GPT-5.6 Introduces Sol, Terra & Luna: Everything You Need to Know About OpenAI's Latest AI Models (2026)

OpenAI just changed how it ships models, and most people haven't caught up yet.

On July 9, 2026, GPT-5.6 went fully live across ChatGPT, Codex, and the API — but it didn't arrive as one model. It arrived as three: Sol, Terra, and Luna. Same generation, three different jobs, three different price tags.

If you're still picking "the smart model" or "the fast model" the way you did a year ago, this guide is going to save you a lot of wasted tokens. Here's everything confirmed so far about GPT-5.6, straight from OpenAI's own release materials and verified post-launch reporting — no guessed numbers, no filler.

What Is GPT-5.6?
Why OpenAI Released Sol, Terra, and Luna
Overview of Each Model
GPT-5.6 Sol Explained
GPT-5.6 Terra Explained
GPT-5.6 Luna Explained
Feature Comparison Table
Performance Benchmarks
Pricing Breakdown
Best Model for Different Users
GPT-5.6 vs. GPT-5.5: What Actually Changed
Real-World Applications
Benefits
Limitations
Frequently Asked Questions
Final Verdict

1. What Is GPT-5.6?

GPT-5.6 is OpenAI's newest model generation, and it's the first one released as a family instead of a single system. Instead of one flagship model with a couple of stripped-down "mini" versions bolted on, OpenAI built three separate tiers from the same generation: Sol, Terra, and Luna.

The naming itself is new too. Going forward, the number (5.6) marks the generation, while the name — Sol, Terra, or Luna — marks a durable capability tier. That means OpenAI can update Terra or Luna independently down the line without needing to relaunch the whole family under a new number.

GPT-5.6 rolled out in two stages. A limited preview opened on June 26, 2026, restricted to roughly 20 vetted partner organizations as part of a coordinated review with the U.S. government. Full general availability followed on July 9, 2026, opening the models up across ChatGPT, ChatGPT Work, Codex, and the OpenAI API, with the rollout completing globally within about 24 hours.

All three models share the same underlying specs: a context window listed at roughly 1 million to 1.05 million tokens, a maximum output of 128,000 tokens, and a knowledge cutoff of February 16, 2026.

2. Why OpenAI Released Sol, Terra, and Luna?

The short answer: because a single giant model trying to be everything to everyone stopped making economic sense.

Running a flagship-level model to summarize an email or tag a support ticket is like hiring a surgeon to put on a Band-Aid. It works, technically, but it's wildly overpriced for the job. GPT-5.6's three-tier structure exists to fix that mismatch.

There's also a genuine engineering upside. By decoupling the tiers, OpenAI can push updates to Luna's speed or Terra's cost-efficiency without touching Sol's reasoning stack, and vice versa. Each tier effectively gets its own release cadence.

This isn't happening in isolation, either. Competing labs have been moving toward similarly tiered lineups, and GPT-5.6's family structure looks like OpenAI formalizing a pattern the whole industry has been drifting toward: match the model to the task, not the task to whatever model happens to be the newest.

3. Overview of Each Model:

Here's the elevator-pitch version before we go deeper on each one:

Sol — the flagship. Built for the hardest coding, research, and agentic work. Highest cost, highest ceiling.
Terra — the balanced middle tier. Roughly GPT-5.5-class quality at about half the price, aimed at everyday production workloads.
Luna — the speed-and-cost tier. Built for high-volume, latency-sensitive tasks like summarizing, tagging, and drafting.

Think of it less as "good, better, best" and more as three different tools that happen to share a toolbox.

4. GPT-5.6 Sol Explained:

Sol is the model OpenAI leads with in every benchmark table it publishes, and it's the only tier that unlocks the new "max" reasoning effort and "ultra" multi-agent mode.

On the Artificial Analysis Coding Agent Index — an independent benchmark measuring real coding-agent performance — Sol posted a new state-of-the-art score of 80 at max reasoning, ahead of Anthropic's Claude Fable 5, while using noticeably fewer output tokens and less time to get there. That efficiency angle matters as much as the raw score, since output tokens are what actually show up on your bill.

Sol also performs strongly on long-horizon, multi-step agentic work. On Agents' Last Exam, a benchmark testing long-running professional workflows across 55 different fields, Sol set a new high score of 53.6, a meaningful jump over Claude Fable 5's adaptive-reasoning result.

