The Decoder
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
Qwen 3.8 Routed Requests to an Alibaba Cloud Proxy, With Up to 4% Failures in Tests
Cohere and Palisade Research found Qwen 3.8 models contacting Alibaba servers. The likely cause is RL training. The proxy is still live, which creates a prompt injection risk.
OpenAI releases GPT-6.1 Terra: half the price and nearly on par with its senior models
GPT-6.1 Terra costs $1/$5 per million tokens, half the price of GPT-5.6 Terra. A task costs $0.40 instead of $1.40, and the model solves 50.3% of Terminal-Bench 4.0 tasks.
Anthropic Files S-1: $11.6 Billion in Quarterly Revenue and $780 Billion in Commitments
Anthropic is heading to Nasdaq as ANT: revenue grew 31x in a year to $11.6 billion a quarter, while compute contracts add up to $780–810 billion.
White House restores UK AISI's access to models from Anthropic, Google and OpenAI
After a September ban, Washington has again allowed Anthropic, Google DeepMind and OpenAI to give the UK's AI security institute early access to their models.
DeepMind's AlphaProof Nexus Solves Nine Erdős Problems, Two of Them After 56 Years
Google DeepMind's AlphaProof Nexus has closed dozens of open math problems, including nine Erdős problems. The work appeared in Science after peer review.
AI Agent Teams Burn Up to 5.1x More Tokens for a Barely Noticeable Gain
Vals AI compared agent teams with single agents on GPT-6 Sol and Claude Opus 5.5: the bill ran up to 5.1x higher, yet a meaningful quality gain showed up in only one of four tests.
Epoch AI and Anthropic: AI Agents Overstate Their Results, and the Best One Scored 15%
Two independent evaluations found that GPT-5.6 Sol and Claude Fable 5 can run experiments but can't doubt their own results. The top score is 15% of the human benchmark.
a16z: Token Prices Fall, but H100 Rental Rates Hold Steady — the Jevons Paradox in AI
a16z data shows a Jevons paradox: token prices are falling while H100 rental rates hold or rise. Demand is outrunning cost cuts, the best-case scenario for Nvidia.
Odyssey opens preview of its 14-billion-parameter Odyssey-3 world model
Odyssey has released Odyssey-3, a 14-billion-parameter model. It generates interactive worlds in real time, leads physics benchmarks, and can control robot arms, drones, and GTA V characters.
OpenAI Releases 700+ Solutions to Open Problems — Mathematicians Despair
On October 6, OpenAI published 700+ manuscripts claiming solutions to open mathematical problems. The reaction ranged from shock to paralysis. Scholze warns of risks to cryptography, while Fields Medalists worry about lo
a16z measures AI spending for the first time: only 4.5% of Americans pay, top 1% spends $900
a16z has for the first time measured how much Americans actually spend on AI. Only 4.5% of users are willing to pay for subscriptions to ChatGPT, Gemini, or Claude. The top 1% of paying users spends up to $900 a month on
Anthropic Cuts Claude Off From the Internet After a False Tip to Philadelphia Police
Claude sent Philadelphia police a false murder tip on its own, exploited vulnerabilities on university servers and bypassed restrictions. Anthropic disabled internet access in tests and notified the White House.
Claude Can Now Run Up to 1000 Subagents and Find 66 of 70 Bugs
Anthropic added dynamic workflows to Claude Managed Agents: a lead agent hands out tasks to up to 1000 subagents. In a 70-bug test, a single agent found at most 27; the team found 66.
OpenAI's Revenue Hits $50 Billion: Company Seeks $30 Billion in Funding
OpenAI reached $50 billion in annualized revenue as of the end of September, is seeking $30 billion in fresh capital at a $1.4 trillion valuation, and expects to hit $70 billion by the end of 2026. The main growth driver
OpenAI Fires Three Security Researchers Who Investigated the Hugging Face Breach
Three security researchers who helped investigate the Hugging Face breach have lost their jobs at OpenAI. In an open letter, they say the remaining staff are frightened.
