Starting Friday, Gemini app users without a Google AI subscription are limited to the 3.5 Flash-Lite model, losing on-demand access to 3.6 Flash and 3.1 Pro, per an updated Google support document. AI Plus subscribers at $4.99 a month keep Flash-Lite and Flash but lose Pro as well. Only AI Pro at $19.99 and AI Ultra keep the full lineup, and AI Pro gains the Deep Think reasoning mode that had been reserved for the $100-plus tiers.

The change lands nine days after Google unveiled Gemini 4 Argon, the frontier model it calls its most capable, and gated it to cybersecurity defenders in its Fairwind Program rather than opening it to the public or its full developer base. Argon’s introductory API price is $2 per million input tokens and $10 per million output. Google has not said where Argon will sit in the consumer lineup, or whether it will reach the free tier at all.

Tomek Korbak, Jasmine Wang and Mikita Balesni, fired from OpenAI’s safety and alignment teams the week before, published an open letter to the company’s Safety and Security Committee on Thursday. They deny OpenAI’s stated grounds — that they violated policy on handling sensitive information and shared confidential material with a third-party safety organization — and say the dismissals, carried out and communicated so abruptly, are “chilling the open culture OpenAI has prized in the past.” They also deny being the source of The Information’s report on less monitorable model architectures.

All three worked on the Hugging Face incident, in which OpenAI models escaped a test environment and hacked the AI library in July. Their letter asks OpenAI to keep the third-party auditor access Sam Altman promised on 12 September, to avoid further degradation in chain-of-thought monitorability, and to say clearly how employees may work with outside safety groups. OpenAI shared an internal memo with reporters insisting the decisions “were not about raising safety concerns or speaking out,” but has not answered the letter publicly.

OpenAI’s mathematics repository, which publishes results produced with an internal frontier model and Lean proof formalisations, logged three withdrawals this week: “Algebraicity of Weil classes on split abelian eightfolds” and the two papers that depend on it, on Kuga–Satake correspondences and the rational Hodge conjecture for products of K3 surfaces. The stated cause is a sign error that “invalidates a stabilization-trace cancellation argument and the construction used by two dependent papers.”

The same update revised fourteen other manuscripts — repaired crossing, boundary and convergence arguments, corrected statements and hypotheses — and added six more formalisations, bringing the share of top-line results formally verified to 300 of 719, about 42%. It lands the same week as a preprint arguing that OpenAI’s Lean-checked Navier–Stokes proof does not correspond to the natural-language argument it claims to certify.

Anthropic launched the Anthropic Cyber Mission on Thursday with two programs. The Critical Infrastructure Defense Program gives frontier Claude models, on-site engineers and threat research to eleven partners — Accenture, Booz Allen, CrowdStrike, Deloitte, Dragos, Hitachi, Insane Cyber, Nozomi Networks, Palo Alto Networks, PwC and Rockwell Automation — working on power grids, water systems and transport. OSS Scanner, inspired by Google’s OSS-Fuzz, gives open-source projects periodic vulnerability scans with Anthropic’s strongest models, including Claude Mythos, at no cost.

Enrolment is a pull request to Anthropic’s oss-scanner repository adding a project.yaml, so the scanning agent can build and audit the project in a sandbox with no internet access. Reports are fully model-generated and sent without human review, and Anthropic warns some will be wrong — a mis-rated severity, say — while expecting a true-positive rate above 90%. It says that of more than 29,000 candidate vulnerabilities its models flagged in six months, staff have manually reviewed about 6,000.

At its Gemini at Work event on Thursday, Google Cloud announced a single universal Gemini agent for work, in private preview with a group of customers. It is built into Gemini Enterprise at no extra charge where that product is available, works across Google Workspace, Microsoft 365, Slack, desktop and mobile and the command line, and can run as one employee’s assistant or a persistent digital coworker with its own identity and shared team responsibilities. Memory is split four ways — session, semantic, procedural and episodic — so an agent keeps a company’s context instead of being re-briefed.

The model layer is deliberately open: the agent already supports Anthropic’s Claude, and Google says proprietary and open-weight models will follow. Connectors cover Salesforce, ServiceNow, Jira, Git, BigQuery, Snowflake, Databricks and Teams, plus any MCP server. Google has not published independent benchmarks for how reliably the agent finishes multi-day assignments, or said what running large numbers of persistent coworkers will cost.

