When was GPT-5.6 released?

GPT-5.6 was previewed to select partners on June 26, 2026, and released publicly on July 9, 2026, following clearance from the U.S. Department of Commerce.

What are Sol, Terra, and Luna?

They are the three tiers of the GPT-5.6 generation. Sol is the flagship, Terra is the balanced everyday model, and Luna is the fast, low-cost model. The number marks the generation; the names mark the capability tier and are designed to persist across future generations.

How much does GPT-5.6 cost?

Per million tokens (input/output): Sol is $5/$30, Terra is $2.50/$15, and Luna is $1/$6. A heavy 2M-in/500K-out agent session costs about $25 on Sol, $12.50 on Terra, and $5 on Luna.

Which GPT-5.6 tier should I use?

For most people, Terra — it offers near-flagship quality at half the flagship price. Use Luna for high-volume simple tasks and Sol for genuinely hard work or huge-context jobs.

What is the GPT-5.6 context window?

Sol has a 1.5-million-token context window, up roughly 43% from GPT-5.5 Pro's 1.05M and the largest of any flagship model at launch.

What is the difference between "max" and "ultra" modes?

Max deepens single-model reasoning for one hard problem; ultra spawns parallel sub-agents for work that can be split up. Both cost more per task than a standard call.

Is GPT-5.6 better than GPT-5.5?

Yes on several axes — bigger context, a new parallel "ultra" mode, better token efficiency, and higher agentic-coding benchmarks — while holding the flagship price flat and cutting the mid-tier price in half.

Is GPT-5.6 better than Claude Fable 5?

It's a split. Sol leads on price, context, and Terminal-Bench 2.1; Fable 5 leads on SWE-Bench Pro (real-world software repair) and multi-day autonomy. See our full comparison for details.

Can I use GPT-5.6 for free?

ChatGPT's plan-level availability (which tiers free users get, and any Sol caps) is being finalized after launch. Check OpenAI's current release notes for what your plan includes.

What are the API model names for GPT-5.6?

Following the GPT-5.5 precedent, the expected identifiers are gpt-5.6-sol, gpt-5.6-terra, and gpt-5.6-luna. Route your code on the tier names rather than version numbers.

Does GPT-5.6 support zero data retention?

Yes — OpenAI retains zero-data-retention options for GPT-5.6, which can be decisive for regulated buyers whose compliance rules forbid data retention.

What You Should Know About the GPT-5.6 Release (2026 Guide)

Q: Why did the government have to approve GPT-5.6?

The launch was gated by U.S. Department of Commerce oversight — the most explicit government involvement in a frontier-model launch so far — reflecting growing treatment of top-tier AI as strategically sensitive.

Introduction

For nine months, OpenAI shipped a new model almost every few weeks — GPT-5, then 5.1, 5.2, 5.3, 5.4, 5.5 — each with its own tangle of Instant, Thinking, Pro, mini, and nano labels. If you found it hard to keep track of which model you were even talking to, you were not alone. GPT-5.6 is OpenAI's attempt to end that confusion, and it changes more than a version number.

Previewed to a small group of partners on June 26, 2026 and made publicly available on July 9, 2026, GPT-5.6 is not one model. It is a family of three — Sol (the flagship), Terra (the balanced everyday model), and Luna (the fast, cheap one) — with a naming system designed to survive future generations. It also arrives with the largest context window OpenAI has ever shipped, a genuinely new "ultra" mode, an aggressive pricing strategy aimed squarely at Anthropic, and — unusually — a launch that required clearance from the U.S. Department of Commerce.

This guide is the plain-English, no-hype explainer of everything you should know about the GPT-5.6 release: what it is, what actually changed, what each tier costs, how the benchmarks really look (including the number OpenAI conspicuously hasn't published), how to start using it today, and which tier is right for you — whether you're a curious beginner, a working developer, a business buyer, a student, or an agency. Everything here is grounded in OpenAI's launch materials and high-authority coverage, current as of July 11, 2026, two days after public launch.

If you want the head-to-head with Anthropic's flagship, that lives in our dedicated GPT-5.6 vs Claude Fable 5 comparison. This article is about GPT-5.6 on its own terms.

