Fable, GLM and Hy4 landed in the same week. The test that matters is the one you run yourself

Fable 5.2

Anthropic released Claude Fable 5.2 on November 9, 2026, and the headline numbers are about scale and speed. Claude Fable 5.2 supports a context window of up to 1 million tokens and a response length of up to 128,000 tokens, and generates output up to 30 percent faster than Fable 5.1. Claude Fable 5.2 replaces Claude Fable 5.1, and offers five effort levels for controlling reasoning depth. m:…2663

The price moved too. Claude Fable 5.2 costs $8 per million input tokens and $40 per million output tokens, and pricing represents a 20 percent reduction compared to Fable 5.1. Claude Fable 5.2 is available on Amazon Web Services, Google Cloud, and Microsoft Azure, and requires 30-day data retention to be enabled. Anthropic also released Claude Mythos 5.2, a companion model with fewer safety guardrails. m:…2663

Anthropic's own positioning is narrower than the launch suggests. Claude Opus 5.5 is recommended for most daily coding and knowledge tasks, which means Fable 5.2 is the specialist: the model for complex reasoning and long-running agent work, not the default. The five effort levels are the tell. A model that lets you dial reasoning depth up and down is a model built for agentic loops where the cost of thinking is the budget. m:…2663

GLM-5.3 and Hy4

GLM-5.3 is Z.ai's flagship model for complex software engineering and long-horizon agent tasks, and it is available on SambaCloud at over 200 tokens/s. GLM-5.3 reaches state-of-the-art results among open-weight models on Terminal-Bench 3.0 and Agents' Last Exam, scoring 28.3 on Terminal-Bench 3.0 and 66.9 on DeepSWE v1.1. On Z.ai's internal Bench, GLM-5.3 reaches 31.4% at around 50K output tokens per task, while Claude Opus 4.8 reaches 29.5% with 120K tokens. m:…7942

The security angle is the interesting part. GLM-5.3 holds the best published score on CyberGym, and the US Center for AI Standards and Innovation rated GLM-5.3 the most advanced open-weight model for cybersecurity. Strix is an open-source security agent that uses GLM-5.3 to find and validate vulnerabilities, and ZCode is Z.ai's agentic development environment built for long-horizon tasks. Claude Code is the top app routing tokens to GLM-5.3 on OpenRouter with 2.5T tokens, which is the kind of adoption number that says more than a benchmark. m:…7942

Tencent's Hy4 arrives through a different route: production first. Tencent and Inferact announced a strategic collaboration on the Hy4 model, and Inferact operates Hy4 in production, managing capacity, routing, caching, observability, and failure recovery. Hy4 production serving supports long-context reasoning, and Hy4 Preview supports a context window exceeding one million tokens. The teams are building native vLLM support for Hy4 Preview, and findings from serving Hy4 in production can be incorporated into Hy4 Preview's public implementation or upstreamed into vLLM. Tencent retains control of the checkpoint, serving requirements, and validation standard. m:…7925

How to tell them apart

The week's most useful contribution is not a model. It is a testing method. Fair comparison of coding models requires consistent project instructions, tool permissions, and application versions, and the evaluation methodology uses a scoring system of eight observable requirements. Visual judgment should be kept separate from the functional behavior score, and landing pages can appear finished while containing functional errors like invalid form rejection. m:…4415

The method is specific enough to copy. Each model should be run five times to assess consistency, and the landing page brief requires tune-ups to cost $85 and flat repairs to start at $20, with the page responsive at a viewport width of 390 pixels without horizontal scrolling, and functioning without external libraries, fonts, or images. GPT-6.1 Sol requires the Responses API for tool calling when using a direct OpenAI connection, and Anthropic retains Fable 5.2 data for 30 days unless otherwise authorized. Those are the kinds of details that decide whether a model works in your stack, and no benchmark score captures them. m:…4415

The lesson of the week is to run the test yourself. The vendors will give you the numbers that flatter them: the context window, the benchmark score, the tokens per second. The fair test gives you the numbers that matter to you: whether the form rejects bad input, whether the page works at 390 pixels, whether the model does the same thing five times in a row. The models are getting better fast. The testing is what has to catch up. m:…4415