GPT-5.6 Three-Model Launch: Soul/Terra/Luna Tiered Strategy and the AI National Security Game

The GPT-5.6 Family Arrives: Soul, Terra, and Luna Three-Line Strategy

OpenAI has given an early preview of its next-generation GPT-5.6 model series, launching three models with distinct positioning all at once: Soul, Terra, and Luna. This release was originally expected in June, then reportedly pushed back to July, only to quietly drop two days ahead of schedule — a rather intriguing cadence.

Looking at the product line, the three models form a clear tiered strategy:

• GPT-5.6 Soul: The flagship frontier model, a major upgrade over GPT-5.5, priced the same as its predecessor — $5 per million input tokens and $30 per million output tokens.
• GPT-5.6 Terra: Focused on balance and efficiency, claiming to deliver performance close to GPT-5.5 at roughly half the cost, making it suitable for high-frequency daily work scenarios.
• GPT-5.6 Luna: A lightning-fast lightweight model aimed at high-frequency tasks, with the lowest price — just $1 per million input tokens and $6 for output.

OpenAI's three-line strategy reflects the typical evolution of product strategy as large model commercialization enters maturity. The early AI model market was dominated by "single flagship" offerings, but as enterprise customer needs diverged, the "Model Family" strategy gradually became mainstream — similar to NVIDIA's H100/A100/L40S tiering in the chip industry, or AWS's instance type matrix in cloud computing. The core logic of this strategy is the engineering implementation of Price Discrimination: by segmenting users along a capability gradient into research users willing to pay premium prices for top-tier performance, developers seeking cost-effectiveness, and batch task operators who only need high-frequency low-complexity processing, revenue coverage is maximized. GPT-5.6 Luna's pricing ($1 per million input tokens) already approaches the deployment cost range of open-source models, indicating that OpenAI is actively compressing the price advantage of open-source alternatives.

On Token Pricing: A token is the basic unit large language models use to process text. Typically, one English word is roughly 1–1.5 tokens, and a Chinese character is about 1–2 tokens. Per-million-token pricing has become the industry-standard billing method, making it easier for developers to estimate API call costs. This tiered pricing strategy allows enterprises to select the appropriate model based on task complexity — Soul for high-precision tasks, Luna for high-frequency low-complexity tasks — thereby achieving fine-grained management between cost and performance.

Even more noteworthy is that the entire series is equipped with a massive context window of up to 1.5 million tokens. Achieving this breakthrough required overcoming multiple technical bottlenecks: the standard Transformer's attention mechanism has a computational complexity of O(n²), meaning that doubling sequence length leads to a quadrupling of computation and memory usage — the fundamental reason early GPT models were forced to stay within 4,096 tokens. Breaking through this limit relies on the coordinated advancement of three technical approaches: first, Sparse Attention mechanisms, such as the FlashAttention family of algorithms that process attention computation in blocks, reducing memory usage from O(n²) to O(n); second, improvements in positional encoding, where techniques like RoPE (Rotary Position Embedding) allow the model's positional awareness to extrapolate beyond its training length; and third, compression and quantization of the KV Cache, reducing memory pressure by lowering the precision requirements of cached data during long-sequence inference. 1.5 million tokens means the model can process roughly one million words of text in a single conversation — equivalent to dozens of books, or about 12 million Chinese characters — enough to accommodate a one-shot full-text analysis of an entire codebase or complete legal contract set. By comparison, the early GPT-3.5 only supported 4,096 tokens; in just two years, context capacity has expanded nearly 200-fold, reflecting the Transformer architecture's continuous breakthroughs in attention mechanism optimization and memory management. This gives developers the flexibility to trade off among intelligence, speed, and cost.

Real-World Performance: Significant Leaps in Coding and Simulation

Due to access restrictions, comprehensive benchmarks cannot yet be run, but some Codex users have obtained preview access, and early feedback has been quite impressive.

In several representative demos, GPT-5.6 Soul demonstrated powerful capabilities in generating long-form code and complete games:

• Minecraft clone: Generated in about 90 minutes, including a main landing page, game mode selection, block-breaking animations, desert terrain, monsters, a crafting system, and day-night cycles. While the polish didn't fully match GPT-5.5 Pro, the overall result felt "more alive."
• SpaceX Starship booster recovery simulation: In a direct comparison with GPT-5.5 Pro, Soul reproduced the booster recovery process accurately and realistically.
• Pokémon-style game: In just 31 minutes with an extremely short prompt, it generated a complete simulator featuring 8 gyms, badge collection, starter partner selection, and a late-game league system.

