Item: OpenAI
Author: Karl Strauchman

OpenAI unveiled GPT-5.6 on June 26 as a three-tier family, Sol, Terra, Luna, and for the first time in the company's flagship history, you can't buy any of it. Per Axios, the initial cohort is roughly 20 organizations "whose participation has been shared with the government," with broader release expected in the coming weeks.

The gating traces to a June 2 executive order establishing a voluntary 30-day government preview window for "covered frontier models." Reuters reports OpenAI presented capabilities to the administration ahead of launch. In its own post, the company is careful to call the arrangement a "short-term step" and explicitly states it doesn't believe "this kind of government access process should become the long-term default." That's a sentence written for a future hearing transcript.

The tier reshuffle is the cleaner story. Sol holds the line at GPT-5.5's headline pricing, $5 per million input tokens, $30 per million output, while Terra delivers what OpenAI calls "competitive performance to GPT-5.5 while being 2x cheaper" at $2.50 and $15, and Luna brings "strong capability at our lowest cost" at $1 and $6. Sources told VentureBeat the Sol/Terra/Luna naming is meant to retire the "nano" and "mini" labels; tiers reflect use cases, not raw size. OpenAI's own framing: "The number identifies a model's generation, while Sol, Terra, and Luna identify durable capability tiers that can advance on their own cadence."

Sol gets the new toys. TechCrunch confirms a "max" reasoning-effort setting and an "ultra" mode that coordinates sub-agents on complex tasks. A Cerebras deployment lands in July at up to 750 tokens per second, and per VentureBeat, Sol is also launching on Amazon Bedrock, the first model in the series to ship day-one on a competing cloud. Prompt caching has been reworked around explicit cache breakpoints, a 30-minute minimum cache life, cache writes billed at 1.25x the uncached input rate, and the 90% discount preserved on reads.

Then there's the system card, which is where the rollout politics start to make sense. All three models are classified "High" for both cyber and biological/chemical capability under OpenAI's Preparedness Framework, though below threshold for AI self-improvement. OpenAI tells Axios that Sol is "better at helping people find and fix vulnerabilities than reliably carrying out end-to-end attacks," and that its capabilities don't reach the "critical" level.

The August executive-order milestone is the one to watch: a classified process to designate "covered frontier models" by cyber capability. The Bush-era export-control architecture for cryptography in the 1990s took roughly that shape, classified thresholds, voluntary industry preview, then statutory teeth. GPT-5.6 is the first model launched inside that scaffolding, and OpenAI is on record asking that it not be the template.