Inside the Groundbreaking GPT-5.6 Launch: Models, Pricing, and Safety Restrictions

On June 26, OpenAI quietly released the next‑generation GPT-5.6 series, but access is currently limited to roughly 20 government‑approved partners, leaving ordinary developers and ChatGPT users unable to use the new models.

Unlike previous Pro/Mini naming, the release adopts a celestial theme: the numeric suffix denotes the generation while the suffixes Sol (Sun), Terra (Earth), and Luna (Moon) indicate capability tiers, forming a full‑family matrix. Sol (Sun): flagship tier representing OpenAI’s highest performance level. Terra (Earth): everyday tier, performance close to GPT‑5.5 but at half the price. Luna (Moon): economical tier focused on low cost and high throughput.

The API pricing reflects this strategy:

Sol: $30 per million input + output tokens.

Terra: $15 per million input + output tokens.

Luna: $6 per million input + output tokens.

For enterprises, Terra offers unbeatable cost‑effectiveness—performance remains strong while costs are halved—whereas Luna serves scenarios that demand extreme throughput and tight budgets.

Sol introduces two advanced modes:

Max mode: forces the model to spend more time on deep reasoning for complex problems.

Ultra mode: a “master‑builder” AI that spawns multiple sub‑agents to parallelize tasks, automatically decomposing, assigning, and aggregating results.

On the Terminal‑Bench 2.1 (command‑line workflow benchmark), Sol Ultra achieved a 91.9% high‑score, surpassing its own Sol baseline (88.8%), Claude Mythos 5 (88%), and Google Gemini 3.1 Pro Preview (70.7%). In the ExploitBench security benchmark, Sol reached Mythos Preview’s level using roughly one‑third of the tokens.

This performance suggests developers may soon rely on Sol Ultra to handle multi‑step tasks without building custom agent orchestration frameworks, echoing similar directions from Anthropic’s Claude agents and Cursor’s background agents.

OpenAI also announced a forthcoming Cerebras‑accelerated version of Sol in July, promising inference speeds of up to 750 tokens per second.

The primary reason for the throttled rollout is safety. OpenAI invested over 700,000 A100‑equivalent GPU hours in automated red‑team testing to discover cross‑scenario jailbreaks. The model includes a refusal mechanism and a real‑time classifier that pauses suspicious outputs for review by a larger model. According to OpenAI’s own assessment, Sol’s network‑security capability is rated “high” but not “critical,” meaning it can assist security experts in finding browser vulnerabilities and exploit primitives but cannot autonomously execute a full attack chain.

OpenAI frames this as a positive signal— the model helps defenders “find holes and patch them” rather than enabling attackers—yet the real‑world validity of this claim remains to be proven, which is why the 20 approved partners are tasked with testing in live environments.

For regular ChatGPT users, OpenAI has said a full public release is planned “in the coming weeks,” with media reports indicating more customers will be added next week, though no concrete timeline exists for the web or app interfaces.

Overall, GPT‑5.6’s launch is more than a model upgrade; it previews a new AI industry paradigm where powerful capabilities are paired with strict regulation, cost‑tiered offerings, and emerging “AI‑manages‑AI” architectures.