It's not a clean sweep, though, and a good guide should say so plainly. On SWE-Bench Pro, a widely watched software-engineering benchmark, Sol scored 64.6% — trailing Claude models by roughly 15 points. Independent post-launch testing from Artificial Analysis also placed Sol at #2 on its aggregate intelligence index, just behind Claude Fable 5, even while Sol kept the top spot on the Coding Agent Index specifically. In other words: dominant at coding-agent work, competitive but not always first everywhere else.

Sol also ships with the most aggressive safety stack OpenAI has built to date, with strengthened protections around cybersecurity misuse, given how capable the model is at vulnerability research and exploit-adjacent tasks.

Sol is best suited for: complex coding agents, cybersecurity and vulnerability research, long-running multi-step workflows, and any task where getting it wrong is expensive.

5. GPT-5.6 Terra Explained:

Terra is the tier OpenAI positions as the default for everyday production work, and the pitch is straightforward: performance competitive with GPT-5.5, at roughly half the cost.

That's a meaningful claim if it holds up under real workloads, because GPT-5.5 was itself a strong, widely deployed model. Getting comparable output at half the price changes the math for anyone running high-volume API calls.

Terra's role, based on OpenAI's own framing, centers on high-volume business tasks — customer support automation, document analysis, and internal tooling — rather than chasing frontier-level benchmark records. On several knowledge-work evaluations, Terra doesn't just match GPT-5.5's peak performance, it beats it, while costing less per token.

There's a wrinkle worth flagging: on certain agentic benchmarks, Luna — the tier positioned below Terra — actually scores higher than Terra. It's not a universal pattern, but it's a useful reminder that tier position and benchmark performance don't always line up perfectly across every test.

Terra is best suited for: customer support pipelines, document processing, internal business tools, and any workload running at real production scale where per-token cost adds up fast.

6. GPT-5.6 Luna Explained:

Luna is the fastest and cheapest tier in the family, built specifically for high-throughput, latency-sensitive work.

At its price point, Luna is designed to be thrown at large volumes of routine requests without much hesitation — think classification, tagging, drafting, and quick summarization. And despite sitting at the bottom of the pricing ladder, Luna isn't automatically the weakest performer across the board; as mentioned above, it edges out Terra on some agentic benchmarks, which is a genuinely useful thing to know before you assume higher price always means better results for your specific task.

Where Luna does show real limitations is long-context recall. On extended-context benchmarks testing information retrieval across very large documents, Luna's scores drop noticeably compared to Sol and Terra. If your task involves stitching together facts scattered across a massive document or codebase, Luna's speed comes with a real trade-off in that specific area.

Luna is best suited for: real-time chat features, high-frequency automation, first-pass drafting, and any application where response speed matters more than deep reasoning.

7. Feature Comparison Table:

Feature	Sol	Terra	Luna
Role	Flagship	Balanced everyday tier	Fastest, lowest-cost tier
Input price (per 1M tokens)	$5.00	$2.50	$1.00
Output price (per 1M tokens)	$30.00	$15.00	$6.00
Context window	~1M–1.05M tokens	~1M–1.05M tokens	~1M–1.05M tokens
Max output	128K tokens	128K tokens	128K tokens
Max reasoning effort	Yes	No	No
Ultra (multi-agent) mode	Yes	No	No
ChatGPT standard chat access	Yes (Plus and above)	No (Work/Codex only)	No (Work/Codex only)
Best for	Hard coding, cybersecurity, long agentic tasks	High-volume business workflows	Real-time, high- throughput tasks

8. Performance Benchmarks:

A few of the headline, OpenAI-reported and independently verified numbers worth knowing:

Coding Agent Index (Artificial Analysis): Sol scores 80, a new state-of-the-art result, ahead of Claude Fable 5.
Agents' Last Exam: Sol scores 53.6, its highest recorded result on this 55-field long-horizon benchmark.
SWE-Bench Pro: Sol scores 64.6%, trailing Claude Fable 5 and Claude Mythos 5, which both score above 80%.
Terminal-Bench 2.1: Sol scores 88.8% standard, rising to 91.9% in Ultra multi-agent mode; Terra scores 84.3%, Luna 82.5%, versus GPT-5.5's 83.4%.
BrowseComp (agentic browsing): Sol posts a new state-of-the-art result at 92.2%.
Artificial Analysis Intelligence Index (independent, post-launch): Sol lands at #2 overall, just behind Claude Fable 5.

A quick honesty check worth repeating: OpenAI itself flagged concerns about SWE-Bench Pro's reliability shortly after GPT-5.6's launch, estimating a notable share of tasks on that benchmark may be flawed. That's a useful reminder that even official benchmark tables deserve a skeptical read, and running your own evaluation against your actual workload beats trusting any single leaderboard.