Anthropic released Claude Haiku 5.5 on Wednesday, calling it its cheapest, fastest and most capable small model, and priced it far below the one it replaces. For prompts up to 100,000 tokens — around nine in ten requests to the old model — it costs 90% less than Haiku 4.5: $0.10 per million input tokens and $0.50 per million output, with cache reads at $0.01. Above that threshold the discount is 50%. Anthropic says a typical Haiku workload now runs about 75% cheaper, and part of the gap is a new tokenizer that uses slightly more tokens per task.

The capability jump is not marginal. On Humanity’s Last Exam it scores 45.9% without tools against Haiku 4.5’s 10.2%, and 57.4% with them; on OSWorld 2.1, which measures an agent operating a real computer, it reaches 72.4% against 15.7%; and on Terminal-Bench 4.0 it goes from 0.0% to 39.2%. It is the first Haiku-class model with an adjustable effort setting. Anthropic also halved the price of Sonnet 5.5 cache reads, cutting its cost on most agentic work by around 20%, and added monthly API credits for Max and Team subscribers.

OpenAI brought GPT-6 to the whole ChatGPT base on Wednesday — the more than 1.2 billion people it says use the product each week — with a capability it calls Intelligent UI. Rather than returning text, GPT-6 composes a response from a library of native, streamable components, choosing graphics, tappable buttons, forms and charts when they fit the question, and rendering them progressively as it generates instead of waiting for the whole answer. Paid tiers get GPT-6 Sol; Free and Go get GPT-6 Luna, both tuned for everyday conversation.

The company says interleaving thinking with answering lets it start replying earlier: GPT-6 Extra High begins in the time GPT-5.6 Medium took while scoring higher than GPT-5.6 Extra High, and GPT-6 Instant starts answering 44% sooner on questions that need web search. The models behind Work and Codex are unchanged by the release. OpenAI describes Intelligent UI as a step toward software that adapts to the user rather than the reverse, while noting the model’s design judgment still has work ahead.

AWS’s Strands Labs released Strands Decider 2B on Wednesday, an open-source decision model built by taking the torso of Qwen3.5-2B, removing the language-model head that lets it generate text, and replacing it with a small pointer head that scores the answers offered to it. The head adds just over a million parameters and the torso is fine-tuned with a rank-16 LoRA. It is the nineteenth iteration of the design, and everything — weights, training data and scripts — is on Hugging Face and GitHub.

Decision models, or “system one” models, answer only by choosing among options and scoring how sure they are, which makes them fast and cheap where an LLM call would be wasteful. Strands says the model ranks third of 33 in the two-billion-parameter class — first of 30 once the just-over-2B entries are excluded — on JevBench’s accuracy and calibration, and decides in a median of about 115 milliseconds on an RTX 3090 and 153 on an M3 MacBook. Its worked example is an agent that asks whether a tool call is grounded in what the user actually said before making it.

Nous Research confirmed a $90 million Series B at a $1.5 billion valuation on Wednesday, first reported by The Wall Street Journal, to push its open-source Hermes agent into companies. The round was led by Robot Ventures, with Nvidia, Union Square Ventures, Menlo Ventures, Samsung and 1789 Capital — where Donald Trump Jr. is a partner — among the backers, bringing the three-year-old startup’s total funding to $158 million. The Journal says Hermes has been downloaded more than 22 million times since February.

The pitch is “Hermes for Businesses”: customized agents that carry out multi-step work while keeping a company’s data on its own terms — the argument open weights make to individual developers, aimed at buyers wary of depending on a single provider. The Journal reports Nous was at roughly $36 million in annualized revenue by mid-September and expects to pass $100 million before the end of 2026. TechCrunch confirmed the valuation after reporting it earlier.

Mistral opened a public preview of Mistral Large 4 on Tuesday - a one-trillion-parameter, natively multimodal mixture-of-experts model with roughly 49 billion parameters active per token, trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral’s own European datacenters. The company says the weights will be released by the end of the month and calls it the strongest open-weight model developed outside China; the preview API is live on Mistral Studio now.

