GPT-5.6 Sol, Terra, and Luna Go Public

GPT-5.6 Sol Ultra scored 91.9 percent on Terminal-Bench 2.1, and the model behind that number moved from a government-only preview to a public API in thirteen days.

From a two-week government preview to public API

OpenAI publicly released the GPT-5.6 family of models on Thursday, July 9, 2026, ending a limited-preview window that began on June 26, 2026. The company's own June-to-July sequence is what makes the launch procedurally unusual. That initial preview was restricted to vetted government partners at the request of the U.S. government, following a June 2026 executive order signed by President Donald Trump asking major AI developers to voluntarily submit leading models for government safety evaluations. Thirteen days later, the same three models became broadly available across web, mobile, desktop, and the API.

The lineup swaps the old numeric-only naming for a capability-tier system with three names: Sol, Terra, and Luna. Sol is the frontier model, described by OpenAI as tuned for long-horizon agentic work, coding, biology, chemistry, and cybersecurity workloads. Terra is positioned as the balanced everyday model, matching GPT-5.5 quality at roughly half the API cost. Luna is the fastest and cheapest tier in the family, aimed at high-volume production workloads. In its July 8, 2026 statement, OpenAI wrote: "GPT-5.6 Sol, along with Terra and Luna, will launch publicly this Thursday. We're expanding preview access globally now."

The benchmark numbers OpenAI led with

The headline result sits at the top of one benchmark. On Terminal-Bench 2.1, a benchmark testing command-line planning, iteration, and tool coordination, GPT-5.6 Sol Ultra scored 91.9 percent. The rest of the tier ladder follows in order. GPT-5.6 Sol scored 88.8 percent, edging past Claude Mythos 5 at 88.0 percent and topping GPT-5.5 at 88.0 percent. GPT-5.6 Terra scored 84.3 percent, matching its predecessor GPT-5.5 while costing half as much. GPT-5.6 Luna scored 82.5 percent, surpassing Anthropic's Claude Opus 4.8 at 78.9 percent.

The scientific evaluations sit in a different frame. On SecureBio's Virology Capabilities Test, GPT-5.6 Sol reached 53.5 percent, a top reported score in that evaluation. On the same lab's Molecular Biology benchmark, Sol reached 60.0 percent, and on its Human Pathogen Capabilities test, Sol reached 68.4 percent. On the World-Class Bio benchmark, GPT-5.6 Sol scored 68.3 percent, approximately 9 percentage points above GPT-5.5. OpenAI expects Sol to push past 40 percent on FrontierMath Tier 4, compared with GPT-5.5's 35.4 percent.

OpenAI paired those results with a safety line. In its July 9, 2026 announcement, the company said the models feature "strong safeguards that can withstand real-world adversarial pressure."

The quieter story: pricing and context window

The eye-catching benchmark is Ultra's 91.9 percent, but the business story is on the price sheet. API pricing for Sol is 5 dollars per million input tokens and 30 dollars per million output tokens, matching the GPT-5.5 launch price for the flagship tier. Terra costs 2.50 dollars per million input tokens and 15 dollars per million output tokens, exactly half the Sol rate. Luna costs 1 dollar per million input tokens and 6 dollars per million output tokens, the cheapest tier in the family.

That pricing pairs with a substantially larger context. The GPT-5.6 context window is approximately 1.5 million tokens, an increase of roughly 44 percent over the 400,000-token effective ceiling on GPT-5.5. The training data cutoff is approximately May 2026. And for customers who need the ceiling on inference throughput, Cerebras hardware deployment of GPT-5.6 will run at up to 750 tokens per second for select customers in July 2026.

Read together, the release targets a workload class that was previously borderline: long-running, tool-using agents that stay resident in production. Terra delivers GPT-5.5-level quality at half the API cost. Luna undercuts every previous OpenAI reasoning-capable model on a per-token basis. The larger window means longer traces, larger tool outputs, and more code stayed in scope. The older GPT-5.4 model will be retired on July 23, 2026, while GPT-5.5 models remain available for existing integrations.

The launch also ships new surfaces

Two of the release's product changes matter for how the models are meant to be used. OpenAI introduced a new "max reasoning effort" setting exclusive to Sol, allowing the model longer internal chains of thought before answering. It also introduced a new "Ultra" acceleration mode that dispatches multiple subagents in parallel across a single task, going beyond single-agent architecture. Ultra is the setting behind the 91.9 percent Terminal-Bench 2.1 score.

Alongside the models, OpenAI shipped a new ChatGPT Work agent for web, mobile, and desktop, and an upgraded ChatGPT desktop application that now integrates Codex directly on macOS and Windows. The GPT-5.6 family also launched alongside a rebranded real-time model called "GPT Live 1," formerly known as GPT-Bidi 1. OpenAI added "hosted sites" as a new feature for paid ChatGPT users, letting them publish output as standalone web pages. The rollout of all three models is happening over roughly 24 hours from the July 9, 2026 announcement.

The tension OpenAI itself flagged

The government evaluation program is the piece OpenAI addressed directly, and the company's language is worth quoting rather than summarizing. In a statement to Nextgov/FCW dated July 8, 2026, OpenAI said the pre-release government evaluation approach should not "become the long-term default." The company complied with the June 2026 executive order and submitted the models before release, then said publicly that the arrangement is not one it wants entrenched for individual customers.

That leaves the launch sitting on top of an unresolved question. Sol, Terra, and Luna cleared a government pre-release review and are now in the API. GPT-5.4 turns off on July 23, 2026. Customers get the price cut, the larger context window, and the new agentic surfaces immediately. What is not yet decided is who evaluates the next frontier release, how long that evaluation takes, and whether the June-to-July sequence becomes a template. OpenAI's public position is that it should not. What the next release looks like is what will show whether that position holds.

Sources

• OpenAI announcement, July 8 to 9, 2026
• Nextgov/FCW coverage of OpenAI statement, July 8, 2026