1. The 60-Second Summary

GPT-5.6 is OpenAI's newest model generation, released July 9, 2026, and its biggest change is structural: instead of one model with confusing sub-labels, it ships as three named tiers — Sol, Terra, and Luna. Sol is the flagship ($5 input / $30 output per million tokens) with a 1.5-million-token context window and the top published score on the Terminal-Bench 2.1 agentic-coding benchmark (88.8%, or 91.9% in its parallel "ultra" mode). Terra is the balanced everyday model ($2.50 / $15), positioned as "GPT-5.5-class at half the price." Luna is the fast, high-volume option ($1 / $6). Under the new system, the number marks the generation and the name marks the capability tier — so Sol, Terra, and Luna are meant to persist across future releases.

If you remember three things: (1) it's a family, not a single model; (2) the flagship got cheaper relative to rivals while the context window grew to a class-leading 1.5M tokens; and (3) OpenAI is betting on frontier intelligence at commodity prices.

✅ Recommendation: Most people should start on Terra — it delivers near-flagship quality at half the flagship price and is the right default for everyday writing, analysis, and coding. Reach for Sol only when a task is genuinely hard, and Luna when volume and speed matter more than depth.

2. What Is GPT-5.6?

GPT-5.6 is the newest generation of OpenAI's large language models, released publicly on July 9, 2026 after a two-week private preview that began June 26. In raw capability it is an incremental-but-real step beyond GPT-5.5, with stronger agentic coding, better token efficiency, and a bigger context window. But the headline story is not a single benchmark — it's the product structure.

Every previous point release in the GPT-5 line was, at heart, one model with optional behaviors bolted on: an "Instant" fast path, a "Thinking" reasoning path, a "Pro" heavy variant, and "mini"/"nano" shrink-downs. GPT-5.6 replaces that with a clean, permanent ladder of three tiers you choose between explicitly:

Sol — the flagship, for the hardest reasoning and the most demanding agentic work.
Terra — the balanced model, tuned for everyday professional tasks.
Luna — the fast, inexpensive model for high-volume, latency-sensitive workloads.

The naming is deliberate. The number (5.6) identifies the generation; the name (Sol / Terra / Luna) identifies the capability tier. OpenAI's stated intent is that these three names outlive 5.6 — so a future "GPT-6 Sol" would be the flagship of that generation, and your routing logic wouldn't have to change. It is, in effect, OpenAI adopting the Haiku / Sonnet / Opus pattern that Anthropic has used for two years, just with celestial names instead of poetic ones.

💡 Expert Tip: When someone says "I'm using GPT-5.6," always ask which tier. The gap between Luna and Sol in both cost and capability is large — comparable to the gap between a compact car and a flagship sedan. "GPT-5.6" alone is no longer a precise statement.

3. Meet the Family: Sol, Terra, and Luna

Here is the family at a glance, then a closer look at each member.

Tier	Role	Price (per 1M tokens, in/out)	Best for
Sol	Flagship — hardest reasoning, advanced agents	$5 / $30	Complex coding, huge-context analysis, research
Terra	Balanced — everyday professional work	$2.50 / $15	Most writing, analysis, and coding
Luna	Fast & affordable — high-volume, low-latency	$1 / $6	Classification, extraction, chat, high-traffic apps

Sol — the flagship

Sol is where OpenAI concentrated its best work. Three things define it:

A 1.5-million-token context window — up roughly 43% from GPT-5.5 Pro's 1.05M, and the largest of any flagship model available at launch. That's on the order of 1,100 pages of text, or a very large codebase, ingested in a single request.
Two heavy modes — "max" (deeper single-model reasoning) and "ultra" (parallel sub-agents working a problem simultaneously). More on both in Section 8.
Efficiency as a design goal. In OpenAI's own evaluations, Sol reaches competitive scores using dramatically fewer output tokens — on the ExploitBench security benchmark it matched a top competitor while spending roughly one-third of the tokens.

Terra — the quiet star

Terra is the tier most people should actually use, and it may be the most strategically important model in the release. OpenAI positions it as competitive with GPT-5.5 at half of GPT-5.5's price — near-flagship quality for $2.50 per million input tokens. For the overwhelming majority of everyday tasks — drafting, summarizing, routine coding, data cleanup — Terra is fast, cheap, and more than good enough. It is the sensible default.

