GPT-5.6 Public Launch: Sol, Terra, Luna, Pricing, Access

Four weeks under a government hold

OpenAI said it would publicly release the GPT-5.6 family, called Sol, Terra and Luna, on Thursday, July 9, 2026, roughly two weeks after it had limited the models' rollout to a small group of trusted partners at the request of the U.S. government. OpenAI had rolled the series out in June and agreed to restrict access to organizations whose participation was shared with the government, while saying it believed in broad access and would work to widen availability over the following weeks. [2]

The restriction traced to a Commerce Department directive in June that suspended access to frontier models on national-security grounds, and to a vetting arrangement OpenAI and Anthropic agreed to two weeks after that. The public GPT-5.6 launch followed by about two weeks Anthropic's restoration of access to its own Fable 5 and Mythos 5 models after the Commerce Department lifted a comparable export-control directive on those models. [2]

Atlas interpretation: CNBC frames the two companies' releases as parallel cases of the same new mechanism: an administration executive order from June asked model developers to voluntarily share their most advanced models with the government for capability assessment ahead of a full release, with federal agencies given 60 days to build an evaluation process. OpenAI's own framing, quoted by CNBC, was pointed about the cost of that arrangement: it said it did not believe such a government access process should become the long-term default, because it keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them. That OpenAI said this while still complying with the process is the tension the quote is doing work around. [2]

Three durable tiers, priced by capability

GPT-5.6 launched as three model tiers rather than one: Sol, the flagship; Terra, described as a balanced model for everyday work; and Luna, the most cost-efficient. OpenAI's announcement said the number identifies the generation while Sol, Terra and Luna are durable capability tiers meant to advance on their own cadence, rather than a single model line. At launch, API pricing per million tokens was $5 input and $30 output for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna. The rollout began globally the same day across ChatGPT, Codex and the API, with OpenAI saying it would reach full availability gradually over the following 24 hours. [1]

OpenAI's own benchmark claims centered on performance per dollar rather than performance alone. On Agents' Last Exam, an evaluation of long-running professional workflows across 55 fields, the company reported Sol scoring 53.6, 13.1 points above its stated comparison point, at roughly one-quarter the estimated cost at medium reasoning effort. On the Artificial Analysis Coding Agent Index, it reported Sol setting a new high of 80 while using less than half the output tokens, less time, and about a third less cost than that comparison point. OpenAI also introduced 'ultra,' a setting that coordinates four agents in parallel by default to trade higher token use for faster results on demanding tasks. [1]

Atlas interpretation: Every comparison in OpenAI's post is against its own prior model or a named competitor's model, evaluated on OpenAI's own harnesses, so the specific point gaps are a vendor's self-reported figures rather than independent measurement. The consistent shape of the claims, tokens down, latency down, cost down, at comparable or better scores, is the more durable part of the pitch: OpenAI is arguing that GPT-5.6's advance is efficiency as much as raw capability, which matters for a company whose customers were, within the month, pushing back on AI spending. [1]

Stronger cyber capability, gated by verification

OpenAI described GPT-5.6 as its strongest cybersecurity model yet, reporting large gains over its predecessor on exploit-development benchmarks it built internally, while stating the model does not cross what it calls the Critical capability threshold in cybersecurity or biology. It said its testing suggests GPT-5.6 is better at finding and fixing vulnerabilities than at reliably carrying out autonomous, end-to-end attacks against hardened targets, and that its GPT-5.6 Sol cyber safeguards block roughly ten times more potentially harmful activity than prior models. [1]

More of the model's defensive cyber capability, including vulnerability triage and validation, malware analysis, detection engineering and patch validation, is reserved for verified individuals and organizations through OpenAI's Trusted Access for Cyber program. OpenAI said individual members of that program would need to enable hardware-backed passkey security by September 1, 2026 to retain access to its most cyber-capable frontier models, with those who did not reverting to default access. [1]

A price cut once the cheap tier faced real competition

On July 30, 2026, three weeks after the public launch, OpenAI cut Luna's price by 80 percent, to $0.20 input and $1.20 output per million tokens, and cut Terra's by 20 percent, to $2 and $12. Sol's pricing did not change. OpenAI said in a release that its strategy remained focused on advancing both capability and efficiency so each generation of intelligence could accomplish more work at a lower cost. [3]

CNBC tied the cut to enterprise customers growing more cost-sensitive about AI spending after a period of unrestrained adoption, and to competition from Google, Microsoft, and Chinese developers touting cheaper models. It specifically noted Moonshot AI's open-weight Kimi K3, released days earlier and reported to outperform leading American models on some benchmarks, and Anthropic's response with, priced at half of Claude Fable 5 while performing comparably on coding and knowledge work. [3]

Atlas interpretation: The pattern of the cut, hitting the cheapest tier hardest and leaving the flagship untouched, matches where CNBC says the competitive pressure actually landed: Chinese open-weight models compete on cost at the low and mid tiers, not on the frontier capability Sol is priced against. That the reduction arrived within three weeks of launch, rather than as a planned part of the original pricing, reads as OpenAI reacting to a specific competitor's release rather than following a pre-set cadence. [3]