A Larger Reasoning Budget Brings Qualitative Change

OpenAI claims that GPT-5.6 sets a new SOTA on the Agentic Coding Workflow evaluation of Terminal Bench 2.1, clearly surpassing GPT-5.5, Claude Mythos 5, Fable 5, and Gemini 3.1 Pro.

The limitation of traditional code generation benchmarks (such as HumanEval and MBPP) is that they mainly test the model's completion ability at the single-function level, which differs significantly from real-world software engineering scenarios. The Agentic Coding Workflow evaluation system represented by Terminal Bench 2.1 was designed precisely to fill this gap: the system simulates real developer workflows, requiring the model to autonomously complete a full closed loop of requirement analysis, technology selection, code implementation, dependency installation, test execution, and error fixing within a sandboxed Linux terminal environment. Scoring considers not only the correctness of the final code but also tracks token consumption efficiency and the number of task-completion rounds. "SOTA" (State of the Art) refers to the best performance level in current public evaluations; setting a new SOTA means surpassing all publicly released competitors on that test. The rise of this evaluation paradigm reflects an upgrade in capability expectations for AI applications — from "code completion assistant" to "autonomous software engineer."

You may not have noticed, but Soul's token efficiency in standard mode is not superior to GPT-5.5. However, when the reasoning level is set to Max Reasoning mode, it actively consumes more tokens in exchange for better output — the reasoning budget increases from around 768 to nearly 1,000, and the model performs deeper internal reasoning before finalizing.

The engineering realization of the "Reasoning Budget" concept can be traced back to the Chain-of-Thought Prompting paper released by Google Research in 2022. That study first systematically demonstrated that guiding models to generate intermediate reasoning steps could substantially improve accuracy on math and logic tasks. Subsequently, OpenAI's o1 series pushed this idea to a new stage: the model no longer relies on prompts to trigger reasoning but instead internalizes the "think first, answer later" behavior pattern at the training level, optimizing the quality of reasoning chains through Reinforcement Learning (RL). Traditional models generate output directly from input, whereas reasoning-capable models first generate a series of internal "thinking steps" before giving a final answer — these steps consume extra tokens but significantly improve accuracy on complex reasoning tasks. The adjustability of the reasoning budget (from 768 to 1,000) embodies a dynamic computational resource allocation mechanism — similar to how humans switch between "rough estimation" and "precise derivation" when solving problems — with particularly notable advantages on tasks requiring multi-step verification, such as code debugging, mathematical proofs, and security vulnerability analysis.

There is also an Ultra mode that can invoke multiple sub-agents to collaboratively handle complex tasks. In this Multi-Agent System architecture, a main coordinating Agent (Orchestrator) breaks down complex tasks into subtasks, distributes them to multiple specialized sub-agents for parallel processing, and finally aggregates the results. In large-scale software engineering scenarios, the main agent can simultaneously dispatch dedicated sub-agents responsible for frontend, backend, database design, and testing, significantly reducing the time to complete complex projects — marking the transition of AI applications from "conversational tools" to "autonomous work systems."

Soul's improvements in cybersecurity and hard sciences (especially biology) are particularly prominent. OpenAI claims it performs comparably to Mythos in advanced vulnerability research while consuming only about one-third of the output tokens. Overall, Soul still slightly trails Fable 5, but the gap has become very small.

The Fable 5 Delisting Controversy: A National Security Game

The most thought-provoking backdrop to this release is the delisting and negotiation controversy surrounding Anthropic's frontier models Fable 5 (and Mythos 5) — which also directly explains why GPT-5.6 is currently only in limited preview.

Fable 5 was alleged to have already acquired the capability to breach a large number of U.S. government security systems, constituting a serious national security warning signal, and was therefore halted for about two weeks. Currently, GPT-5.6 is likewise only opened to a small group of trusted partners approved by the U.S. government for API and Codex access, with broader availability expected to take several more weeks.