Luna — the volume workhorse

Luna replaces the old mini/nano floor. At $1 / $6 it's built for jobs where you call the model thousands or millions of times: classification, tagging, extraction, routing, autocomplete, and high-traffic consumer chat. You trade some depth for speed and cost, which is exactly the right trade at volume.

📌 Best Practice: Don't put everything on Sol "to be safe." The smartest teams route by difficulty: Luna for high-volume simple calls, Terra as the everyday default, and Sol reserved for the genuinely hard escalations. This single habit can cut an AI bill by 60–80% with no visible quality loss.

4. Why OpenAI Changed the Naming System

To understand why this release matters, you have to appreciate how confusing the previous year had become. Here is the actual cadence that preceded GPT-5.6:

GPT-5 — August 2025
GPT-5.1 (Instant / Thinking) — November 2025
GPT-5.2 — December 2025
GPT-5.3 Instant — February 2026
GPT-5.4 (+ Thinking, Pro, mini, nano) — March 2026
GPT-5.5 (codename "Spud") — April 2026

Six releases in nine months, each layering on its own Instant/Thinking/Pro/mini/nano labels. Developers maintained brittle routing tables. Non-technical users had no idea whether "GPT-5.4 Thinking mini" was better or worse than "GPT-5.3 Instant." The model picker had become a maze.

Sol, Terra, and Luna are the fix. By separating the generation number from a small set of stable tier names, OpenAI gives everyone a durable mental model: pick your tier once (by how hard your work is), and let the generation number advance underneath it. It mirrors Anthropic's long-running Haiku / Sonnet / Opus ladder — an implicit acknowledgment that the competitor's naming was simply clearer.

⚠️ Common Mistake: Assuming the tier names imply a speed order like the old "Instant/Thinking" labels. They don't map to a single axis. Luna is fastest and cheapest, Sol is most capable, and Terra balances the two — but all three can "think." Choose by task difficulty and budget, not by a speed label.

5. What's Actually New vs GPT-5.5

Beyond the naming, here is what genuinely changed under the hood.

1. A bigger context window. Sol's 1.5M tokens is a ~43% jump over GPT-5.5 Pro's 1.05M and the largest flagship window at launch. If your bottleneck was "I can't fit the whole codebase / data room / discovery corpus in one prompt," that ceiling just rose meaningfully.

2. The new "ultra" mode. GPT-5.5 had a heavy "Pro"-style path; GPT-5.6 introduces genuine parallel sub-agents under Sol's "ultra" mode. Instead of one model thinking longer, ultra spins up several workers that split a problem and reconcile — which is precisely how Sol converts an 88.8% Terminal-Bench score into 91.9%.

3. Better token efficiency. This is underrated. Reaching a given quality with fewer output tokens means lower cost and lower latency for the same result. Sol's one-third-tokens result on ExploitBench is the clearest example, and it applies across the family.

4. Raw speed on specialized hardware. For select customers, Sol runs on Cerebras hardware at up to ~750 tokens per second — an order of magnitude faster than typical flagship serving. In agent loops that call the model hundreds of times, that compounds into real wall-clock savings.

5. Inherited computer-use skills. The GPT-5.4 generation introduced OpenAI's native computer-use capability (operating real GUI applications), and GPT-5.5 scored 78.7% on OSWorld-Verified. GPT-5.6 inherits and extends this — relevant if your agents must drive apps and browsers, not just terminals and APIs.

6. Aggressive repricing. GPT-5.5 launched at $5/$30, a 2× jump over GPT-5.4 that developers complained about for weeks. GPT-5.6 holds Sol at that same $5/$30 while delivering a new generation — and drops Terra to $2.50/$15 as, effectively, "GPT-5.5-class at half price." More on this next.

🚀 Pro Tip: If you upgraded a workload to GPT-5.5 in April and swallowed the price hike, re-benchmark it on Terra now. In many cases you'll get equal-or-better quality at half the cost — the single easiest AI cost win available this quarter.

