Anthropic Launches Claude Duet: An Always-On AI Agent That Works While You Don't
Anthropic opened a research preview of Claude Duet on October 21 — a standalone app where Claude operates as a proactive background agent. Unlike the familiar chat interface, Duet connects to services like Gmail, Slack, and Stripe, delivers a daily briefing, and works on tasks autonomously between conversations.
The app is built on Claude Opus 5.5 and runs on iOS, Android, and the web. US-based Pro and Max subscribers can join a waitlist. Duet asks for user approval before sensitive actions such as payments or sending messages, and bank transfers are blocked entirely.
Anthropic acknowledges the current limitations openly: Duet can misread conversation tone, misjudge urgency, and its background browser sessions aren't yet visible to users. Background checks run every few hours rather than continuously, partly to manage subscription usage limits.
The launch directly targets similar offerings from competitors — OpenAI's Dots, Meta's Muse Wonder, and xAI's Grok Bot — making always-on AI agents the defining battleground of late 2026.
Source: 9to5Mac, The Decoder
Google's Gemini 4 Argon Reclaims Benchmark Lead — With Caveats
Google's Gemini 4 Argon, first announced on September 30 and now broadly available, tops 13 of 19 benchmarks in Google's own evaluation table — including a 77.9% score on DeepSWE v1.1 (vs. Opus 5.5's 74.2%) and an industry-low 0.7% attack success rate on Gray Swan's prompt injection benchmark.
But independent evaluations tell a more nuanced story. Artificial Analysis places Argon in the top three labs but below Opus 5.5 on its composite quality index (52.6 vs. 57.6), and the model lags behind GPT-6 Astra on FrontierSWE v2 (55.0% vs. 65.5%) and behind Opus 5.5 on Terminal-Bench 4.0 (57.4% vs. 66.4%). On Arena AI's leaderboard, Argon ranks first for text but only ninth on Code Arena for web development.
Google is offering introductory API pricing through November 30 and has expanded the model's context window to 1 million output tokens per response — a major jump from the 64K limit in the prior generation.
Source: VentureBeat, Artificial Analysis
OpenAI Nears $30 Billion Bridge Round at $1.25 Trillion Valuation
OpenAI is close to finalizing a bridge funding round of up to $30 billion at a $1.25 trillion pre-money valuation, according to multiple reports from October 22. UAE investors and Coatue are leading the round, which would put the post-money valuation near $1.28 trillion.
The price represents a $150 billion haircut from the $1.4 trillion figure OpenAI presented to backers including BlackRock and MGX earlier this month. OpenAI frames the round as "opportunistic" bridge financing ahead of a potential IPO, with CFO Sarah Friar telling employees a listing could happen in 2027.
The round would follow OpenAI's $122 billion financing in March that valued the company at $852 billion post-money — meaning the company's paper valuation has grown by roughly 47% in seven months. OpenAI has also been expanding its distribution, recently bringing its models and Codex coding tools to Google Cloud, making them available on all three major US cloud platforms.
Source: Yahoo Finance, SiliconANGLE
Anthropic Pledges 5% of Shares to Public Benefit Fund Ahead of IPO
In a move that could reshape how AI companies approach going public, Anthropic announced on October 18 that it will commit 5% of its shares to a new Anthropic Public Benefit Fund. The commitment includes a mechanism for growth: the board may allocate an additional 2% of shares for each 50% rise in company value above an undisclosed baseline.
The fund will make grants to nonprofits and direct money toward groups affected by AI disruption — with labor market displacement named as one area of focus. Anthropic has explicitly ruled out lobbying or political donations from the fund.
The announcement arrives just weeks before a potential November IPO, which Reuters describes as potentially one of the largest in history. Anthropic plans to list as a public benefit corporation, with roughly 20% of shares going toward public good purposes including the new fund. Details on which organizations will receive funding remain scarce.
Source: Al Jazeera, ABC News / AP
Mistral Previews Its Largest Model Yet: 1 Trillion Parameters, Open Weights Coming
French AI lab Mistral launched a public API preview of Mistral Large 4 on October 6, its most ambitious model to date. The model uses a granular mixture-of-experts architecture with approximately 1 trillion total parameters and 49–52 billion active per forward pass, along with a 1.6 billion parameter vision encoder.
The model is natively multimodal with a 1 million token context window, and Mistral has promised open weights by the end of October — with some sources pointing to October 27 as the target date. Until then, the model is served under Mistral's proprietary terms through Mistral Studio.
If the open release happens on schedule, Mistral Large 4 would become one of the largest openly available models in the world, reinforcing the European lab's position as a leading open-source alternative to the US-based frontier labs.
Source: TestingCatalog, CellCog