Two powerful open-source AI models have rapidly filled the gap left by the US government's unprecedented ban on Anthropic's Fable 5 model, offering developers and companies frontier-level capabilities without the restrictions that now apply to closed systems. GLM-5.2, released by Z.ai, and Kimi K2.7 Code, from Moonshot AI, both launched this week and are now competing at the highest levels of AI performance benchmarks, while remaining freely available under MIT licenses.

Why Did the US Government Ban Fable 5?

The ban represents the first direct government intervention in the US AI space to restrict access to a frontier model. The US government issued an order preventing Anthropic from providing access to Fable 5 and its more powerful sibling, Mythos 5, to any foreign national. According to reports, the government received a tip, reportedly from Amazon, that Fable 5 could be jailbroken to perform cybersecurity tasks despite Anthropic's safety guardrails. Anthropic responded by pulling both models entirely, even from internal employee access, while negotiating with government officials on a safe release framework.

The decision highlights a tension in AI development: Anthropic had spent weeks building safety measures into Fable 5, yet when asked whether the model could be made completely secure against jailbreaks, the answer was no. This reality prompted government action, though it also sparked debate about the broader implications of restricting access based on nationality.

What Makes GLM-5.2 a Credible Replacement?

GLM-5.2 is a 753-billion-parameter open-weight model designed specifically for long-horizon coding tasks, where an AI system must work through complex engineering problems over extended periods without losing context. The model scored 74.4% on FrontierSWE, a benchmark measuring whether an agent can complete full engineering projects over hours, trailing only Opus 4.8 by about one percentage point and beating GPT-5.5. On Terminal-Bench 2.1, a specialized coding benchmark, it jumped to 81% from GLM-5.1's previous 63.5%.

What sets GLM-5.2 apart is its practical efficiency. The model supports a 1-million-token context window, roughly equivalent to processing 100,000 words at once, allowing developers to feed entire codebases, design documents, test logs, and migration plans into a single work session. Z.ai engineered this using a technique called IndexShare, which reduces computational cost by 2.9 times at full context length, making the massive model more affordable to run. Early users report approximately 8 times cost savings compared to Claude Opus 4.8 for comparable quality on real coding tasks.

How Does Kimi K2.7 Code Compare?

Moonshot AI's Kimi K2.7 Code also launched this week and is now available on CoreWeave Inference with notably fast speeds. Like GLM-5.2, Kimi K2.7 Code is MIT-licensed and available without regional restrictions, making it accessible to developers globally. The model has climbed into top rankings for open-source performance, positioning itself as another serious alternative for teams seeking frontier-level capabilities without depending on closed systems.

What Are the Practical Implications for Developers and Companies?

The emergence of these two models addresses a critical gap in the AI market. Before Fable's ban, many developers and companies had begun relying on closed frontier models from major labs. Now, with GLM-5.2 and Kimi K2.7 Code both available as open weights, teams have credible options for building AI systems they can control, deploy locally, and modify without depending on a single vendor or government policy.

• Cost Efficiency: GLM-5.2 delivers approximately 8 times cost savings versus Claude Opus 4.8 for comparable coding quality, making frontier-level AI more accessible to smaller teams and startups.
• Context and Continuity: The 1-million-token context window allows models to maintain coherence across large codebases, design documents, test failures, and tool traces without losing critical details from earlier in the conversation.
• No Regional Restrictions: Both GLM-5.2 and Kimi K2.7 Code are available globally under MIT licenses, eliminating the nationality-based access barriers that now apply to Fable 5 and other closed models.
• Deployment Flexibility: Open-weight models can be deployed on private infrastructure, giving companies control over data, latency, and customization without relying on external APIs.

How to Evaluate Open-Source Models for Your Team

• Test on Real Workloads: Benchmark GLM-5.2 and Kimi K2.7 Code against your actual coding tasks, not just public benchmarks, to understand how they perform on your specific codebase and tool stack.
• Assess Context Window Usability: Run long code-review sessions, multi-file refactors, and tool-heavy agent loops to verify that the 1-million-token context window actually helps your team maintain continuity across complex tasks.
• Calculate Total Cost of Ownership: Compare API pricing, inference latency, and deployment costs across hosted services like Z.ai and Fireworks, as well as the infrastructure costs of running the model locally.
• Verify Tool and Framework Support: Confirm that your existing development tools, version control systems, and CI/CD pipelines integrate smoothly with the model's API and function-calling capabilities.

The broader context matters here too. The Fable ban signals that governments are now willing to intervene directly in AI development, creating uncertainty for companies that depend on closed models. This shift has accelerated interest in open-source alternatives, particularly among enterprises concerned about supply-chain risk and regulatory exposure. For developers and CTOs, the message is clear: open-weight models are no longer just a cost-saving option; they are increasingly a strategic necessity for building resilient AI systems.

Z.ai's pricing for GLM-5.2 stands at $1.40 per million input tokens and $4.40 per million output tokens, with cached input tokens priced at $0.26 per million, making it competitive with other frontier models while offering the flexibility of open weights. Kimi K2.7 Code's availability on CoreWeave Inference suggests similar accessibility and pricing competitiveness.

The question now is whether these open-source models can sustain their momentum. GLM-5.2 and Kimi K2.7 Code have arrived at a moment when the AI market is fragmenting, with developers increasingly skeptical of single-vendor dependency. If these models continue to deliver on their benchmarks and remain freely available, they could reshape how teams build AI systems for the next several years.