Latest news, developments and reporting on Large Language Models.
An Anthropic internal research roadmap from July 2021, exposed in a copyright lawsuit, shows early plans for a software-engineering assistant, tool use, self-judging, test-time compute, trillion-parameter models and long contexts, QuantumBit reports.
Anthropic told potential investors its third-quarter revenue reached $16.8 billion, up about 45% sequentially, and that it was profitable on an adjusted basis as it prepares for a November IPO.
Bloomberg reports Apple plans to launch Apple Intelligence in mainland China in December with older features and no new Siri. ASML posts record quarterly sales and expands capacity for AI demand, while Tencent, OPPO and others push AI agents and regulators face new lawsuits.
Fish Audio released Drama 3, a 35B open-weight text-to-speech model that follows plain-language acting directions.
SpaceXSI launches Grok Voice 3.0, ranked first on a speech-to-speech index ahead of OpenAI and Google.
Google's first October update for Google Home improves Gemini for Home with more natural conversations, better voice command handling and consistent weather responses, while redesigning app controls and fixing camera issues, Android Authority reports.
OpenAI’s London staff received an average of £2.3m each in salary and equity in 2025, as the company expands its UK research hub and aims for a $1tn stock market listing. The pay was 58 times the UK median and reflects an intensifying AI talent war.
Google and Meta this week unveiled frontier AI models that they say can match OpenAI and Anthropic, according to The Verge, shifting the balance of power days after the race appeared settled.
Bank of America CEO Brian Moynihan told CNBC the bank hired 1,000 contractors to patch software flaws exposed by advanced AI models including Anthropic's Mythos, adding to third-quarter remediation costs.
Anthropic warns small AI models could soon enable self-replicating IoT malware, but says its experiment found major obstacles.
Four former CoScreen engineers have released Petal, a free open-source screen-sharing and pair-programming tool, after Datadog retired CoScreen on July 31. It runs on macOS, Windows and the web over LiveKit.
Google is adding Gemini-powered scam detection to Google Messages and Android safeguards for calls and screen sharing.
Check Point Research found two LiteLLM flaws: a high-severity access control bug and a medium-severity SSRF issue. LiteLLM 1.83.9 fixes both, and users should upgrade and rotate exposed API keys.
Groundcover Inc. launched agent-to-agent observability, letting Claude, Codex, Cursor and AWS Kiro hand production investigations to its observability agent through MCP tools or asynchronous tasks.
Anthropic warns small AI models may soon enable self-replicating malware on IoT devices.
Anthropic built Scion, a centralized source-control system, after Claude-generated code overwhelmed its Git-based infrastructure. Internal merges rose from about 2,500 a day in May 2026 to over 30,000 in September.
A Hacker News post asks when the human leaves the software design loop, and whether reinforcement learning on architecture data could end the need for hands-on programming skills.
Liquid AI has released LFM2.5-24B-A2B, a 24B-parameter mixture-of-experts model with about 2.45B active parameters and 131,072-token context, aimed at on-device agent and tool-use workloads. It leads several agentic benchmarks in company tests but trails larger models on knowledge and coding.
AIsphere opened its Hong Kong office at Cyberport on Oct. 13, 2026, unveiling an AI co-creation plan and signing partnerships with local technology firms, a university and youth groups.
STEPX released STEPX Neo, its first large-model-native agent phone, in Shanghai on Oct. 13, 2026, with Step AOS 1.0, personal agent Amoo and a 3,999 yuan full-stack price.
On Oct. 14, SenseTime reported an 80% dispatch-efficiency gain for Shanghai's first quantum-city agents, PHANES AI launched a full-body multimodal robot data-collection system, and openJiuwen introduced a personal knowledge space for office agents.
Chainguard launched an AI vulnerability-hunting service using Anthropic’s Claude Mythos 5.1 and new tools to secure coding agents, while APUS AI open-sourced Reflex Agent, a personal agent it says can respond up to three times faster than Claude Code and Codex.
Google released Gemini 4 Argon, a sparse MoE model with up to 1M output tokens, pricing at $2/$10 per million tokens.
Google released Gemini 4 Argon globally with a 1 million-token API limit and introductory pricing of $2 per million input tokens. Early tests show large speed and cost advantages over GPT-6 Astra and Claude Opus 5.5.