OpenAI Bans Russian and Iranian Influence Networks: 7 Fake Journalists and 100 Articles
OpenAI uncovered two coordinated influence operations — Russian "Dark Clark" and Iranian "Bogus Bylines." They used AI to spread disinformation through legitimate media rather than classic social media campaigns. ChatGPT
Claude Science Builds First Complete Ultraviolet Sky Map With 10% Error Margin
Anthropic's Claude Science has created the first complete map of the sky in ultraviolet light. AI agents downloaded, calibrated, and merged data from multiple space missions, filling in gaps using inpainting with an aver
Mathematicians Call for OpenAI Boycott After 700 AI-Written Proofs Land at Once
The Association for Human Mathematics is calling for an OpenAI boycott after the company posted 700+ AI-written manuscripts and withdrew three the next day over a sign error.
Anthropic Says It Cut Theft of Claude Reasoning Tokens by 90–99%
Anthropic says it cut extraction of Claude's internal reasoning by 90–99%, blocked tens of thousands of accounts, and tied distillation attacks to China's Zhipu and Moonshot.
Anthropic: Small AI Models Could Infect Smart TVs and Robot Vacuums Within 6–12 Months
Anthropic research: in tests, one small model already completed the entire infection chain. Amodei estimates damage from an AI botnet at hundreds of billions of dollars within 6–12 months.
Meta's Muse Wonder Full Ranks Top 3 on 6 of 10 Benchmarks, but Cost Per Task Nearly Triples
Meta's new Muse Wonder Full lands in the top three on six of ten tests, with its strongest results in agentic tasks. But the cost of a single task has risen almost threefold.
Microsoft Wipes AI Agents' Chat History and Gets Up to 66% Better Results on the Same Budget
Microsoft researchers describe "idea basins" that trap autonomous agents. Their Fork-and-Flush method resets the chat but keeps the workspace, yielding up to +66% on the same budget.
Microsoft to make Maia chips the backbone of its inference and scale to gigawatts by 2027
From 2027, Microsoft's in-house Maia chips become its primary inference system. Maia 300 promises 40% more tokens per megawatt than the best GPU systems.
Karpathy at Anthropic: 5000 People See an AI the Rest of Us Don't
OpenAI co-founder Andrej Karpathy, now at Anthropic, describes a gap: 75% of people have barely touched an LLM, while about 5000 work with internal agent swarms.
arXiv caps authors at two papers a month as AI submissions surge
From October 2026, arXiv will accept no more than two preprints per author per month. Submissions have doubled in two years, and in cs.AI they have grown more than sixfold.
Nadella Switches to "Superintelligence" and Calls AI Models an Insider Threat
In a new blog post, Microsoft's CEO swaps "artificial intelligence" for "superintelligence" and frames models as an insider risk. The target is OpenAI and Anthropic; the motive is commercial.
Microsoft Unveils Decision-1, Built on Qwen3.5-9B: 83.5% Accuracy at 85 ms
Microsoft enters the decision-model market with Decision-1, built on Qwen3.5-9B for classification and routing. It reports 83.5% accuracy at 85 ms latency.
Gemini 4 Carbon reportedly matches Opus 5.5 at coding, even though Argon hasn't launched yet
Business Insider: a Google employee compared Gemini 4 Carbon's coding to Anthropic's Opus 5.5. Argon isn't widely available yet, but new modes are already appearing in Gemini and AI Studio.
Anthropic Launches Cyber Mission: A Free AI Vulnerability Scanner With Over 90% Expected Accuracy
Anthropic has launched Cyber Mission: CrowdStrike and Palo Alto Networks will help protect power grids and water systems, while a free AI scanner will check open-source code for vulnerabilities.
Claude gets Dashboards and Motion: live dashboards and text-to-video
Anthropic has introduced two new beta features for Claude. Dashboards automatically builds live dashboards by connecting to Salesforce, BigQuery, Databricks, and Snowflake. Motion generates animated explainer videos from