The U.S. system for assessing the dangerous capabilities of frontier AI models can, at the institutional level, be traced back to the AI Executive Order (EO 14110) issued by the Biden administration in October 2023, which required that models trained with computing power exceeding 10²⁶ FLOP report their safety testing results to the government. The evaluation framework is typically jointly implemented by agencies such as CISA (Cybersecurity and Infrastructure Security Agency), the NSA, and NIST, with Red Teaming as the core tool — professional security researchers simulate malicious actors and attempt to induce the model to provide genuinely usable attack plans. The key criterion for distinguishing "theoretical risk" from "substantial risk" is usually whether the model's output significantly lowers the technical barrier to carrying out an attack (i.e., the "Uplift" metric). The assessment of a model's "dangerous capabilities" mainly focuses on three dimensions: assistance in cyberweapon development, guidance in biochemical weapon synthesis, and the ability to deceive or evade human oversight. The Fable 5 halt shows that the U.S. has established a fairly sensitive triggering mechanism, representing an important shift in AI governance from reactive response to proactive prevention.

Gradual Lifting Rather Than Full Release

According to the latest reports, the Trump administration is reportedly about to allow Anthropic to restore access to Fable 5, with restrictions possibly lifted as early as the near future. Anthropic officially confirmed that it has been in close negotiation with the government since June 12: the government has approved the redeployment of Mythos 5 to a small group of U.S. agencies responsible for critical infrastructure, with about 100 agencies having regained access.

This indicates that the Department of Commerce is adopting a strategy of gradual relaxation rather than full lifting. Even if Fable 5 returns, it will most likely come with stricter security safeguards, tighter access controls, and targeted weakening of capabilities in certain high-risk cybersecurity domains.

Is the CEO Manufacturing Panic, or Is the Government Making Independent Decisions?

Around Anthropic CEO Dario Amodei, a narrative has circulated online: that it was precisely his "manufacturing of panic" that led to restrictions on Fable 5 and future U.S. models.

This interpretation is debatable. The U.S. government has the NSA, intelligence agencies, cybersecurity experts, and scientific advisors, and is unlikely to act against a frontier model worth billions of dollars and influential to the entire AI race landscape simply because of one CEO's statements. These decisions are almost certainly based on independent assessments within the government.

The more likely truth is: the new generation of frontier AI models has already reached a level that poses a substantial risk to cybersecurity — once they fall into the wrong hands, or are replicated and spread via model distillation, they would directly threaten the security of national infrastructure. Model distillation was first systematized by Hinton et al. in 2015, and its core principle lies in the fact that a large model's (the teacher model's) output is not just a final label, but a probability distribution over all possible outputs (soft labels). These probability distributions carry information about inter-class relationships learned by the large model, and the small model, by minimizing the KL divergence with the soft labels, can achieve performance close to the teacher model at a cost far below the original training cost. In the era of large language models, distillation takes even more diverse forms: beyond traditional knowledge distillation, it also includes supervised fine-tuning (SFT) using conversational data generated by frontier models — with an extremely low barrier to entry, as DeepSeek-R1's success has already demonstrated the effectiveness of this path. In the AI safety context, distillation technology creates a severe proliferation dilemma: even if the original frontier model is under strict access control, any third party with sufficient interaction permissions can theoretically distill a derivative model with similar dangerous capabilities through systematic sampling — this poses a fundamental challenge to the strategy of keeping frontier capabilities "in the box," which is the core driver of regulatory intervention.

GLM 5.5 Closing In on Claude Mythos? The U.S.-China AI Race Heats Up

Zhipu AI has released news that GLM 5.5 is coming soon, showcasing a new Chinese model that reportedly matches Claude Mythos in security vulnerability discovery capabilities.

If true, this would be an important signal. Mythos has long been regarded as one of the world's strongest models for cybersecurity and long-horizon vulnerability research. If GLM 5.5 can genuinely compete with it on real-world security tasks, it would indicate that Chinese AI labs have not slowed their pace on this critical battlefield.

Of course, one core question remains to be verified: does the model truly possess combat-level security capabilities, or is it only effective in specific scenarios? Regardless of the answer, this confirms that the current U.S.-China AI race has deeply pivoted toward cybersecurity — the most sensitive and strategically valuable dimension.