6. Pricing: What Each Tier Costs

Pricing is where GPT-5.6's strategy is clearest. Here is the family alongside Anthropic's ladder, per million tokens (input / output):

OpenAI (GPT-5.6 family)	Price (in/out)	For reference: Anthropic	Price (in/out)
Sol	$5 / $30	Claude Fable 5	$10 / $50
Terra	$2.50 / $15	Claude Opus 4.8	$5 / $25
Luna	$1 / $6	Claude Sonnet 5	$3 / $15
—	—	Claude Haiku 4.5	$1 / $5

Read it diagonally and the strategy jumps out: OpenAI priced its flagship against Anthropic's second tier. Sol at $5/$30 sits nearer Opus 4.8 ($5/$25) than Anthropic's flagship Fable 5 ($10/$50). Terra undercuts Sonnet 5. OpenAI didn't just match on price — it shifted the whole family one competitive rung down.

What a real workload actually costs

Per-token prices are abstract; agent workloads are output-heavy, so the mix matters. Take a representative heavy coding-agent session of 2M input tokens and 500K output tokens:

Model	Input cost	Output cost	Session total
GPT-5.6 Luna	$2.00	$3.00	$5.00
GPT-5.6 Terra	$5.00	$7.50	$12.50
GPT-5.6 Sol	$10.00	$15.00	$25.00

Run twenty such sessions a week and Luna versus Sol is a difference of roughly $20,000 per seat, per year. That is why tier routing isn't a micro-optimization — at scale it's a budget line.

⚠️ Common Mistake: Comparing only flagship prices when you shop. For most production traffic, the meaningful comparison is mid-tier — Terra ($2.50/$15) against rivals' mid models — with the flagship reserved for hard escalations. Price the tier you'll actually run at volume, not the one in the headline.

7. Benchmarks: How Good Is It, Really?

Two things are true at once: GPT-5.6 Sol posts the best published number on a major agentic benchmark, and OpenAI has left one important number blank. Both facts matter.

Where Sol leads: Terminal-Bench 2.1

Terminal-Bench 2.1 measures agentic command-line work — planning, tool use, and multi-step execution in a terminal. Published results:

Model	Terminal-Bench 2.1
GPT-5.6 Sol Ultra (parallel sub-agents)	91.9%
GPT-5.6 Sol	88.8%
GPT-5.5	88.0%
Claude Mythos 5	84.3%
Claude Fable 5	83.4%
Claude Opus 4.8	78.9%
Gemini 3.1 Pro Preview	70.7%

On terminal-agent orchestration, Sol is at the top of the published charts — and its "ultra" mode extends the lead.

The number OpenAI didn't publish: SWE-Bench Pro

SWE-Bench Pro measures something harder and arguably more real: resolving actual GitHub issues end to end — reading an unfamiliar codebase, finding the fault, and shipping a fix that passes tests. Published results:

Model	SWE-Bench Pro
Claude Fable 5	80.3%
Claude Opus 4.8	~69%
GPT-5.5	58.6%
GPT-5.6 Sol	not published

That empty cell is the most interesting number in the release. Anthropic's Fable 5 leads SWE-Bench Pro by a wide margin, and OpenAI has not put Sol on the board. Until it does — or until independent labs run it now that the preview is over — the honest reading is: Sol demonstrably leads on terminal-agent orchestration; Fable 5 demonstrably leads on real-world software repair.

The efficiency angle

One result deserves more attention: on ExploitBench (a security-research benchmark), Sol matched a top competitor's scores using roughly one-third of the output tokens. Combined with the Cerebras speed path, OpenAI's pitch isn't just "smart" — it's "smart, terse, and fast."

How much should you trust any of this?

The responsible caveat: vendor-published benchmarks are marketing artifacts until independently reproduced. Three cautions specific to this launch:

All of GPT-5.6's launch numbers came from the closed preview, when outside labs couldn't verify them. Independent scores are only now (post-July 9) beginning to appear.
"Sol Ultra" is a mode, not the base model. Its 91.9% uses parallel sub-agents at higher cost per task. For a like-for-like read, compare base Sol (88.8%).
Benchmark selection is strategic. Each lab leads with the chart it wins. Assume the same of both.

