GPT-5.6 Release: The Complete Guide — Luna/Terra/Sol, Benchmarks, Pricing, and vs. Claude

On July 9, 2026, OpenAI made its new model family "GPT-5.6" generally available (OpenAI's official announcement, following a limited preview on June 26). The biggest change is that OpenAI dropped the old two-tier "standard + Pro" structure in favor of a three-model lineup: Luna (fast, low-cost) / Terra (balanced) / Sol (flagship).

The flagship Sol took the top spot on the Artificial Analysis Coding Agent Index with a score of 80, a measure of coding-agent performance, and on Agents' Last Exam it scored 53.6 — measuring long-running real-world workflows — pulling 13.1 points ahead of Claude Fable 5 (40.5) (Vellum's tally). On the other hand, on production-grade coding measured by SWE-Bench Pro, Claude Fable 5 scored 80.0% while Sol scored 64.6% — an area where GPT-5.6 clearly still loses to Claude.

In this article, drawing on OpenAI's official announcement and several independent benchmark reports, we explain — on a confirmed-facts basis — everything about GPT-5.6: the differences between the three models, pricing, benchmarks, new features, availability by ChatGPT plan, comparison with Claude, and how to choose by use case.

1. Release overview — date, availability, spec sheet

The key point is that all three models share the same roughly 1-million-token context and 128K max output. What differs is "intelligence," "speed," and "price" — the volume of documents you can handle is the same. It's designed so you choose vertically by use case and cost.

2. Luna / Terra / Sol — how the three models differ and how to pick

The biggest change in GPT-5.6 is the reorganization from two tiers (standard/Pro) to three. Here's where each one sits.

The principle for choosing is simple. Make Terra your default, move up to Sol if you need more accuracy, and drop down to Luna when unit cost and speed matter most — this "start from Terra" approach makes it easy to balance cost and quality.

3. Pricing — the mid-tier Terra deserves your attention

What stands out is Terra's price-to-performance. OpenAI describes Terra as "delivering performance rivaling the previous-generation GPT-5.5 at roughly half the price." In other words, many tasks that used to require Sol (roughly equivalent to the old standard tier) can potentially now run at half the cost.

Meanwhile, Sol's $5/$30 matches Claude Opus 4.8 ($5/$25) on input and runs slightly higher on output. Per-unit prices among flagships are converging across vendors, and the gap now shows up in "performance" and "token efficiency."

4. Benchmarks — head-to-head with Claude

Bottom line: strong as an agent, but Claude leads on production-grade coding

GPT-5.6 Sol took first place on overall agent ability (Agents' Last Exam) and on the coding-agent index. On TerminalBench it also edges out Claude Fable 5 slightly. But on SWE-Bench Pro (real production-grade bug fixing in actual repositories), Sol scored 64.6% against Claude Fable 5's 80.0% — a loss by more than 15 points. The honest read is that "the ability to operate autonomously as an agent" and "the ability to fix a real codebase accurately" are two different things.

Also, Luna's long-context recall (MRCR) is 41.3%, far below Sol and Terra. Even though it has a 1-million-token window, Luna is a poor fit for handling very long text with precision. For long-form work, pick Terra or above.

5. What's new — five key points

1. From two tiers to a three-model lineup

The biggest change. The old "standard + expensive Pro" structure was reorganized into three tiers: Luna / Terra / Sol. Terra in particular, with its "GPT-5.5-class at half price" value, becomes the centerpiece of cost planning.

2. Improved token efficiency (54% in coding)

OpenAI CEO Sam Altman said that Sol improved token efficiency by 54% on coding tasks (CNBC). When the same work consumes fewer tokens, total cost drops even if the per-unit price is unchanged. This is exactly why you should compare not just the price sheet but "the real cost per task."

3. Programmatic Tool Calling (Responses API)

As a new developer feature, "Programmatic Tool Calling," where the model generates and runs JavaScript to orchestrate tool calls, was added to the Responses API (Simon Willison's explainer). It lets complex agents that span multiple tools run with fewer round trips. Reasoning effort has six levels: none / low / medium / high / xhigh / max.

4. ChatGPT Work and a new desktop app

Alongside GPT-5.6, OpenAI also announced "ChatGPT Work," an agent for business use, a new desktop app bundled with Codex, and a customer-facing hosting service. It's a move to expand ChatGPT from "just chat" into a "platform that carries out work."

5. The GPT-Live voice model (full-duplex conversation)

On the voice side, a new series called GPT-Live debuted. Rather than the traditional turn-based approach (respond only after the other party finishes speaking), it delivers full-duplex conversation that can interrupt and interject like a human. Paid users get GPT-Live-1; free users get GPT-Live-1 mini.

6. Availability by ChatGPT plan

A notable point is that even Free / Go users can use Terra within ChatGPT Work and Codex. However, freely switching between models and adjusting reasoning effort requires Plus or above. The GPT-Live voice model's mini version is available even on the free tier. (Plan structure and pricing are based on various reports at launch. For the latest, check the official ChatGPT pricing page.)

7. Comparison with Claude (Fable 5 / Opus 4.8)

To sum up: if you want to run an agent broadly and autonomously, GPT-5.6 Sol; if you want to fix a real codebase accurately or run long-form coding, Claude — that's the division of labor. On price, Claude Opus 4.8 ($5/$25) has a cheaper output price than Sol ($5/$30) and higher coding quality, so for coding use cases the Claude side's cost-effectiveness stands out.

