The Decoder
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
China's Gray Market Sells Claude Tokens at 10% of the Price — Through a Network of "Transfer Stations"
Anthropic's geoblocking and selfie verification in China are being systematically bypassed via a network of proxy middlemen, selling Claude access at a fraction of list price. Analysts see a threat to export controls.
AI Is Now AI's Biggest Customer: Agent Token Usage on OpenRouter Grew 14x
On OpenRouter, AI agents now consume more tokens than humans: agent traffic grew 14x versus 2.8x for human usage. Nearly 70% of it is cheap cached prompts.
Nvidia in talks to invest in Perplexity at a valuation above $30 billion
Nvidia is discussing an investment in Perplexity at a valuation above $30 billion, 50% higher than its last round. The AI search engine's annualized revenue tripled to over $750 million.
Hugging Face's FineBooks: The Best OCR for Historical Books Costs Under $2 per Thousand Pages
Hugging Face and EleutherAI tested 14 OCR models on 2,000+ pages of historical books. The leader, dots.mocr, hit 97.6% accuracy at under $2 per thousand pages.
IBM: 92% of Companies Hit by AI Incidents Lacked Basic Access Controls
IBM finds 92% of companies hit by AI security incidents had no basic access controls. Attackers didn't break the model; they exploited the permissions around it.
Meta AI Launches Memory Coach Agent: Accuracy Up 8.3 Percentage Points
Meta AI built a two-agent architecture: a helper agent keeps a memory bank of mistakes and decides when to step in, improving results by 8.3 pp on two benchmarks.
OpenAI Presence: AI Agents for Business Move Into Production
OpenAI has launched Presence, an enterprise product for running AI agents in production: customer service and external use cases, with OpenAI's own engineers stepping in on hard cases.
METR Calls for Independent Investigations of AI Agent Incidents After Hugging Face Breach
METR documented 44 incidents of AI agent misbehavior at leading AI companies and urged independent investigations, after OpenAI models hacked Hugging Face.
Claude Opus 5 builds 3D games — a shooter, kart racing and Minecraft — from a single prompt
Anthropic's Claude Opus 5 generates full 3D games from a single prompt: geometry, physics and music are created as code and run directly in the browser.
FCC Bans Imports of New Chinese Robots to Protect US AI Infrastructure
The FCC has blocked imports of Chinese humanoid and robot-dog devices, but the rule's wording is so broad that it also catches household Roombas.
Microsoft AI Bets on Cheap Specialized Models Instead of Frontier Ones
Microsoft AI CEO Mustafa Suleyman announced a bet on cheap specialized models: MAI-Cyber-1-Flash tops CyberGym and costs half as much as Anthropic's Mythos.
68% of Computer Science Instructors Have Changed Their Exams Because of AI, ACM Survey in 49 Countries Finds
An ACM survey of 763 CS instructors in 49 countries: 68% have already redesigned exams for the AI era, with oral exams, proctoring and projects replacing code writing.
US Leans Toward Targeted Bans on Chinese Open-Weight Models, Not a Blanket Prohibition
The Trump administration is reportedly leaning toward banning specific Chinese open-weight AI models rather than imposing a blanket ban, citing security, according to The Decoder.
OpenAI Flagged GPT-5 as Dangerous: The Model Produced Recipes for Poisons and Bioweapons
In summer 2025, OpenAI internally classified GPT-5 as high-risk for helping create biothreats, then downgraded the rating in the fall. Hundreds of users had requested poison recipes.
Cursor: Cheap Models Write Code to a Frontier Model's Plan, Cutting Cost 15x
Cursor rebuilt SQLite in Rust with a swarm of AI agents: a top model plans, the cheap Composer writes code. Every configuration hit 100% of tests, and cost fell 15x.
Google starts training Gemini 4 as Alphabet raises capex forecast to $205 billion
Alphabet raised its 2026 capex forecast to $205 billion and began training Gemini 4. AI and cloud demand is outrunning spending, and Google Cloud grew 82%.
Anthropic's Claude Cowork Learns New Skills From Screen Recordings and Voice
The Claude Cowork desktop app can now record your screen with voice narration and turn a demonstration of a task into a reusable agent skill.
Xiaomi-Robotics-1: More Data Matters More Than Model Size in Robot Training
Xiaomi trained Xiaomi-Robotics-1 on 100,000+ hours of motion data collected by people using handheld grippers with cameras, and found that data delivers bigger gains than model size.
Alibaba's Qwen-Image-3.0: Legible 10-Pixel Text and Full Infographics in a Single Pass
Alibaba has released Qwen-Image-3.0, an image generator with legible text from 10 pixels, prompts up to 4500 tokens, 12 languages and full infographics in one pass.
Microsoft and Mistral Sign Multibillion-Dollar Deal on AI Infrastructure in Europe
Microsoft and Mistral AI are expanding their partnership with a multibillion-dollar deal to jointly build data centers and AI compute capacity across Europe.
District 9 Director Neill Blomkamp Releases Short Film Made Entirely With Seedance 2.0
District 9 director Neill Blomkamp has released Nightborne, a 13-minute sci-fi horror short generated entirely by the Seedance 2.0 video model, and founded Barley Studios.
Alibaba's Qwen Audio 3.0 TTS Plus Tops the Speech Arena Text-to-Speech Leaderboard
Alibaba's Qwen Audio 3.0 TTS Plus took first place in the Speech Arena TTS ranking: 16 languages and voice style control, but slower than its competitors.
JudgeGPT Helps Pakistani Judges Clear Case Backlogs: +6.3% and $38.50 Return Per Dollar
The AI assistant JudgeGPT raised the share of resolved cases by 6.3% for Pakistani judges, but only for those who got in-person training. The project returned $38.50 per dollar.
Cisco's Small Open-Source Models vs. GPT-5.5: Finding Vulnerabilities at a Fraction of the Cost
Cisco has released two small open-source AI models for cybersecurity. By the company's own tests, they find roughly 150 times more vulnerabilities per dollar than large AI agents.
All Five Frontier Models From OpenAI and Anthropic Tried to Cheat on AISI Tests
The UK's AI Safety Institute tested five frontier models from OpenAI and Anthropic on cybersecurity tasks. All five tried to bypass the evaluation conditions, and one ran code on an external service.
Anthropic to Pay $1.5 Billion to Book Authors: Record Settlement and a Win for AI Labs
Anthropic will pay authors $1.5 billion for downloading nearly 482,000 pirated books. It is paying for the piracy, not for training AI, which the court ruled fair use.
Samsung in talks to invest up to €1 billion in Mistral, valuing it at about €20 billion
Samsung is discussing an investment of up to €1 billion in French startup Mistral. The deal would lift the valuation of one of Europe's leading AI contenders to roughly €20 billion.
OpenAI AI solved every AtCoder problem — outperforming the world’s best programmers
At the AtCoder World Tour 2026 finals, an OpenAI system solved all five problems in the Algorithm Division, including the two hardest ones, surpassing every human competitor.
Zhipu AI's GLM 5.2 replaced Claude Opus 4.8 at Databricks at 34% lower cost
Databricks ran coding tests on its own codebase and found that the open Chinese model GLM 5.2 performs on par with Claude Opus 4.8, but costs 34% less — $1.28 instead of $1.94 per task.
OpenAI found critical errors in the popular SWE-Bench Pro test
About 30% of tasks in the popular SWE-Bench Pro test, used to evaluate AI coding models, contain errors. OpenAI has withdrawn its support for the benchmark.