OpenAI just previewed the GPT 5.6 model series — and it's a big one. The new family includes three models: Soul, Terra, and Luna. Soul is the flagship next-generation frontier model, Terra is the balanced efficiency play, and Luna is the fast, affordable option for high-volume tasks. If you've been wondering what GPT 5.6 is and whether it actually delivers, here's the full breakdown — plus everything going on with the Fable 5 ban, the Dario Amodei controversy, and China's push to match US AI in cybersecurity.
What Is GPT 5.6? Meet Soul, Terra, and Luna
OpenAI's GPT 5.6 is a three-tier model family designed to give developers and enterprises more flexibility depending on what they need. Here's how each model breaks down:
02:15
GPT 5.6 model lineup — Soul, Terra, and Luna pricing and positioning compared side by side
Watch at 02:15 →
- GPT 5.6 Soul — The flagship. A step-function improvement over GPT 5.5, priced at $5 per million input tokens and $30 per million output tokens. This is OpenAI's most capable model yet, with a new max reasoning mode and an ultra mode that uses multiple sub-agents to tackle complex tasks.
- GPT 5.6 Terra — The balanced option. Delivers performance competitive with GPT 5.5 at roughly 2x lower cost. Built for everyday enterprise workflows where raw power isn't always needed.
- GPT 5.6 Luna — The flash model. Fast, cheap, and built for high-volume tasks. Priced at just $1 per million input tokens and $6 per million output tokens — making it the most cost-efficient option in the lineup.
What's wild is that all three models reportedly ship with a 1.5 million token context window. That's enormous, and it opens up serious possibilities for long-form reasoning, large codebases, and complex document analysis.
OpenAI claims GPT 5.6 Soul sets a new state-of-the-art score on Terminal Bench 2.1 for agentic coding workflows — outperforming GPT 5.5, Claude Mythos 5, Claude Fable 5, and Gemini 3.1 Pro. They also report major gains in biology and cybersecurity, with the model achieving strong results while using fewer tokens than before. In some cases it does use more tokens on higher reasoning settings, but the outputs seem to justify it.
Why Did the US Government Ban Claude Fable 5?
This is where things get really interesting. Claude Fable 5 was taken offline roughly two weeks ago after it demonstrated the ability to identify and exploit vulnerabilities in US government security systems. That was a massive red flag, and the US government moved quickly to restrict access.
08:40
GPT 5.6 Soul generating a playable Minecraft-style world complete with terrain, mobs, and day/night cycle
Watch at 08:40 →
But here's the update: the Trump administration is reportedly close to allowing Anthropic to restore access to Fable 5. Insiders familiar with the discussions expect the restriction could be lifted as early as this week. Anthropic confirmed that since June 12th, it has been working closely with the government to restore access to both Mythos 5 and Fable 5. As a first step, around 100 organizations focused on critical infrastructure defense have already regained access to Mythos 5.
11:20
SpaceX Starship booster catch simulation built by GPT 5.6 Soul vs GPT 5.5 Pro side-by-side comparison
Watch at 11:20 →
The likely catch? Fable 5 will probably come back in a more restricted form — think tighter safety guardrails, reduced capabilities in high-risk cybersecurity areas, and stricter access controls. It won't be the same model we saw at launch, but it'll still be one of the most capable models available. We'll have to wait for the final terms between Anthropic and the government to know exactly what changes.
When Will GPT 5.6 Be Available to Everyone?
Right now, GPT 5.6 is in a limited preview for API and Codex users only, available to a small group of US government-approved trusted partners. The early launch — which happened just two days before this video — was unexpected, and it looks like national security concerns around frontier models accelerated OpenAI's timeline.
Broader availability is expected within two to three weeks. If you're a Codex user, you can check your backend analytics dashboard to see if you've been flagged into the early rollout of GPT 5.6 Soul. Otherwise, sit tight — it shouldn't be long before everyone gets access.
18:05
Anthropic enterprise market share chart showing Anthropic overtaking OpenAI in corporate AI spending
Watch at 18:05 →
What Can GPT 5.6 Soul Actually Do? Real Demos
The early demos are genuinely impressive. Here's a quick look at what GPT 5.6 Soul has been tested on:
- Minecraft clone — Soul generated a full playable Minecraft-style environment complete with terrain variety (including desert biomes), mob characters, cloud animations, a day/night cycle, crafting mechanics, and a working block-breaking animation. It built this in about 90 minutes. Some features like picking up blocks didn't work perfectly, but the overall output was striking.
