Sol, Terra, Luna

OpenAI previewed the GPT-5.6 series on June 26: Sol as the flagship, Terra for balanced everyday work, and Luna at the cheap-and-fast tier. The release isn't a single flagship spec; it's a three-tier portfolio launched under US government coordination, and the news is the packaging.

The pricing sets the tier structure. Sol runs at $5 per million input tokens and $30 per million output. Terra is $2.50 in and $15 out, with OpenAI saying it has competitive performance to GPT-5.5 at half the cost. Luna is $1 in and $6 out, putting frontier-lab quality at edge-deployment prices. OpenAI is explicit that the number 5.6 marks the generation, while Sol, Terra, and Luna are durable tier names that evolve on their own timeline going forward. That structure is itself a strategy commitment.

Capabilities came without the usual benchmark numbers. OpenAI's preview describes Sol as setting "a new state of the art on Terminal-Bench 2.1" and reports broad improvements on GeneBench v1, both without specific scores. The most concrete capability claim is ExploitBench, where Sol comes in competitive with Anthropic's Mythos Preview, the cyber-specific model gated to Project Glasswing partners, at roughly one-third the output tokens. Every other claim in the announcement cites a benchmark without giving the number. The benchmark-first launch framing got dropped.

Government coordination is the second move. OpenAI says it briefed the US government on the model's capabilities ahead of launch, and is starting with a limited rollout to a trusted-partners list cleared with the government before wider access in the coming weeks. OpenAI also says this kind of pre-release process shouldn't be the long-term default, framing it as a short-term step taken while it works with the Administration on the cyber Executive Order framework.

Anthropic has been operating under similar government coordination since June 12, when its Fable 5 and Mythos 5 went offline under an export directive and Project Glasswing partners got access to Mythos Preview under managed conditions. What's new is OpenAI making the coordination step a public part of the release cycle, not a post-hoc constraint. The release didn't happen and then get reviewed; the review happened and then the release was scoped.

Two practical changes follow. Pricing-tier strategy is now explicit across labs. Sol at $5 input is aggressive at the frontier, Terra at half the GPT-5.5 cost is real savings in the middle, and Luna at $1 input changes what's feasible to run at high volume. The lab-vs-lab decision now has to be made at the tier level. Teams pick across labs' Sol-equivalents for absolute best capability, across the Terra tier for balanced quality at scale, and across the Luna tier for high-volume cheap. Three model decisions where last quarter there was one.

The other change is timing. With government coordination on the release cycle, frontier model availability is no longer just a function of training completion. Teams planning around model availability dates need to factor in a coordination window between training-done and broad-access. For mission-critical deployments, that's a real schedule input.

The capability story will come at general availability. What shipped on June 26 was the tier architecture, the pricing ladder, and the coordination disclosure. The model release got more shapes attached to it on the way out the door.

What to Do With This

If you're picking AI tooling as an individual or a small team, switch the question from "which model is best" to "which tier of which lab matches my workload." Sol-class for one-shot frontier work, Terra-class for steady-state, Luna-class for high-volume cheap. Lab choice now happens inside each tier, not above them.

If you lead a team that uses AI across several workflows, your annual vendor review should compare apples-to-apples by tier rather than by flagship. The Sol vs Opus 4.8 vs Gemini 3.5 Pro comparison is one decision; the Terra vs Sonnet vs Flash comparison is another; they don't have to favor the same lab.

If you decide what AI tools the whole company runs, factor government coordination into your release-cycle expectations. The window between a lab announcing a model and broad availability is now longer than the technical-readiness gap alone implies. Build coordination delay into multi-quarter planning.

Want More Than This Newsletter?

Alcreon publishes a daily AI briefing, long-form dossiers, and an analysis feed for the teams actually shipping AI in production. This newsletter is one read out of the full library.

Read the daily feed or browse the editorials.

Also on the Radar

Anthropic's June Economic Index: Hourly Usage Data and a Counter-Intuitive Survey Finding

Anthropic published the June 2026 Economic Index on June 26, introducing hourly-resolution usage data and the first Anthropic Economic Index Survey. The arresting finding: people delegating the most work to Claude report the highest optimism about future pay and job security, contradicting the usual displacement-fear surveys. Computer and mathematical occupations are 30% of Claude survey respondents versus 4% of US employment.