📌 Best Practice: Build a 20-task evaluation set from your own backlog — real tickets, real documents, real spreadsheets — and run the tiers on it before you commit spend. A private eval of your actual work beats every public leaderboard for a buying decision.

8. "Max" and "Ultra": The Two Heavy Modes

Sol offers two ways to spend more compute for a better answer, and understanding the difference saves money.

"Max" mode deepens single-model reasoning — the model thinks longer and harder along one line of attack. Use it for a single, coherent, hard problem: a tricky proof, a subtle bug, a nuanced piece of analysis.
"Ultra" mode spawns parallel sub-agents that split a problem, work in parallel, and reconcile their results. Use it for work that decomposes: sweep a codebase for a class of bug, process a large batch, or run a matrix of experiments. Ultra is how Sol turns 88.8% into 91.9% on Terminal-Bench — and it costs more per task because it's doing more work.

The practical rule: max for depth on one thread, ultra for breadth across many. Both cost more than a standard Sol call, so reserve them for tasks where the marginal quality is worth the marginal spend.

💡 Expert Tip: Ultra mode is not "always better." On a single indivisible problem, parallel sub-agents can waste compute reconciling redundant work. Match the mode to the shape of the task: divisible → ultra, indivisible-but-hard → max, everything else → standard Terra or Sol.

9. Speed: The Cerebras Fast Path

For most users, GPT-5.6 runs at typical cloud-serving speeds. But for select customers, OpenAI is deploying Sol on Cerebras hardware at up to ~750 tokens per second — roughly an order of magnitude faster than standard flagship serving.

Why does this matter beyond a bragging number? Because latency compounds in agent loops. When a model is called hundreds of times in a single autonomous task, shaving each call from seconds to a fraction of a second turns hours of wall-clock time into minutes. For interactive products — a coding assistant, a live support agent, a real-time voice mode — that speed is the difference between "usable" and "delightful." (Speed is also central to OpenAI's new real-time product; see our GPT Live guide.)

🚀 Pro Tip: If your use case is latency-sensitive and high-volume, ask whether the Cerebras fast path is available for your account — and design your agent loops to exploit it (more, smaller calls) rather than fighting it (few giant calls). Speed changes the optimal architecture, not just the clock.

10. Why the Government Was Involved

One detail set the GPT-5.6 launch apart from every prior release: it was gated by the U.S. Department of Commerce. The private preview ran under Commerce Department oversight, and the public release proceeded only after regulatory clearance — the most explicit government involvement in a frontier-model launch to date.

For everyday users, this changes nothing about how the model behaves. But it's a meaningful signal about where the industry is heading: frontier AI is increasingly treated like a strategically sensitive technology, with export-control-style scrutiny applied to the most capable systems. If you're a business planning long-term AI strategy, expect more of this — regulatory checkpoints, regional availability differences, and compliance surfaces that vary by model tier and jurisdiction.

📌 Best Practice: Bake regulatory variability into your AI roadmap. Don't assume a model available in one region on launch day is available everywhere, or that access rules won't change. Keep your stack portable across at least two providers so a single policy shift can't strand a critical workload.

11. How to Access GPT-5.6 (Step by Step)

In ChatGPT (for everyone)

Step 1 — Open ChatGPT and click the model selector. In the web app, the model name sits at the top-left of the conversation view (e.g., "GPT-5.5"). Click it and a dropdown opens.

12. Which Tier Should You Use?

Match the tier to who you are and what you're doing.

Beginners and students. Start with Terra (or Luna for the fastest, cheapest experience). You get near-flagship quality without paying flagship prices, inside the ChatGPT interface you likely already use. Only reach for Sol when a specific task is genuinely stumping the cheaper tiers.

Professional developers. Use Terra as your daily driver and Sol for hard, cross-file work — turning on "ultra" for parallelizable jobs and "max" for single deep problems. The pick_model routing pattern above is the pragmatic default. For a broader tool view, see our best AI coding assistants roundup.

Businesses. Route by task shape. Put high-volume simple traffic on Luna, everyday work on Terra, and reserve Sol for the escalation tier. Run a private 20-task eval before standardizing, and remember the ~$20K/seat/year gap between tiers makes routing a real budget lever.

