MiniMax's M3.1 Flash launch leads the October 19 AI briefing: in the past 24 hours, MiniMax turned the Flash model it quietly shipped in late September into a formal release with published benchmarks, Anthropic said it will commit 5% of its shares to a new public benefit fund, Disney moved to hire most of bankrupt Character.AI's staff, DeepSeek's new Flash model topped the Code Arena web development board, and SemiAnalysis forecast 50 gigawatts of new global data center capacity in 2027 alone.
MiniMax formally launches M3.1 Flash: nearly 3x faster inference, 70% lower cost
MiniMax formally updated M3.1 Flash on October 19, 2026 through its official WeChat account, teasing it the same day on X with a "welcome back, Space Bunny" post — Space Bunny, the anonymous model that had quietly processed more than 200 trillion tokens, was this model's early version. By the company's published figures, the new release's online inference runs nearly three times faster, overall cost on agentic tasks is down 70% versus the previous generation, and everyday output speeds exceed 200 tokens per second with a median holding above 170; it scores 73.7% on the DeepSWE v1.1 coding benchmark, 61.7 on the OSWorld V2 computer-use benchmark, and 82% on the CyberGym cybersecurity benchmark. The new version uses a more aggressive sparse-attention architecture, and its benchmark charts report scores across five thinking-effort levels from low to max. Pricing is published: ¥1 per million input tokens, ¥4 per million output tokens and ¥0.1 per million cached-read tokens; the model is live on the MiniMax API, M Plan and MiniMax Code, with open weights to follow and a capability-first Pro version also on the way. The pacing matters: on September 27 it first appeared as a Preview inside MiniMax Code, with no benchmark table and no pay-as-you-go price. This release adds the formal version, the scores and the price, putting the Flash tier properly on stage. All figures are vendor-published with no third-party verification yet, but the direction is clear: leading labs are now using near-flagship agent ability at a fraction of the price to compete for everyday coding and computer-use work.
Anthropic commits 5% of its shares to a public benefit fund that grows with its valuation
Anthropic announced on its official blog on October 18, 2026 that it is creating the wholly owned Anthropic Public Benefit Fund, which will receive 5% of the company's shares subject to stockholder approval; for every 50% increase in the company's value above a set baseline, the board may add a further 2%. At the company's $965 billion valuation from May, the initial 5% works out to roughly $48 billion. The fund is structured as an LLC rather than a charity, which the company says keeps it flexible between large donations and direct distributions to specific populations, and it states that none of the equity held by the fund will be used for lobbying or political giving. The company also tallied the bigger picture: its co-founders have pledged to donate 80% of their personal equity, about half of employees take part in an equity donation match program, and with the new fund roughly 20% of Anthropic's shares will be dedicated to the public good; details on governance and valuation milestones will follow in the coming months. The company is on the verge of going public, with multiple outlets reporting an IPO as soon as November. Redirecting the escalating equity usually reserved for executive pay into public benefit is its institutional answer to the industry's trust problem — though who ultimately receives the money, and who decides, remains undisclosed.
Disney to hire most of Character.AI's staff: at least 60 people, about $60 million in annual pay
TechCrunch reported on October 18, 2026 that a bankruptcy court filing spotted by Bloomberg revealed more detail on Disney's agreement with Character.AI: Disney will make employment offers to at least 60 Character.AI employees — the majority of its workforce — at an estimated $60 million in combined annual compensation, and the filing says the vast majority of those offered jobs have accepted them. Character.AI had previously announced that Disney, as the stalking-horse bidder at $20 million, is seeking a non-exclusive, perpetual license to part of its core intellectual property; that bid can still be topped, and any deal must go through the bankruptcy auction and court approval. Character.AI's former CEO Karandeep Anand has already become Disney's first chief technology officer. Character.AI filed for Chapter 11 protection last week; the filing says that despite a large, engaged user base, the company was hobbled by high costs and potential litigation and regulatory pressure, including a lawsuit and government investigations over teen safety. A star companion-chat company ending in bankruptcy, with its people and technology packaged off to an entertainment giant, is a complete case study in how high costs plus compliance risk can sink a consumer AI product; what happens to the platform itself awaits the auction.
DeepSeek-V4.1-Flash tops the Code Arena WebDev board at roughly one-tenth the cost
Arena announced through its official account on October 18, 2026 that DeepSeek-V4.1-Flash-20261017 (Max) is now the top-ranked open model on its Code Arena: WebDev leaderboard, ahead of Kimi K3 (Max), at roughly ten times lower cost. Code Arena rankings come from users voting on blind head-to-head outputs, and the WebDev board measures how well a model generates usable web pages and front-end apps in one shot. This is not a general-capability ranking, so do not extrapolate too far; but an open model at a Flash price point reaching the top of a high-frequency task like front-end generation is a direct signal for teams choosing a coding-assistant model: retest it on your own real page tasks before deciding whether to switch.
SemiAnalysis: 50GW of new data centers in 2027 alone, with leases signed first
SemiAnalysis published a report on October 19, 2026 forecasting that global data center capacity will grow by 50 gigawatts in 2027 alone — roughly equal to everything built worldwide through 2024, rebuilt in a single year. Two leading indicators back the call: just the top 40 campuses already in advanced construction account for more than 24 gigawatts, most of them more than halfway built; and more than $1 trillion in binding, non-cancellable data center leases have been signed in 2026 year to date, with pre-leases running at a total contract value of $30–50 billion per gigawatt. The report also notes that modular designs have cut some campus build times to under 10 months, that operators are bypassing grid queues with behind-the-meter power plants, and that construction outside the US accelerates sharply in 2027. It warns that in 2027–28 the bottleneck could shift from buildings and power to memory and logic chips. For users, a 2027 supply surge means tokens keep getting cheaper; for investors, trillion-dollar leases and debt come first, and returns wait until those campuses actually light up.
