Skip to content
News Model release

OpenAI launches GPT-5.6 Sol, Terra, and Luna in limited preview with new model naming system and tiered pricing

· by Pondero Newsdesk

The short version

OpenAI began a limited preview of GPT-5.6 on June 26, 2026, shipping three models at distinct capability and price tiers, plus a new 'ultra' sub-agent mode on Sol that routes complex tasks across multiple agents.

OpenAI launches GPT-5.6 Sol, Terra, and Luna in limited preview with new model naming system and tiered pricing

OpenAI shipped GPT-5.6 on June 26, 2026, as a limited API and Codex preview. Three models arrived at once: Sol (flagship), Terra (balanced), and Luna (fast). The launch also introduced a new naming system where tier names (Sol, Terra, Luna) can advance independently of the version number, giving developers clearer signals about capability versus cost tradeoffs across future releases.

What shipped

Sol is priced at $5 input / $30 output per 1M tokens. Terra lands at $2.50 input / $15 output, with OpenAI claiming competitive performance to GPT-5.5 at 2x lower cost. Luna is $1 input / $6 output per the OpenAI announcement.

Sol carries two new inference options not present in prior models. A max reasoning effort setting gives the model more time to think through complex requests. An ultra mode routes work across multiple sub-agents in parallel, designed for long-horizon tasks that a single agent context struggles with.

On benchmarks cited by OpenAI, Sol sets a new result on Terminal-Bench 2.1 (command-line workflows requiring planning and tool coordination), improves on GeneBench v1 (genomics and quantitative biology) while using fewer tokens than GPT-5.5, and outperforms prior models on ExploitBench for long-horizon vulnerability research. Separately, OpenAI is launching Sol on Cerebras in July at up to 750 tokens per second, initially limited to select customers.

GPT-5.6 also introduces more predictable prompt caching: explicit cache breakpoints, a 30-minute minimum cache life, cache writes billed at 1.25x the uncached input rate, and a 90% discount on cache reads per the pricing section.

The government gating context

The preview is not public. Participation is limited to "a small group of trusted partners and organizations," and there is no public waitlist or application. That arrangement follows a request from the Trump administration, covered separately here on June 26. OpenAI reiterated its position in the launch post: "We don't believe this kind of government access process should become the long-term default."

OpenAI stated that GPT-5.6 Sol does not cross the Cyber Critical threshold under its Preparedness Framework. In browser-engine testing (Chromium and Firefox), Sol identified vulnerabilities and exploitation primitives but did not produce a functional full-chain exploit under the conditions tested per the system card linked in the announcement. A layered safeguard stack includes real-time classifiers that can pause output and route flagged content to a larger reasoning model for review before it reaches the user.

OpenAI dedicated over 700,000 A100-equivalent GPU hours to automated red-teaming ahead of this release, targeting universal jailbreaks rather than narrow attack patterns.

Why it matters

The tiered naming system has a concrete implication for buyers: Sol, Terra, and Luna are now positioned as durable capability bands that OpenAI can update on their own cadence. A Sol-class model in six months need not wait for GPT-5.7 to ship. That architecture makes pricing and capability roadmap commitments slightly more predictable, though OpenAI has not published a specific update schedule.

For operators currently on GPT-5.5, Terra is the migration candidate to watch. If the 2x cost reduction claim holds in production (OpenAI's own caveat: "real-world results may vary substantially"), Terra may serve the majority of agentic workloads at meaningfully lower spend than Sol.

The Cerebras partnership is the other operationally relevant signal. 750 tokens per second on a frontier-class model would change latency economics for real-time coding assistants and interactive agents. That is a preview claim, not a production commitment, but it names a specific speed target to hold OpenAI to when the July rollout begins.

What to watch next

OpenAI has said broad ChatGPT, Codex, and API access is "coming weeks" away. The pace of that rollout depends in part on how the U.S. government's cyber Executive Order framework develops. Watch for any formal announcement naming the initial ~20 trusted partner organizations, and for the Cerebras Sol speed to go live in July as a real benchmarkable data point.

Sources