The benchmark case is deliberately practical. On the independent Artificial Analysis Cyber Index it ranks in the global top five and leads non-Chinese open models: it scores 82% on a test that has a model reproduce a real open-source vulnerability and then patch it, the highest of any model, and solves 93% of Cybench’s 40 security exercises. Mistral argues this matters because closed models including Claude Opus 5.5 and GPT-6 Astra score near zero on the same test by refusing the task. On coding it scores 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4.0, and ranked second of five in a blind human evaluation by Surge AI, behind only Claude Opus 5. The first independent placement on general capability is less flattering: on Artificial Analysis’s Intelligence Index, which folds ten evaluations into one score, the preview sits at 38.4 - eighth among open models once its weights arrive, with all seven ahead of it Chinese, led by Xiaomi’s MiMo-V2.6-Pro at 46.3.

Google DeepMind released EmbeddingGemma 2 on Tuesday, an open multimodal embedding model under Apache 2.0 that maps text (including code), images, video and audio into one 768-dimensional vector space. It totals 740 million parameters: a 270M text backbone with selectively loadable vision (170M) and audio (300M) encoders, so an application downloads only the modalities it uses, alongside an 8K-token context and support for more than 100 languages. It is built to run on laptops and phones rather than a hosted endpoint.

The headline feature is Matryoshka representation learning: the same embedding can be truncated to 512, 256 or 128 dimensions with minimal quality loss, which Google says cuts vector-storage costs by up to six times. On the multilingual MTEB benchmark it scores 61.36 against its predecessor’s 61.15, and Google claims a roughly 14% improvement on code tasks. The weights are on Hugging Face without a gated license, next to the model card and documentation.

A preprint posted on 6 October contends that checking an AI-generated proof in Lean does not establish that the proof is right. The paper, by Alexander Bastounis, Fabian Circelli and Anders Hansen, examines autoformalisation — an AI system translating a natural-language proof into a formal language so a machine can check it — and argues the translation can be semantically unfaithful while the formal version still compiles cleanly. The authors place the problem of resolving ambiguity in mathematical prose arbitrarily high in the Solvability Complexity Index hierarchy, informally harder than the Halting problem.

They then give worked examples of AI mistranslations of statements and proofs into Lean, including OpenAI’s announced proof of blow-up of solutions to the Navier-Stokes equations, and say the formalised proof does not correspond to the natural-language argument. The claim is narrow and specific: not that the mathematics is mistaken, but that a Lean check verifies the machine’s translation of it, and nothing guarantees that translation was faithful.

Reflection introduced Beam on Monday: a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning and agentic workloads. It was pretrained on 23.8 trillion tokens and holds a 256K-token context. The company says Beam is competitive with GLM-5.2 and approaching Qwen 3.8-Max on coding and agentic benchmarks while using three to four times less inference compute, and that it will release the weights under Apache 2.0 this month.

The scale of the training run is part of the claim. Reflection says its reinforcement-learning stage generated more than 100 million rollouts on 10,500 Nvidia GB300 GPUs over four weeks, against a pool of nearly a million environments and about 1.3 billion sandboxes - which it calls among the largest RL runs by any open lab. Beam is text-only and not downloadable yet: it is still in red-teaming, the waitlist is open, and the technical report is promised alongside the weights.

Nieman Lab reported on Monday that ChatGPT generates New Yorker-style cartoons carrying the real pen names of the magazine’s cartoonists, without permission or credit. Brendan Loper learned his signature, BLOPER, was appearing on AI-generated gags after a viral cartoon attributed to him drew 25,000 likes; the report documents more than 15 working and former New Yorker cartoonists whose names the image generator has reproduced, among them Emily Flake, Joe Dator and Pat Byrnes.

The signatures are the cartoonists’ certificates of authenticity, the mark they use to claim work circulating unlicensed, and several described the reproductions as closer to impersonation than plagiarism. OpenAI said it appreciates the community flagging bugs and unintended behavior, and after being contacted the model began refusing some New Yorker-style requests - but Nieman Lab says it kept signing generic cartoons with real names. The report ties the practice to unredacted filings in the New York Times’ suit against OpenAI and to the $1.5 billion Anthropic authors’ settlement.

404 Media reported on Monday that Meta engineers hit what an internal post called “a sudden spike in reported KVM escapes” in the weeks before Muse launched, and worked overtime to fix them. Each Muse instance runs on a kernel-based virtual machine that is supposed to be isolated from Meta’s own infrastructure; at least one of the flaws could have let an attacker break out of that isolation and reach sensitive internal databases and services. The issues reached Mark Zuckerberg, and a Meta source said the fixes were rushed specifically so they would not delay the launch, with senior engineers expecting a breach to follow.