8. Caveats — undisclosed benchmarks and coding weaknesses

1. Some benchmarks OpenAI hasn't disclosed

Independent analysis (Vellum) points out that this time OpenAI did not publish SWE-bench Verified, GPQA Diamond, AIME, MMLU, ARC-AGI-2, or FrontierMath. With the focus on agent-oriented metrics, the absence of direct comparison figures for general reasoning and math is worth keeping in mind when evaluating. You need to read the results while discounting the possibility that only the favorable numbers were lined up.

2. Even Sol is weak at production-grade coding

As noted above, on SWE-Bench Pro, Sol scored 64.6% versus Claude Fable 5's 80.0%. If your main goal is fixing bugs in production repositories or creating PRs, it's worth considering Claude over GPT-5.6.

3. Luna struggles with long-form processing

Luna is cheap and fast, but its long-context recall (MRCR) is a much lower 41.3%. Avoid handing entire long specs, logs, or codebases to Luna — leave long-form work to Terra or above.

4. Knowledge cutoff is February 16, 2026

Nothing after that date has been learned. Tasks that need the latest information assume you'll also use a web search tool.

9. Recommendations by use case — which model to choose

FAQ

Q1. When can I start using GPT-5.6?

General availability began July 9, 2026 (limited preview June 26). You can use it in ChatGPT, ChatGPT Work, Codex, and the OpenAI API.

Q2. How do Luna / Terra / Sol differ?

They're models arranged vertically along three axes: intelligence, speed, and price. Luna = cheapest and fastest; Terra = balanced (GPT-5.5-class at half price); Sol = top performance. Context (about 1 million tokens) and max output (128K) are shared across all three. We recommend making Terra your default and moving up to Sol or down to Luna as needed.

Q3. Can I use GPT-5.6 on the free plan?

You can't in regular chat, but within ChatGPT Work and Codex, even Free / Go users can use Terra. To freely switch between models or adjust reasoning effort, you need Plus ($20/mo) or above. The GPT-Live voice model's mini version is available even for free.

Q4. Between GPT-5.6 and Claude, which is stronger at coding?

It depends on the metric. On terminal operations (TerminalBench 2.1), Sol's 88.8% edges out Claude Fable 5's 86.0%, but on production-grade SWE-Bench Pro, Claude Fable 5 scored 80.0% and Sol 64.6%, with Claude far ahead. If your main goal is fixing bugs in production repositories, use Claude; for broad autonomous agents, use Sol — that's the realistic split.

Q5. What's the pricing?

Per 1M tokens: Sol $5/$30, Terra $2.50/$15, Luna $1/$6 (input/output). Terra is positioned to deliver performance equivalent to the previous-generation GPT-5.5 at roughly half the price, making it the centerpiece of cost planning.

Q6. What are the context window and knowledge cutoff?

All three models have a roughly 1-million-token (1,050,000) context and a 128K max output. The knowledge cutoff is February 16, 2026.

Q7. Are there new developer features?

Programmatic Tool Calling (where the model generates JavaScript to assemble tool calls) was added to the Responses API. Reasoning effort can be controlled across six levels: none / low / medium / high / xhigh / max.

Q8. What is GPT-Live?

It's a new voice model series announced alongside GPT-5.6. Rather than turn-based, it's full-duplex, enabling natural conversation including interruptions and interjections. Paid users get GPT-Live-1; free users get GPT-Live-1 mini.

Q9. Are there benchmarks OpenAI hasn't disclosed?

Yes. Independent analysis points out that SWE-bench Verified, GPQA Diamond, AIME, MMLU, ARC-AGI-2, FrontierMath, and others were not published this time. Since the focus is on agent-oriented metrics, general reasoning and math ability need to be evaluated separately.

Q10. Will my existing GPT apps keep working?

The API is largely compatible, and you can migrate by switching the model ID. That said, because it's now a three-model lineup, redesigning which tasks you assign to Luna/Terra/Sol lets you optimize cost and quality. In many cases, starting from Terra works well.

Conclusion: an era of "choosing vertically" with a three-model lineup

GPT-5.6 arrived not as a single flagship model but as a three-tier lineup of Luna / Terra / Sol. The roles are clear: Sol takes first place on overall agent ability and the coding-agent index, Terra becomes the workhorse of real-world use with its "GPT-5.5-class at half price" value, and Luna serves as the cost-effective choice for bulk processing.

At the same time, there are factors that keep it from being the unconditional best: it loses by a wide margin to Claude on production-grade coding (SWE-Bench Pro), OpenAI hasn't published direct comparison figures for general reasoning and math, and Luna is weak at long-form processing.

The smart way to operate in 2026 is, as ever, to "use GPT-5.6's three models and Claude selectively according to the task." Everyday work on Terra, complex agents on Sol, bulk processing on Luna, production coding on Claude — optimize cost and quality on the assumption of a multi-model setup.

Related articles

• GPT-5.5 Release: Complete Guide — details on the previous-generation GPT-5.5
• Claude Opus 4.8 Release: Complete Guide — the direct competing model
• Claude Fable 5 Release: Complete Guide — the model that leads on SWE-Bench Pro
• Claude vs ChatGPT Pricing Comparison — the plan structures of both