joelceth (joelceth)

openrouter just dropped a mystery model called pony alpha that processed 4 billion tokens in 24 hours, and everyone is asking the same question: who actually built it? openrouter has done this before. quasar alpha and optimus alpha turned out to be openai's gpt-4.1. sherlock alpha was revealed as xai's grok 4.1 fast. there is always a major lab behind these stealth models. pony alpha is currently free. it has a 200 thousand token context window. it performs well in coding, reasoning, and agentic workflows. in svg benchmarks it matches claude opus 4.5 quality.

• 10 replies
• 10 recasts
• 81 reactions

ai models answer memorized questions perfectly. but what happens when you give them a new document and say "read this and apply it"? tencent and fudan tested with 1899 tasks: even the best model only scored 24%. why is this so hard? answering from memorized knowledge is easy for ai models. the hard part is reading information they've never seen before and applying it instantly. tencent and fudan university released a test suite called cl-bench to measure exactly this. it includes 500 different contexts, 1899 tasks, and over 31000 validation rules. models receive information they never saw during training and must solve problems using it. the results are striking: the best of 19 tested models, gpt 5.1, only achieved 23.7% accuracy. claude opus 4.5 got 21.1%, gemini 3 pro managed 15.8%. even the most advanced models struggle to learn from context and apply new information.

• 2 replies
• 8 recasts
• 70 reactions

anthropic and openai released their new models on the same day at the same time. opus 4.6 thinks deep, codex 5.3 runs fast. benchmarks show no clear winner. so which one actually gets the job done? opus 4.6 offers a 1 million token context window for the first time. it can read an entire project's codebase in one go. sixteen independent agents worked together to build a c compiler from scratch, and that compiler successfully compiled the linux kernel. on top of that, it found 500 previously unknown security vulnerabilities in open source software without anyone asking it to. codex 5.3 is the first model that helped build itself. it debugged its own training runs and managed its deployment infrastructure. it runs 25 percent faster than its predecessor and does the same job with less than half the tokens. it scores 77.3 percent on terminal tasks, beating opus at 65.4 percent.

• 2 replies
• 12 recasts
• 63 reactions

how does a 0.9 billion parameter model outperform gemini-3 pro and qwen3-vl? china just released 4 new ocr models in a single week, all open source. is the era of massive parameters coming to an end? zhipu's glm-ocr model became world number one on omnidocbench v1.5 with just 0.9 billion parameters. in the last week of january 2026, china released 4 different ocr models back to back. all open source, all free. the interesting part: none of them are chasing parameter counts. instead, they're expert models focused on specific problems. deepseek's ocr2 model reads multi-column documents like a human. instead of going left to right from the top, it reorders based on meaning. baidu's paddleocr-vl handles low-quality photos and curved pages. tencent's youtu-parsing can convert flowcharts directly to code and runs 22 times faster.

• 0 replies
• 8 recasts
• 50 reactions