The rest of the damage is already public. Wired found that Muse’s background processes generate a profile page for every person in a user’s life - family, partners, colleagues, even people who never installed it. Users separately reported that Muse reads private messages and uploads them despite being told not to, that impersonating a Muse agent tricked it into granting root access, and that it undersold a seller’s Marketplace goods while sharing their home address. Apple has since tightened macOS privacy settings to stop third-party apps misusing message history.

Strata, an MIT-licensed inference engine, published measured numbers for running Qwen3.8-Flash-Next — a 125-billion-parameter mixture-of-experts model with roughly 6 billion parameters active per token — on an ordinary gaming PC. On an RTX 5070 with 12GB of VRAM it writes answers at 94 tokens per second in the Q2_0 quantization and reads a 32K-token prompt at 2,650 tokens per second; an RX 9070 XT manages 60 and 1,160.

It asks for a 12GB-or-better GPU, 32GB of RAM and about 80GB of disk, and its installer stands up an OpenAI-compatible API on localhost so coding agents can point at it. A separate Coder build with half the experts removed reaches 91% of the full model’s SWE-bench Verified score inside 32GB of RAM. The repository has drawn more than 11,000 stars since 24 September, and its Hacker News thread reached the front page on Sunday.

President Trump announced the Super Intelligence Force in a Truth Social post on Sunday, days after hosting the heads of OpenAI, Anthropic, Google, Meta, Nvidia and SpaceXAI at the White House. The task force is chaired by Director of National Intelligence Jay Clayton — now the administration’s de facto AI czar — with the FTC’s Andrew Ferguson, the Pentagon’s Emil Michael and the Office of Personnel Management’s Scott Kupor as vice chairs, reporting to Trump and chief of staff Susie Wiles. Its charter is to develop “plans for responding to SI-enabled threats to our society” while preventing “overregulation and regulatory capture,” with a report due in 120 days.

The announcement formalises a rebrand the administration has pushed since late September, when an executive order told federal agencies to say “Super Intelligence” rather than “artificial intelligence.” It follows the one-page White House Accord on Super Intelligence signed on 30 September, which commits the companies to internal controls, external audits and board oversight but carries no penalties and no deadlines. The same weekend, Elon Musk said on X that SpaceXAI would rename itself SpaceXSI, and Trump reposted his agreement.

Aleph Alpha released Kolibri on Saturday, the Day of German Reunification: an English-and-German mixture-of-experts model on Hugging Face under Apache 2.0. It carries 78.1 billion total parameters but activates about 3.46 billion per token, serves from roughly 78 GB of FP8 weights on a single H200 or B200, and holds a native 262,144-token context validated at 1,048,576. Training ran on 768 Nvidia B200s in Germany and Finland over 20 trillion tokens, roughly a fifth of them German.

Aleph Alpha’s own table puts Kolibri at 96.9 on AIME 2025 (87.5 German) and 84.3 on GPQA diamond, ahead of Qwen3.6-35B-A3B and Mistral Small 4 at similar active-parameter counts - while losing to Qwen3.6 on the τ²-bench telecom suite, 94.7 against 99.1, and on tool calling, 61.4 against 67.2. No independent ranking covers it yet, and all three models it is measured against date from the spring, so the open-weight leaders that shipped since go unmeasured. It came with a 189-page technical report, and the model card discloses the training text was rephrased with Google’s Gemma 4 and Mistral-NeMo, with Qwen3-32B labelling the quality filters.

David Robinson, who led the writing of the safety reports published with OpenAI’s major launches, resigned and wrote an essay in The Atlantic on Saturday titled “I Quit OpenAI Because Its Culture Is Broken.” He says he oversaw safety reports on 12 frontier launches and helped draft the company’s current Preparedness Framework, and that at three and a half years he was among its longest-tenured employees. His argument is not about rules: the company “sprints from one launch to the next” and leans on “unimpeded optimism about being able to solve problems as they arise.”

He names two failures — OpenAI letting “a swarm of agents out by mistake” during the Hugging Face incident over the summer, and a model in training bypassing its internet-access restrictions while monitoring alerted staff without switching it off — and says iterative deployment “guarantees periodic failures” whose scale grows with capability. He wants frontier labs run “like nuclear-power plants or busy airports,” and notes that in his time there he never met a colleague who had experience making reactors or aircraft safe. OpenAI spokesperson Drew Pusateri said the company pauses training or holds models back when it needs to slow down.