- SpaceX Starship booster catch simulation — Directly compared against GPT 5.5 Pro, Soul produced an accurate, polished recreation of the Starship booster catch sequence — complete with mechanical clamps and realistic physics.
- Pokémon-style RPG — One-shotted from a short prompt in 31 minutes. The game included starter selection, eight gym badges, multiple gym leaders, and a full Pokémon League end-game system. Copyright restrictions meant it couldn't use actual Pokémon, but the structure was all there.
- Voxel 3D world generation — In Boxel, Soul produced high-quality 3D environments in about 44 minutes, showing strong spatial reasoning and aesthetic consistency.
The jump from GPT 5.5 is noticeable, especially in long-form coding, game generation, and front-end design. The model also appears to excel in cybersecurity and hard sciences like biology, where it reportedly performs competitively with Claude Mythos using about one-third of the output tokens.
Is Dario Amodei Fear-Mongering About AI Safety?
There's been a lot of noise on Twitter and YouTube blaming Anthropic CEO Dario Amodei for the Fable 5 ban — essentially claiming that his public warnings about AI safety convinced the US government to restrict frontier models. That narrative is probably wrong.
Think about it: the US government has the NSA, its own cybersecurity experts, scientific advisors, and intelligence agencies. They're not going to restrict billion-dollar AI systems that affect the entire AI race just because one CEO expressed concern. These are decisions with enormous economic and national security implications based on internal government assessments.
Where Anthropic can be fairly criticized is in how it reportedly communicated with the government during the early stages — specifically around responding to national security concerns in a timely and cooperative way. That's a legitimate critique. But it's very different from claiming Dario somehow convinced the government to kneecap US AI.
The more likely explanation is simple: these frontier models have reached a capability level where they pose a real cybersecurity risk — particularly the risk of China obtaining or reproducing these capabilities through model distillation or using them for offensive cyber operations against US infrastructure. That's a national security call, not a PR one.
Is China Catching Up in AI Cybersecurity?
Possibly. Zai, the company behind the GLM model series, is reporting that GLM 5.5 is coming very soon — and they're claiming a new Chinese model is already matching Claude Mythos at finding security vulnerabilities. If true, that's significant. Mythos has been considered one of the world's strongest models for cybersecurity and long-horizon vulnerability research.
The big question is whether this performance holds up in real-world security work or whether it's limited to specific benchmarks. But either way, it confirms that the gap between US and Chinese frontier AI labs is narrowing — and cybersecurity is now one of the most contested battlegrounds in the global AI race.
Why Is Anthropic Winning Enterprise AI Budgets?
Here's a chart worth paying attention to: by the end of 2025 into early 2026, Anthropic reportedly overtook OpenAI in US corporate AI spending by paid transactions. The reason? Anthropic went deep on software development use cases — an area where results are measurable, timelines are clear, and companies can directly see improvements in build times, bug rates, and code review cycles.
This doesn't mean OpenAI has lost enterprise. More likely, we're heading toward a multi-model world where companies use Anthropic for coding, Gemini for multimodal tasks, OpenAI for general productivity, and cheaper models for lower-value work. The next phase of the AI race is about workflow economics — which model reduces friction, which one can be governed at scale, and which one justifies its cost.
What Is Grok 4.5 and How Good Will It Be?
Elon Musk tweeted that Grok 4.5 has entered private beta at SpaceX and Tesla, with performance reportedly around the level of Claude Opus (likely Opus 4.8). The model is built on xAI's new 1.5 trillion parameter V9 foundational model, with additional training using Cursor data to improve coding capabilities. If those numbers hold up in real-world testing, Grok 4.5 could become another serious contender in the frontier AI race — and we should be seeing it fairly soon.
Between GPT 5.6 Soul, the potential return of Fable 5, China's GLM 5.5, and Grok 4.5 on the horizon, the next few weeks in AI are shaping up to be genuinely historic. Stay tuned — full benchmark testing on all of these models is coming as soon as broader access opens up.