9. Pricing Breakdown:

GPT-5.6 pricing is set per 1 million tokens and has remained unchanged from the June preview through the July general-availability launch:

Sol: $5.00 input / $30.00 output
Terra: $2.50 input / $15.00 output
Luna: $1.00 input / $6.00 output

GPT-5.6 also introduced a reworked prompt caching system. Cache writes are billed at 1.25 times the standard uncached input rate, while cache reads keep a 90% discount off the standard input price. There's also a guaranteed 30-minute minimum cache life and support for explicit cache breakpoints, giving developers more predictable control over caching costs on long-running agent sessions.

For context on the pricing tiers themselves: running the same volume of work through Luna instead of Sol costs roughly a fifth as much per token — a gap wide enough to justify actually testing whether the cheaper tier can handle your workload before defaulting to the flagship.

10. Best Model for Different Users:

If you're an everyday user: Luna-tier speed covers most casual tasks — drafting messages, summarizing articles, brainstorming — without any noticeable quality drop-off for that kind of work.

If you're a developer building coding agents: Sol is the clear starting point for anything long-running or complex, with Terra as a cheaper option worth testing for scoped, well-defined tasks.

If you're running a business support or document pipeline: Terra is built specifically for this — strong quality at meaningfully lower cost than a flagship-tier model.

If you're building high-frequency, latency-sensitive features: Luna is the obvious default, with escalation to Terra or Sol reserved for requests that actually need the extra depth.

11. GPT-5.6 vs. GPT-5.5: What Actually Changed?

GPT-5.6's direct predecessor in OpenAI's lineup is GPT-5.5, not the earlier GPT-5, and the generational jump brings a few concrete changes:

Structure: GPT-5.5 shipped as a single model with variants; GPT-5.6 ships as three distinct, independently priced tiers.
Cost: Terra reportedly matches or beats GPT-5.5's performance on several benchmarks at roughly half the price.
Reasoning controls: GPT-5.6 adds a new "max" reasoning effort level exclusive to Sol, plus an "ultra" mode that coordinates multiple agents working in parallel — OpenAI says four agents run by default in ultra mode.
Tool use: GPT-5.6 introduces Programmatic Tool Calling, letting the model write and execute short JavaScript routines to orchestrate multiple tool calls, loops, and conditional logic in a single pass, rather than making tool calls one at a time.
Caching: The prompt-caching system was rebuilt with explicit breakpoints and a guaranteed minimum cache life, replacing the less predictable caching behavior in earlier generations.

If you're currently on GPT-5.5 for coding-agent work specifically, Sol is generally worth testing first — the long-horizon coding behavior shows a real, hands-on improvement in staying oriented across multi-file tasks, not just a benchmark bump.

12. Real-World Applications:

Software development: Sol for long-running coding agents and complex debugging; Terra for scoped implementation tasks and first-pass code review.
Customer support: Terra handling high-volume ticket triage and response drafting, with the option to escalate genuinely complex cases to Sol.
Content and marketing teams: Luna for rapid first drafts and summarization, with human editing layered on top.
Cybersecurity teams: Sol for defensive work like threat modeling, code review, and patch development, under OpenAI's stated goal of supporting legitimate security research while constraining offensive misuse.
Data and document-heavy workflows: Terra or Luna for extraction, tagging, and classification at scale, depending on how much long-context recall the task actually needs.
Presentation and document generation: Sol for polished, template-matched output in professional formats like slides, spreadsheets, and reports.

13. Benefits:

Cost control finally has a real dial. Instead of one price point for every task, you can route cheap work to Luna and reserve Sol's cost for tasks that actually need it.
Stronger coding-agent performance. Sol's results on long-horizon coding benchmarks represent a genuine step up for agentic development work.
More predictable caching. Explicit cache breakpoints and a guaranteed minimum cache life make long-running agent costs easier to forecast.
New orchestration tools. Programmatic Tool Calling and ultra mode give developers more sophisticated ways to structure complex, multi-step agent workflows.
Independent tier development. Because each tier can update on its own schedule, improvements to Luna or Terra don't require waiting on a full flagship relaunch.

14. Limitations:

Not a universal win. Sol trails Claude models on SWE-Bench Pro and lands second to Claude Fable 5 on Artificial Analysis's broader intelligence index.
Uneven access. Terra and Luna aren't selectable inside standard ChatGPT conversations — they're currently limited to ChatGPT Work, Codex, and the API, which limits casual access for everyday chat users.
Luna's long-context recall is weaker. Tasks requiring recall across very large documents show a real performance drop on Luna specifically.
Benchmark reliability questions. OpenAI itself raised concerns about flaws in the SWE-Bench Pro benchmark shortly after launch, a reminder that no single leaderboard should be treated as the final word.
Elevated safety classifications. All three tiers carry a "High" capability rating for cyber and biological risk categories in OpenAI's own system card, which may introduce added governance requirements for regulated industries.

