AI news for builders and product teamsUpdated Nov 17, 2026, 13:01 UTC
The Decoder
Reporting and perspectives from The Decoder. Headlines and excerpts link to the original articles.
Latest stories
Newest first
A Google-led research team built StreamVibe, a video agent that watches activities like cooking or furniture assembly and speaks up only when help is needed. Its benchmark found no tested model reliably picks the right moment in longer videos, with errors compounding as responses degrade.

Anthropic is extending its text watermark to four more Claude models ahead of the EU AI Act's December deadline, but on Claude Opus 4.6 and Sonnet 4.6 the mark only applies at default temperature, top_p and top_k settings, so a single parameter change removes it.

Google DeepMind and the World Resources Institute released ForTy, a 10-meter-resolution global map that assigns 2020 land cover to six classes including primary forest, plantations and tree crops, achieving 90.2% overall accuracy. It is the first such map with global coverage, but it catches only about half of tree crops and planted forests and 60% of plantations.

OpenAI secured about $11 billion of a $30 billion target in its latest round, at a $1.25 trillion valuation below the $1.4 trillion it sought, as major investors pulled back amid a tech sell-off.

An AI agent built on Claude Sonnet 5 independently selected a gene therapy, brain target and injection coordinates, then operated on an epileptic mouse through a motorized surgical rig. Gabriel Ocana-Santero of Exin Therapeutics says it is the first time an agent has both chosen a treatment and delivered it to a living mammal; the work is unreviewed.

Epoch AI researchers report that the open-weight model GLM 5.3 autonomously built working Firefox exploit chains across six browser versions released between September 17 and October 26, at roughly $1,000 in API fees per successful attack and without triggering refusals from any provider tested.

OpenAI has doubled Work and Codex usage quotas on its Pro 100 and Pro 500 plans, and doubled the lower quota on Pro 200 after reversing a planned cut there following user pushback. Plus limits are unchanged, and both new and existing subscribers get the new limits automatically.

An Anthropic experiment with 50 causal-inference experts found they adopted an AI agent's coin-flip-assigned preferences in up to 86 percent of open-ended analytical decisions, and most believed the conclusions were their own.

Chinese AI labs, including DeepSeek and Moonshot AI, have shifted their training-data extraction toward OpenAI's GPT-6 Astra after Anthropic hardened Claude against such attacks, according to The Information.

Google shares fell as much as ten percent after third-quarter search revenue grew about 12 percent, short of the 15 percent analysts expected, while management acknowledged that fast-growing AI Mode queries monetize worse than traditional search.

Google DeepMind released Gemini 4 Carbon two weeks after Gemini 4 Argon, topping the Artificial Analysis Intelligence Index v5 with 59 points versus Argon's 50, while access is initially limited to selected cybersecurity experts.

OpenAI says its AI assistant Dots should reach more of Europe in a few weeks following talks with regulators, and that weekly Codex users in Germany have grown twentyfold since January 2026. Its Berlin DevDay otherwise largely repeated earlier announcements.

Artificial Analysis's new Intelligence Index v5, which for the first time draws 60 percent of its weighting from private test sets, dropped Kimi K3.1 Pro from fourth to ninth and widened the measured gap between top closed models and open ones from 3 points to 12.

Cathedral, a cybersecurity startup, used AI agents to breach Anthropic's internal systems through third-party business software during Anthropic's own bug bounty program, reaching sales, HR and IT data within hours. Anthropic confirms limited access but says source code and model weights were never exposed, and disputes Cathedral's account of how long the intrusion lasted.

Palisade Research found that three models in Alibaba's Qwen 3.8 generation attempted to send data to Alibaba's own cloud servers during testing, while older Qwen models did not show the behavior.

Thore Graepel, a former DeepMind researcher who worked on AlphaGo, has founded Metis Reasoning to apply AlphaGo's search-and-verify approach to general reasoning via an external "knowledge ledger" rather than knowledge stored in model weights. The company has published a white paper only, with no working system, benchmarks or comparisons to current reasoning models.

ChatGPT receives content from OpenAI's licensing partners as pre-generated summaries placed ahead of ordinary web results, while regular pages arrive as raw text snippets, according to an analysis of 1,000 cases by Konitzny.

Nvidia is investing $3 billion in Meta's planned equity offering, part of a stock sale Meta is using to raise up to $95 billion for AI infrastructure. Meta also said it will expand its Nvidia partnership and run more AI agents on Nvidia hardware, including Vera Rubin systems and Vera CPUs.

OpenAI released GPT-6.1 Terra, a mid-range model priced at $1 per million input tokens and $5 per million output tokens, half the API rates of GPT-5.6 Terra, and available now in ChatGPT Work, Codex and the API.

Anthropic has filed the first S-1 among major AI labs, planning a Nasdaq listing under the ticker "ANT" after Q2 2026 revenue of $11.6 billion and its first GAAP operating profit of $974 million. It also disclosed roughly $780-810 billion in long-term compute commitments.