A New Dimension of Competition: From Model Intelligence to Workflow Economics

Anthropic's evolving business model is equally worth watching. In enterprise and business scenarios, Anthropic's usage continues to surge, and it has reportedly surpassed OpenAI in U.S. enterprise AI spending driven by paid transactions.

This reveals a profound shift: The AI battle is no longer just about whose model is smarter, but about whose model can genuinely embed into an enterprise's daily workflows. The lock-in effect of enterprise-grade AI workflows belongs to the economic category of "Switching Cost," but its complexity far exceeds that of traditional SaaS products. When an AI system becomes deeply embedded in enterprise operations, switching barriers arise across multiple dimensions: at the data level, accumulated conversation history, custom prompt libraries, and fine-tuning datasets are hard to migrate; at the process level, approval workflows and quality standards built around specific model behavior patterns need recalibration; at the personnel level, the team's accumulated familiarity with a specific model's "habits" is hard to quantify but genuinely exists. This is similar to the logic by which Windows and Office became the default office environment in the 1990s — once an AI system is deeply embedded in an enterprise's CI/CD pipeline, code review process, knowledge base management, and employee training system, the hidden cost of switching vendors often far exceeds the explicit difference in subscription fees, forming a powerful platform lock-in effect.

MCP (Model Context Protocol) is an open protocol introduced by Anthropic in 2024, designed to standardize the interface specifications between AI models and external tools and data sources — akin to a USB standard for the AI domain. By opening up the MCP ecosystem, Anthropic seeks to position Claude as the "hub" of enterprise tool chains rather than a "plugin." Anthropic builds workflow stickiness through Claude.ai's Projects feature, custom system prompts, and the MCP tool chain; OpenAI is positioning itself on the same battlefield through GPT Actions and the Assistants API. Whoever first becomes the enterprise's "default AI environment" holds the most enduring advantage in this race. While alternatives exist, the platform that controls "how work actually happens" is the ultimate winner.

Anthropic's ability to gain a foothold largely stems from its focus on software development, an ideal testing ground: workflows are already highly digitized, improvements in build time, bug count, and review cycles are clearly measurable, and the barrier to value validation is low.

However, this does not mean OpenAI is out of the enterprise market. What's more likely to emerge is a multi-model division of labor landscape:

• Anthropic focuses on coding
• Gemini handles multimodal
• OpenAI covers general productivity
• Cheap models take on low-value batch tasks

The next stage of the AI race is "workflow economics" — which system has less friction, is easier to manage, and delivers measurable business value. As for the coding track, with GPT-5.6's strong performance, OpenAI is clearly making a comeback.

Grok 4.5 Waiting in the Wings

Elon Musk revealed that Grok 4.5 has entered internal testing at SpaceX and Tesla, with performance reportedly reaching Opus-level. The model is based on xAI's new 1.5-trillion-parameter V9 foundation model and has strengthened its coding capabilities using Cursor data, potentially becoming another powerful variable in the frontier competitive landscape.

The 1.5-trillion-parameter scale places Grok 4.5 in the tier of the largest known language models. In current public information, GPT-4's parameter count is estimated at around 1.8 trillion (mixture-of-experts architecture), while the flagship version of the Llama 3 series has around 405 billion parameters. While ultra-large parameter scales bring capability improvements, they also mean exponential increases in inference cost. To cope with this cost, xAI is most likely adopting a Mixture of Experts (MoE) architecture: the model is divided into multiple expert sub-networks, and only a small subset is activated per inference, keeping actual computation far below the theoretical upper limit of the total parameter count. Notably, the detail that Grok 4.5 strengthens its coding capabilities using Cursor data reveals the competitive value of the data flywheel effect in the coding track — as a leading AI coding IDE, Cursor's user interaction data contains dense signals of model successes and failures in real development scenarios, making it a scarce asset for training high-quality coding models. If it can be made available at a competitive price, it will pose a direct impact on the existing landscape.

Conclusion

GPT-5.6's early preview coincides with a sensitive juncture where the U.S.-China AI race and national security regulation are deeply intertwined. From the tiered strategy of three models, to the delisting game around Fable 5, to GLM 5.5's catch-up and the rise of workflow economics, this series of developments collectively sketches a new landscape for the AI industry: technological competition, commercial deployment, and geopolitical security have become deeply bound together, and no single dimension can fully explain the current trajectory.

Key Takeaways