the API and Codex at the direction of the US government, and OpenAI said wider public access is planned in the weeks after the June 26 launch.

## Sources

* OpenAI. (2026). _Previewing GPT-5.6 Sol: a next-generation model_. Primary announcement; the June 26 2026 launch of Sol, Terra and Luna, tier pricing, reasoning modes, benchmarks, and limited-preview access. Verified 2026-06-27\. [openai.com/index/previewing-gpt-5-6-sol](https://openai.com/index/previewing-gpt-5-6-sol/)
* The Decoder. (2026). _OpenAI’s GPT-5.6 Sol launches to rival Claude Mythos under government access rules it calls unsustainable_. Secondary coverage; pricing, Cerebras speed, the unsustainable-access quote. Verified 2026-06-27\. [the-decoder.com](https://the-decoder.com/openais-claude-mythos-competitor-gpt-5-6-sol-launches-under-government-controlled-access-it-calls-unsustainable/)
* MarkTechPost. (2026). _OpenAI previews GPT-5.6 with Sol, Terra and Luna: tiered models, new reasoning modes, limited access_. Secondary coverage; tier positioning, max and ultra modes, Terminal-Bench 2.1 figures. Verified 2026-06-27\. [marktechpost.com](https://www.marktechpost.com/2026/06/26/openai-previews-gpt-5-6-with-sol-terra-and-luna-tiered-models-new-reasoning-modes-limited-access/)
* VentureBeat. (2026). _OpenAI unveils GPT-5.6 Sol, Terra and Luna models, but only accessible to limited preview partners for now, per US Gov_. Secondary coverage; limited-preview access to about 20 partners. Verified 2026-06-27\. [venturebeat.com](https://venturebeat.com/technology/openai-unveils-gpt-5-6-sol-terra-and-luna-models-but-only-accessible-to-limited-preview-partners-for-now-per-us-gov)