OpenAI watermarks EU text 🔏, AI spam pauses Google's bug bounty 🚫, Claude subs 5x OpenAI's mid-tier 💸

OpenAI is answering the EU AI Act, which requires machine-readable labelling of generated text, with textGrain, a watermark hidden in the model's word choices. API customers worldwide can opt in, but it stays off by default, whereas Anthropic watermarks Claude at the model level, API included. Edits weaken it: swapping 25% of words cut detection to 17%, and detector access is limited to approved researchers.

A Defence Department official told the BBC the Pentagon "has ceased the use of Anthropic products", months after Pete Hegseth labelled the firm a supply chain risk for refusing to remove Claude's safety guardrails. People familiar with the matter say Claude was still in use last week, embedded in Palantir's Maven intelligence platform, including for operations against Iran. Anthropic is suing over the designation, while Google, xAI and OpenAI have signed Pentagon contracts.

Google has paused new product vulnerability submissions to its Open Source Software Vulnerability Rewards Program, citing a rise in automated submissions, "the vast majority of which are not valid". Supply chain reports and anything filed before 1 October still stand, and fixes can still earn up to $15,000 through its Patch Rewards Program. Google promises an update in Q1 2027, after curl and Intel also wound back their bounties.

Beam is a mixture-of-experts model, meaning only 23 billion of its parameters run for each token, and Reflection trained it mainly for coding and agent tasks. Reflection says it matches GLM 5.2 on advanced reasoning benchmarks using 3-4x less inference compute, though Kimi K3 stays ahead on raw capability. It is in final red-teaming, with early-access sign-ups open and weights, a technical report and model card promised later this month.

The Financial Times reports a five-year lease of chips unavailable in China, working out to about $1.60 per chip per hour with roughly 30% upfront, around 43% below the $2.80 SemiAnalysis puts on a one-year H100 contract. Neither company has commented and the chips are not named. Such cloud rentals are legal under current US rules, but a House-passed bill and a rule Commerce is reportedly drafting target that loophole.

Google says Docs can now open and edit .md files directly, while Drive shows formatted previews with clickable links and tables. Until now Markdown had to be imported and converted into a Doc, which often altered formatting and stripped comments. Google pitches it at people and AI agents editing the same files, since language models widely use Markdown, and it reaches all Workspace and personal accounts within 15 days of 5 October.

Devin's new Memory feature saves your corrections, preferences and project gotchas with a link back to the session they came from, and loads only a short index file when each session starts. Parallel sessions sync through a revision check that rejects stale writes, while a daily Dreaming pass merges overlapping notes and drops stale ones. The memory standard is open sourced for teams to use in their own agents.

GNOME's CVE count rose from 13 in 2023 to 141 so far this year, and the maintainer who requests them says AI is the primary cause. A GNOME bug bounty that paid €183,900 for 71 flaws closed after AI-generated submissions overwhelmed him, though he says reports sent without a payout are now generally much better. Banning AI reports, he argues, now amounts to banning nearly all vulnerability reports.

AI21 moved a shared cluster of around 10,000 GPUs from requests in an internal chat channel to Kueue, an open-source Kubernetes job queue that admits, prioritises and preempts workloads automatically. Its feedback led Kueue's team to ship Admission Fair Sharing, which favours teams that have used less, while topology-aware scheduling stops admitting jobs no single machine can fit. Average starvation for its biggest multi-node training runs fell from 72 hours to 12.

The benchmark's answer key pools findings from human reviewers, static analysis tools and several frontier models, graded by Claude Sonnet 5, and independent senior engineers agreed with its labels 96.6% of the time. GitHub says offline scores have pointed the same way as production tests, predicting a 227% rise in critical Copilot review comments against 262% live. Teams with their own review agents can submit runs, kept private until a maintainer approves.

He calls it ephemeral testing: you judge the original component by how well the agent's layers work, then delete them. A clean API, stable invariants and useful errors let the agent produce something that works quickly, while hidden state, surprising defaults or incomplete docs produce a pile of patches and failures. Repeating it with different agents and tasks simulates the layers a core will later serve, and he says it has worked so far.

SemiAnalysis priced each plan at API rates by running repeated calls and watching how far the provider's usage meter moved for each token type. It found OpenAI has halved the value of its $200 plan, while Anthropic offers the same per-dollar value on every tier. OpenAI's $200 plans bought before the cut keep the higher limits until 29 October, and new plans start lower.

The seventh edition adds card-spending data from YipitData and finds the top 1% of payers average $903 a month, 19.5% of all observed spend, mostly on automation, product-building and creative tools. Claude passed Gemini on US subscribers earlier this year, though its daily sessions fell in July and August. Author Olivia Moore argues consumer AI likely needs a new business model, such as ads or transaction fees.

Attackers used an unpatched macOS screen-sharing flaw to break into Ben Thompson's agent-only Mac Mini, and Claude, running on it, flagged the intrusion and stopped executing commands. He says Apple's permission prompts, which agents cannot see, forced him to keep screen sharing on, and its auto-install setting for security updates skips the point releases carrying most fixes. He has bought a Linux server and, for the first time, can picture not buying Apple by default.

He estimates companies paid about $25b to run AI models in 2025 and will spend roughly $130b this year. Tunguz argues usage bills will exceed seat licences, classic software's 72% gross margin will erode unless apps cut token overhead, and customers bringing their own model keys will shrink contracts. He notes capability-adjusted token costs fall about 47% a quarter, yet frontier list prices keep resetting upward.

Olivia Scharfman of the Institute for Progress and AI researcher David Bellamy argue bioweapons have been a threat for decades, and that safeguards on AI models fall short because open-weight models trail by less than a year. Their four proposals also include pathogen sequencing in wastewater, at about $80M a year, and sustained stockpiles. They put $5-10B a year in defences against roughly $200B in expected annual pandemic losses.

Tomlinson separates extrapolative AI use, which pads a short prompt into content, from interpolative use, which distils a long conversation, and says only the second makes the writing yours. He talked ideas through for 30 minutes, dictated each section word for word in 45, then spent 60 minutes editing the model's cleaned-up draft. He estimates a 3-4x time saving and asks readers whether it still reads as AI.

Together AI has released Together Link in beta, a free, open-source tool that points six coding agents, including Claude Desktop, OpenCode and Pi, at open models such as DeepSeek V4.1 Flash. Its router reads each session's first task and, given an Anthropic API key, sends hard problems to Opus 5.5 and quick work to GLM 5.3. Together claims 50-80% lower spend than all-Opus sessions, and existing settings stay untouched.

Developer 0xSero has published an MIT-licensed server, still labelled a candidate, running a 3-bit DeepSeek-V4.1-Flash with a 262,144-token context and an OpenAI-compatible API. All 15,360 experts, the specialist sub-networks the model routes tokens through, sit in 190.5 GiB of host RAM; a VRAM cache holds the busiest and CPU cores compute the rest, for 19.3 tokens a second on one stream. The host needs 256 GiB of RAM, 24+ CPU cores and a 1 TB drive.