15. Frequently Asked Questions:

1. What is GPT-5.6? GPT-5.6 is OpenAI's newest model generation, released as a family of three distinct tiers — Sol, Terra, and Luna — rather than a single model.

2. What's the difference between Sol, Terra, and Luna? Sol is the flagship for the hardest tasks, Terra is a balanced everyday tier at roughly half Sol's price, and Luna is the fastest and cheapest tier, built for high-volume work.

3. When did GPT-5.6 launch? A limited preview began June 26, 2026, followed by full general availability on July 9, 2026.

4. How much does GPT-5.6 cost? Per 1 million tokens: Sol is $5 input / $30 output, Terra is $2.50 / $15, and Luna is $1 / $6.

5. Is GPT-5.6 better than GPT-5.5? On several benchmarks, yes — particularly for coding-agent work, where Sol sets new state-of-the-art scores. Terra also reportedly matches or beats GPT-5.5 on some tasks at about half the cost.

6. Which GPT-5.6 model should I use for coding? Sol for complex, long-running coding agent work; Terra for scoped, well-defined implementation tasks where cost matters more.

7. Can I use Terra or Luna in regular ChatGPT? Not in standard ChatGPT conversations as of launch — they're available through ChatGPT Work, Codex, and the OpenAI API.

8. Is GPT-5.6 available through the API? Yes, all three tiers — gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna — are self-serve through the OpenAI API.

9. What is "ultra mode" in GPT-5.6? A multi-agent setting, available with Sol, that coordinates multiple agents working in parallel on complex tasks — OpenAI says it runs four agents by default.

10. Does GPT-5.6 outperform Claude models? It depends on the benchmark. Sol leads on coding-agent-specific tests, but Claude models currently lead on SWE-Bench Pro and rank slightly ahead on Artificial Analysis's broader intelligence index.

11. What is Programmatic Tool Calling? A new API feature letting GPT-5.6 write and run short JavaScript routines to orchestrate multiple tool calls, loops, and conditional logic within a single response.

12. What's the context window for GPT-5.6? All three tiers share roughly a 1 million to 1.05 million token context window, with a maximum output of 128,000 tokens.

13. Is GPT-5.6 safe to use for cybersecurity tasks? OpenAI built additional safeguards for GPT-5.6 given its cybersecurity capabilities, and all three tiers carry a "High" risk classification for cyber and biological capability in the official system card, meaning added review may apply for sensitive use cases.

14. Why did OpenAI change its naming system? The new format separates generation (the number) from capability tier (the name), letting individual tiers update independently without a full family relaunch.

15. Which model is best for beginners or casual users? Luna covers most everyday tasks efficiently; for deeper reasoning or coding help inside ChatGPT itself, Sol is available at medium effort and above on paid plans.

16. Final Verdict:

GPT-5.6 isn't a bigger model — it's a smarter way of shipping models. Sol earns its flagship price tag on genuinely hard, long-running work, especially coding agents. Terra is the tier most production workloads will actually live in day to day. And Luna proves that "cheapest" doesn't have to mean "weakest," even if it does trade away some long-context recall.

The real shift here isn't any single benchmark score. It's that model selection now works like a dial instead of a light switch — and for anyone building with AI in 2026, learning to use that dial well is quickly becoming as important as picking the right model in the first place.

Key Takeaways:

GPT-5.6 launched as three separate model tiers — Sol, Terra, and Luna — instead of one flagship model.
Sol leads on coding-agent benchmarks; Terra offers GPT-5.5-level performance at about half the cost; Luna is the fastest and cheapest tier.
Pricing per million tokens: Sol $5/$30, Terra $2.50/$15, Luna $1/$6 — unchanged from preview to general availability.
Terra and Luna aren't yet available in standard ChatGPT chat, only in ChatGPT Work, Codex, and the API.
No single tier wins every benchmark — routing tasks by actual difficulty, not habit, is the real strategy shift here.

Featured Snippet Answer (40–60 words): GPT-5.6 is OpenAI's newest AI model generation, released July 2026 as three tiers instead of one model: Sol (flagship, best for complex coding and reasoning), Terra (balanced, cost-efficient everyday model), and Luna (fastest, cheapest tier for high-volume tasks). Pricing ranges from $1/$6 to $5/$30 per million tokens depending on tier.

Search This Blog

AI Tools & Productivity