Safety ratings
High cyber. High bio. Below critical.
All three models carry the same ratings — High in cybersecurity, High in bio/chem, under the critical line. Independent evaluator SecureBio warned of real uplift for some actors.
0
Gov-approved partners first
0
Held in a review queue
0
Token context on Sol
0
Bigger context vs 5.5
Point 01 — the map
The names aren't marketing. They're a map.
OpenAI split the generation from the capability. The number — 5.6 — is the generation. The name is a durable tier that moves on its own clock. You don't pick a model anymore. You pick a tier per job.
Flagship
Sol
Frontier reasoning
$5 / $30
The one genuinely hard reasoning step in your pipeline. Adds Max Reasoning Effort + Ultra mode.
Terra
Balanced everyday
$2.50 / $15
RAG, chatbots, production APIs — roughly 5.5-class performance at a fraction of the flagship bill.
Luna
Fast & cheap
$1 / $6
Classification, intent routing, moderation, summarizing. Route this work here and cut your bill.
Find the flaw, describe the flaw, weaponize the flaw. That's a short road.
Source: OpenAI system card
Points 02 & 03 — the gate & the ceiling
Why a government stepped in — and what "good at cyber" really means
A model that can find a vulnerability is most of the way to a model that can write the exploit for it. That dual-use is the whole reason a review happened.
The review
Commerce Dept. ran CAISI
OpenAI flew technical staff to Washington to answer questions. Honest caveat: a White House official pushed back on the word "clearance." The gate was real; who held the key is murkier.
What it did
Multi-day recon, real PoCs
Per NeuralTrust's read of the system card: autonomous vuln-research campaigns, real proof-of-concept inputs, controlled exploitation primitives on a memory-safety bug the last model only crashed on.
The honest ceiling
Not a master burglar. A tireless junior researcher.
- It's speed at recon and flaw-finding — finds every open window on the building, fast, and hands you the list.
- The bottleneck OpenAI names is exploit-development judgment — the tricky window it still can't climb through.
- Anyone selling GPT-5.6 as a push-button hacker hasn't read the system card — or is selling you fear.
The most important sentence
It could not produce a working full-chain exploit against a hardened target. That distinction is everything — speed at recon, not autonomous end-to-end attack.
The orchestration you used to own now happens behind a single API call.
Point 04 — Ultra mode
Points 04 & 05 — the real story for builders
Ultra mode is the feature. The security bill is the catch.
Sol can spawn its own parallel subagents. You give it one complex task; it decides how to split the work, runs the pieces at once, returns one answer. Power with a meter running.
The absorbed layer
Research → draft → verify, no router
One task becomes separate sub-runs and you never wrote the decompose-spin-up-gather dance. For hand-built multi-agent pipelines, the model just absorbed a layer we used to write ourselves.
The cost tell
Every subagent burns tokens
Each subagent runs independently, so a single Ultra call can cost several times a standard Sol call. The convenience is real — so is the meter behind it.
The flip nobody covers
A tree of agents is a chain of trust
Per Augment Code's research: in an orchestrator-subagent setup, one compromised agent can propagate malicious instructions downstream. Hijack the orchestrator via injection and every subagent inherits it.
The gatekeeping, in order
Two weeks in a government waiting room
June 26, 2026
GPT-5.6 finished — and withheld
OpenAI completes Sol, Terra and Luna but ships only to ~20 government-approved partners in a restricted preview.
The review
Commerce Dept. CAISI evaluation
Sol described as shifting the performance-efficiency frontier for long-horizon security tasks. Technical staff answer questions in Washington.
July 8, 2026
Restrictions lifted
The administration clears the model. OpenAI argues this access process shouldn't become the default — a company arguing against its own launch strategy tells you the gate was real.
July 9, 2026
Broad global rollout
GPT-5.6 opens to everyone else — behind, not ahead of, the government-named partners.
OpenAI
VentureBeat
Axios
CNBC
METR
SecureBio
NeuralTrust
Augment Code
Commerce Dept · CAISI
The what is right here. The how is in the Vault.
The tier-routing playbook, the Ultra-mode subagent patterns, and the permission tiers that keep an orchestrator from taking your whole system down — that's inside the AI Founders Vault.