GPT-5.6 Unveiled: Massive Power, Tiered Pricing, and Limited Access

Hello, I'm Ai Learning's Lao Zhang.

OpenAI has released GPT-5.6 in a limited preview, introducing a new naming scheme where the number (5.6) denotes the generation and the labels Sol, Terra, and Luna denote three permanent capability tiers.

Sol / Terra / Luna Explained

Sol : the flagship "most powerful model to date" for pushing the intelligence ceiling.

Terra : a balanced option with performance comparable to GPT‑5.5 but at roughly half the price.

Luna : a fast, cheap variant that delivers respectable ability at minimal cost.

Visually, the three tiers can be compared in the chart below:

Two New Reasoning Modes: max and ultra

max : gives Sol extra time for deep, hard‑problem reasoning.

ultra : schedules a group of sub‑agents to cooperate on complex tasks, effectively embedding multi‑agent orchestration into the model.

Ultra’s multi‑agent approach is notable because it packages what is usually a cumbersome coordination layer into a single model tier.

Where the Strength Lies

OpenAI showcases benchmark gains in three domains:

Programming : GPT‑5.6 Sol achieves a new SOTA on Terminal‑Bench 2.1 , a suite that evaluates planning, iteration, and tool‑use in command‑line workflows.

Biology : On GeneBench v1 , Sol not only outperforms 5.5 but does so with fewer tokens, a crucial cost‑saving for research teams.

Network Security : In ExploitBench , Sol matches the Mythos Preview while using only one‑third of the output tokens. Across ExploitGym , all three tiers improve security performance as reasoning intensity rises.

A hallucination‑rate chart shows GPT‑5.6 Sol (blue square) consistently lower than 5.5 across simulated latency levels.

Security: A Double‑Edged Sword

OpenAI positions GPT‑5.6 Sol as the "strongest security model to date," capable of vulnerability research and exploit building, yet it cannot autonomously execute a full attack chain and does not cross the "Cyber Critical" threshold in the Preparedness Framework.

"Sol is better at finding and fixing bugs than launching end‑to‑end attacks."

During testing on Chromium and Firefox, Sol generated bug‑finding components but failed to run a complete exploit chain.

Layered Safety Guardrails

OpenAI implements a "thousand‑layer cake" of safeguards because no single guard can stop determined adversaries.

Model layer : trained to refuse illicit network‑attack requests, even with disguised intent.

Realtime layer : security and biology classifiers monitor output; suspicious content triggers a generation pause and a larger model re‑examines the context.

Account layer : cross‑dialogue account‑level review distinguishes persistent malicious behavior from legitimate dual‑use research.

Differentiated access : the most sensitive capabilities are not exposed to all users.

During preview, users may experience false positives or slower responses for borderline dual‑use cases.

700,000 GPU‑Hour Red‑Team Effort

OpenAI invested over 700,000 A100‑equivalent GPU hours to run automated red‑team attacks aimed at discovering universal jailbreaks that work across prompts and scenarios.

In a CyberGym robustness test, universal jailbreak success dropped from 83% with no guardrails to 10% after the autoRT guard, and to 0% once the guard was fully applied.

Human experts will also conduct manual red‑team testing during the preview to capture creative attack vectors that automation may miss.

Pricing and Usage

Pricing per 1 M tokens (input + output):

Sol : 30 USD

Terra : 15 USD (a sweet spot with near‑5.5 performance at half the cost)

Luna : 6 USD (optimised for high‑volume, lower‑intelligence workloads)

Cache (prompt caching) upgrades include explicit cache breakpoints, a minimum 30‑minute cache TTL, and a 1.25× charge for cache writes while reads retain a 90% discount.

Speed and Availability

OpenAI announced that GPT‑5.6 Sol will run on Cerebras hardware in July, reaching up to 750 tokens/second, initially limited to a small set of trusted partners via API and Codex before broader rollout to ChatGPT and regular API users.

Overall, GPT‑5.6 represents a simultaneous advance in capability, safety, and commercial strategy; developers should watch Terra and the cache improvements, while security professionals should examine the new security performance and layered guardrails.