Agencies. Terra covers most client production work economically; Sol handles the flagship deliverables where polish is billed. Luna is ideal for high-volume pipelines (bulk content classification, metadata, first-pass drafts).

Regulated industries. GPT-5.6 retains zero-data-retention options, which can be decisive if your compliance rules forbid data retention. Confirm ZDR availability for your account and region before building.

✅ Recommendation: If you take one action after reading this, make it this: move your default workload from Sol (or GPT-5.5) to Terra and measure the result. For most tasks you'll see equal quality at half the cost — and you can always escalate the genuinely hard cases to Sol.

13. GPT-5.6 vs the Competition

GPT-5.6 doesn't exist in a vacuum. Here's the quick landscape as of mid-2026:

vs Claude Fable 5 (Anthropic). The defining rivalry of the release. Sol wins on price (half of Fable 5), context (1.5M vs 1M), and published terminal-agent benchmarks; Fable 5 wins on real-world software repair (SWE-Bench Pro) and multi-day autonomous runs. It's a genuine split — full detail in our GPT-5.6 vs Claude Fable 5 comparison and the Claude Fable 5 guide.
vs Gemini 3.1 Pro (Google). Gemini remains the pick for deep Google Workspace integration and real-time search grounding, and trails the leaders on the headline agentic benchmarks here. See ChatGPT vs Gemini for the ecosystem view.
vs multi-agent orchestrators (e.g., Sakana Fugu). A different bet entirely — systems that coordinate several frontier models rather than being one. If "best possible answer regardless of vendor" is your goal, the Sakana Fugu guide is worth reading alongside this.

The honest framing: there is no single "best model" in mid-2026. There is a best tier for each job, and increasingly, the strongest teams route across tiers and even vendors rather than pledging loyalty to one.

🚀 Pro Tip: The most cost-effective 2026 stacks are hybrid — Luna/Terra for volume, Sol for hard escalations, and a rival flagship on hand for the tasks it wins. Cross-vendor routing is now a normal engineering practice, not a hedge.

14. Limitations and What to Watch

A definitive guide owes you the caveats.

The unpublished SWE-Bench Pro score. Until OpenAI or independent labs post Sol's real-world software-repair number, treat "best coding model" claims with care. Sol leads terminal-agent work; the repair crown is currently Anthropic's.
Preview-period benchmarks. Every launch number came from the closed preview. Independent replication is only just beginning; expect the picture to sharpen over the coming weeks.
Heavy-mode cost. "Max" and especially "ultra" can multiply the cost of a task. Without per-tier budgets and deliberate use, bills can surprise you.
Plan availability is still settling. Which tiers free users get, and any Sol caps, are being finalized post-launch. Confirm before you depend on a specific tier.
Regional and regulatory variability. The Commerce Department gating hints at more jurisdictional differences ahead. Availability may not be uniform worldwide.
Still a language model. GPT-5.6 can be confidently wrong, and it doesn't "know" anything it wasn't trained on or given. Verify important facts and keep a human in the loop for consequential work.

⚠️ Common Mistake: Treating launch-day benchmark charts as settled fact. The most valuable data — independent, reproducible scores on your kind of work — arrives after launch, not on it. Reserve judgment (and big migrations) until you've run your own eval.

15. The Bottom Line

GPT-5.6 is less a leap in raw intelligence than a leap in clarity and value. The three-tier Sol / Terra / Luna structure finally makes OpenAI's lineup legible; the aggressive pricing — a held-flat Sol and a half-price Terra — hands most users a real cost win; and the 1.5M context window plus the new ultra mode give the flagship genuine new capability. The one asterisk is the unpublished SWE-Bench Pro score, which keeps the "best coding model" title contested with Anthropic's Fable 5.

For nearly everyone, the practical takeaway is simple: default to Terra, escalate hard cases to Sol, use Luna at volume, and run your own eval before betting the business on any leaderboard. Do that, and GPT-5.6 is one of the easiest upgrades of the year.

Want the head-to-head that this release is really defined by? Read our GPT-5.6 vs Claude Fable 5 comparison. Curious about OpenAI's new real-time voice-and-vision mode built on this generation? See our GPT Live guide. Or browse the full directory of AI chatbots to see where GPT-5.6 fits in your stack.