OpenAI GPT-5.6: Sol, Terra, and Luna Pricing, Benchmarks July 2026

OpenAI launched GPT-5.6 on June 26, 2026 as three distinct capability tiers: Sol, Terra, and Luna. Sol, the flagship, tops Terminal-Bench 2.1 at 91.9% and costs $5 per million input tokens. As of July 5, all three models remain restricted to roughly 20 vetted partner organizations, with general availability expected in mid-July. Claude Sonnet 5 shipped to all users the same day GPT-5.6 launched and is already available through every Claude plan.

Here is what you need to know before general access opens.

What Is the GPT-5.6 Model Family?

GPT-5.6 introduces a new naming system built around durable capability tiers rather than version suffixes. Sol, Terra, and Luna are not just cost tiers. OpenAI describes them as independent capability levels that can advance on their own release cadence. Sol is the highest capability tier, comparable to the old "Opus-class" framing. Terra is the balanced everyday model. Luna is the fast, cost-optimised option.

This naming change is significant for developers building on the API. GPT-5.6 Sol will stay Sol as the model improves, rather than being renamed GPT-5.7. The tier name becomes a stable reference point for capability expectations.

Sol, Terra, and Luna: What Each Tier Is For

Sol is the frontier model. It is what you reach for when quality is the only constraint, not latency or cost. The 91.9% score on Terminal-Bench 2.1 under the "Sol Ultra" configuration is the highest recorded for any model on that benchmark. Standard Sol runs at 88.8%. The gap between Sol Ultra and standard Sol suggests the Ultra configuration uses extended thinking or multi-pass reasoning, similar to how Claude's extended thinking mode changes output quality on hard tasks.

Terra is the balanced model. At $2.50 input and $15 output per million tokens, it sits in the same pricing neighbourhood as Claude Sonnet 5 and Gemini 2.5 Pro. For most business automation, coding assistance, and document processing, Terra is where the cost-quality tradeoff makes sense. OpenAI describes it as suitable for "everyday work."

Luna is the speed and cost tier. At $1 input and $6 output per million tokens, it competes directly with Claude Haiku 4.5 and Gemini 2.5 Flash. High-volume pipelines where latency and cost matter more than frontier reasoning belong on Luna.

GPT-5.6 Pricing vs Current Alternatives

Sonnet 5's introductory pricing at $2/$10 undercuts GPT-5.6 Terra at $2.50/$15 on both input and output. Until GPT-5.6 reaches general availability, Sonnet 5 is the practical choice for developers who need a capable model today at competitive pricing.

Benchmark Comparison: GPT-5.6 vs Claude and Gemini

GPT-5.6 Sol represents a meaningful improvement over GPT-5.5 on agentic coding tasks. The jump from 88.0% (GPT-5.5) to 88.8% (Sol) and 91.9% (Sol Ultra) on Terminal-Bench is the clearest signal that the model materially improves on the kinds of tasks developers care about most: autonomous code execution, tool use, and multi-step agent workflows.

The missing Claude Sonnet 5 data on Terminal-Bench is because the benchmark results were published before Sonnet 5 could be fully evaluated. That data will arrive as independent researchers run evaluations over the next few weeks.

Why GPT-5.6 Is Government-Gated

OpenAI shared GPT-5.6 Sol with the US government before announcing it publicly. General availability was delayed while OpenAI worked through security review with federal stakeholders. The pattern mirrors what happened with Anthropic's Fable 5 and Mythos 5, where the government imposed export controls after researchers found a jailbreak post-launch.

OpenAI avoided that outcome by coordinating with government before release. The tradeoff is a delayed general rollout for roughly 20 vetted partners to date, with a promised mid-July general availability window.

GPT-5.6 Sol will also run on Cerebras hardware at up to 750 tokens per second, which would make it the fastest frontier model available by token throughput. Cerebras access is launching in July for select customers as capacity expands.

What the Sol/Terra/Luna System Means for Developer Tooling

The new naming system creates a stable routing surface for agent frameworks and middleware. Instead of encoding specific version numbers into routing logic, teams can route by capability tier.

That said, OpenAI has changed its naming conventions multiple times. GPT-4 became GPT-4o became GPT-5 became GPT-5.5 became GPT-5.6, with capability suffixes and mini variants appearing and disappearing. The Sol/Terra/Luna framework is a meaningful step toward stability, but developers who have been burned by deprecation cycles should treat it as a commitment to watch rather than a guarantee.

Practically, this means: do not hardcode "Sol" as a permanent alias in critical production systems until the tier has been stable for at least two release cycles.

What to Do Before General Access Opens

If you are on the waitlist or have API access today, the priority tests are the ones that matter for your specific use case. Benchmarks like Terminal-Bench are useful signals but they measure a particular class of agentic coding tasks. Your pipeline may have different characteristics.

For teams currently on Claude Sonnet 5: the introductory pricing window is an opportunity to build cost baselines before GPT-5.6 Terra is available for comparison. When Terra ships, you will want real cost-per-task data, not benchmark extrapolations.

For teams on GPT-5.5 or GPT-4o: Sol's improvement on agentic tasks suggests it is worth evaluating for coding agents and tool-use workflows. The pricing is higher than 4o but the capability gap is significant.

Key Takeaways

• GPT-5.6 launched June 26, 2026 to approximately 20 vetted partners; general availability expected mid-July 2026
• Sol: $5/$30, Terra: $2.50/$15, Luna: $1/$6 per million input/output tokens
• Sol Ultra scores 91.9% on Terminal-Bench 2.1, the highest recorded for any model on that benchmark
• Claude Sonnet 5 at $2/$10 introductory pricing undercuts GPT-5.6 Terra and is available today across all Claude plans
• Sol/Terra/Luna are durable capability tiers, not version-specific model names, representing a shift toward stable API references
• For developers: build cost-per-task baselines on currently available models now; use them to evaluate GPT-5.6 Terra objectively when it ships rather than benchmarking from scratch
• What to watch: GPT-5.6 general availability date and Cerebras rollout for 750 tokens/second Sol access