GPT-5.6 Explained: Sol, Terra and Luna Models, Pricing and Benchmarks

On July 9, 2026, OpenAI publicly launched GPT-5.6, its new flagship model generation, after a two-week preview limited to vetted API and Codex partners that started on June 26. The release replaces the familiar single-model launch with a family of three tiers, sol, Terra and Luna. The naming change matters, the number (5.6) marks the generation, while Sol, Terra and Luna are durable capability tiers meant to persist and improve across future generations. Sol is the frontier flagship for the hardest reasoning and agentic work, Terra is the balanced everyday model, and Luna is the fast, lowest-cost option for high-volume traffic. In this guide we break down what each tier is for, the pricing, the benchmark numbers published so far, the new agentic features like Ultra mode with subagents, and what all of this means if you are building software or AI agents on top of these models.

Sol, Terra and Luna, the Three Tiers

GPT-5.6 ships as three models that share a generation but target very different workloads:

• GPT-5.6 Sol (flagship), aimed at hard reasoning, agentic coding, science, cybersecurity and complex multi-step workflows. It adds a max reasoning effort setting and an Ultra mode that orchestrates subagents for long-horizon tasks. Priced at $5 per million input tokens and $30 per million output tokens.
• GPT-5.6 Terra (balanced). The default for everyday professional work and the routing target for most traffic. OpenAI positions it as similar in capability to GPT-5.5 at $2.50 per million input tokens and $15 per million output tokens.
• GPT-5.6 Luna (fast and cheap), built for high-volume tasks, drafts, extraction, classification and latency-sensitive traffic, at $1 per million input tokens and $6 per million output tokens.

Benchmarks, what Has Been Published So Far

OpenAI frames the GPT-5.6 gains around long-horizon agentic work, stronger coding, plus notably better results in genomics, quantitative biology and cybersecurity tasks. The headline published numbers come from Terminal-Bench 2.1, which measures command-line agent workflows:

Two caveats worth knowing before you draw conclusions. First, OpenAI has not published SWE-Bench Pro results for Sol at launch, which is the benchmark most teams use to compare real-world agentic coding. Second, the widely repeated claim of a 1.5 million token context window appears only on third-party aggregator blogs and has not been confirmed on any official OpenAI page, so treat it as unverified.

• GPT-5.6 Sol Ultra: 91.9% on Terminal-Bench 2.1 (Sol running in Ultra mode with subagents)
• GPT-5.6 Sol: 88.8% on Terminal-Bench 2.1 (flagship tier, standard mode)
• GPT-5.6 Terra: 84.3% on Terminal-Bench 2.1 (balanced tier)
• GPT-5.6 Luna: 84.3% on Terminal-Bench 2.1 (fast tier, matches Terra on this benchmark)

Pricing and Prompt Caching

The pricing ladder is aggressive, especially at the bottom. Luna undercuts most frontier-tier competitors by a wide margin. On top of the base rates, OpenAI applies its usual caching economics:

• Base rates per million tokens, sol: $5 input / $30 output. Terra: $2.50 input / $15 output. Luna: $1 input / $6 output.
• Cache writes at 1.25x, writing a prompt into the cache costs 1.25 times the uncached input rate, so caching pays off from the second hit onwards.
• 90% discount on cached reads, cached input tokens are billed at a 90% discount, which is decisive for agents and assistants that reuse long system prompts.
• 30-minute minimum cache life, cache entries live at least 30 minutes with explicit breakpoints, enough for most interactive sessions and batch pipelines.

Availability and the July 2026 Timeline

GPT-5.6 lands in the middle of a busy month for OpenAI. The sequence so far, on June 12 the GPT-5.2 models were retired from ChatGPT, with conversations migrated to GPT-5.5. On June 26 GPT-5.6 entered limited preview for selected API and Codex partners. On July 8 OpenAI launched GPT-Live, a full-duplex voice model that replaces the default ChatGPT Voice experience. And on July 9 GPT-5.6 reached public general availability.

During the preview phase GPT-5.6 was API and Codex only, with no ChatGPT access and no public signup. With the July 9 launch, availability expands globally, so expect the usual staged rollout across ChatGPT tiers and API accounts over the following days.

What GPT-5.6 Changes in Practice

Beyond the benchmark race, three things stand out for teams that ship software on top of these models:

• Tiering is now the default architecture, with Sol, Terra and Luna sharing one generation, OpenAI is telling you to route, cheap traffic to Luna, everyday work to Terra, and only the hardest problems to Sol. Single-model architectures leave money on the table.
• Agentic features move into the platform, ultra mode with subagents, computer use and command-line workflows are now first-class product features, not community hacks. Long-running autonomous agents keep getting easier to build and harder to differentiate on plumbing alone.
• Durable tier names simplify upgrades, because Sol, Terra and Luna are meant to persist across generations, future migrations should be closer to a version bump than a re-architecture, as long as your code treats the model as a configurable dependency.

How We Evaluate New Models at Alher Tech

Every model launch arrives wrapped in marketing numbers. We ship AI agents and automations into production, so our job is to separate what a release actually changes from what it merely announces.

If you are deciding whether GPT-5.6 belongs in your stack, we can run that evaluation on your real use cases and design the architecture around the answer.

• Evals on your real tasks, we benchmark new models like GPT-5.6 head-to-head against incumbents on each client's actual prompts and workflows before recommending a switch. Launch-day benchmarks rarely predict your workload.
• Smart routing across tiers, the Sol, Terra and Luna split maps directly to the routing architectures we build, the cheapest model that passes the eval handles each request, and only the hardest traffic pays flagship prices.
• No single-vendor bets. The Claude Fable 5 suspension in June showed what happens to products hardcoded to one model. We design multi-provider architectures with fallbacks so a vendor incident never becomes your outage.
• Guardrails before autonomy, features like Ultra mode with subagents are powerful, but production agents need sandboxed tools, spend caps and human approval on irreversible actions before they run unattended.

Conclusion

GPT-5.6 is less about one big model and more about a pricing and capability ladder, sol for the hardest agentic work, Terra for the everyday, Luna for volume. The published Terminal-Bench numbers are strong, but the missing SWE-Bench Pro results and the unconfirmed context window mean the full picture will take a few weeks to emerge.

For businesses, the practical takeaway is architectural, model tiering and routing are now the default way to consume frontier AI, and teams that treat models as swappable, evaluated dependencies will capture the price drops without the migration pain.

If you want help evaluating GPT-5.6 against your current stack, or you are planning an AI agent or automation on top of it, we do exactly that for our clients.

Related articles

• GPT-5.6 vs Claude Fable 5, benchmarks, Pricing and Which One to Use
• What GPT-5.6 Means for Your Business, aI Agents, Costs and Use Cases
• Claude Opus 4.8, what's New, Benchmarks, Pricing & How to Use It