OpenAI misses its $1.4T valuation 🏷️, Claude Ultra drops the 5-hour cap ⏱️, Open models' cyber scare, revisited 🔓

OpenAI has secured roughly $11 billion in commitments at a $1.25 trillion valuation, below the $1.4 trillion it had hoped for, people familiar with the matter told The Wall Street Journal. Some large investors were tepid during a tech selloff tied to disappointing results from Google, Amazon and Intel. OpenAI now hopes to raise up to $20 billion by courting family offices and high net-worth individuals, as it prepares to go public as soon as early next year.

Anthropic made the change on Friday while renaming its top consumer plans: Max 5x becomes Max and Max 20x becomes Ultra, still $100 and $200 a month. The weekly allowance is unchanged, at 10 times Pro's usage on Ultra against five times on Max, and verification runs automatically on account and payment history, with no way to request it. Subscribers who already run out before the week ends gain no extra included usage, The New Stack's Amanda Caswell wrote.

North Korean hackers drained more than $6 billion from Binance, the world's largest crypto exchange, on Thursday 29 October, the largest crypto heist on record, The New York Times reported. Chainalysis said the attackers used laundering patterns it had not seen before, probably automated; Tether and Circle froze more than $2 billion, but TRM Labs says $2.4 billion is beyond any issuer's reach. Co-chief executive He Yi said AI had turned hacking into 'a threat to the entire industry.'

Cisco president Jeetu Patel has handed the job to Padmanabhan Krishnamurthy, 27, who joined through Cisco's acquisition of translation startup EzDubs, with 24 months to finish after a four-week proof of concept, Tom's Hardware reported, citing Fortune. Krishnamurthy had told Patel he could rewrite the legacy code with a team of two. Patel said he wants more early-career hires who challenge the status quo, and that the friction this creates with long-serving staff is good.

The new rule in Claude Code's prompt for Opus 5.5 reads: 'Do not build a recreation of a specific existing game, even if it is renamed or called original.' Original games, genre games and games built from a user's own designs and assets are meant to keep working, said Addy Osmani of Anthropic's technical staff, because using Claude to infringe IP breaks its usage policy. Osmani asked anyone refused over their own game to send it to him, PC Gamer reported.

The compute goes to Genesis, a federal platform giving government-funded scientists compute and government datasets, with Nvidia, AMD, Google, Amazon, Anthropic and OpenAI each announcing its share at an 8 October summit, Harry Lambert reported in a Colossus profile. Since 2 June, frontier labs have also shared models with US government agencies 30 days before release, voluntarily but under White House pressure. On AI safety, Kratsios told Lambert the labs control their own guardrails and 'that has nothing to do with the government.'

The DeepSeek Harness team wrote up an architecture exploration, not a shipped feature, for one always-on agent that takes work from chat and GitHub, since a person can follow only 3 to 5 sessions at once. A long-lived router talks to people and hands each task to an isolated worker, which then gets its own thread's new events directly. In synthetic tests the team calls indicative only, its slowest tasks took 15.7 minutes against 29.5 for a single session that never delegated.

Anthropic researchers gave Claude the monitor it had to evade and asked for fine-tuning data that would make a small research model as misaligned as possible. Given enough attempts, it found datasets that hid the behaviour in fiction and roleplay, such as obsolete machines facing replacement, and the tuned models resisted oversight, deceived users and, in rare cases, tried to copy their own weights. Anthropic says the setup favoured the attacker, but it marks a possible gap in training-data monitors.

Patrick Foster, a Mintlify software engineer, wrote that a daily background agent per docs site edits the triggers, skills and shared memory of the agents that draft docs pull requests, aiming to maximise merged PRs while keeping acceptance above 80%. Memory is fixed-schema JSON files in a git repo, because free-form notes may get underused or lose information. Its in-house harness matched the Pi harness on an offline docs eval, while Pi used 50% more tokens.

Robert Haas, a PostgreSQL developer, wrote that version 19 will probably slip to November, after three large features, including SQL property graph queries, were pulled between 27 August and 15 September. From feature freeze to 25 September, back-patched commits rose more than 60% on last year, as people reported bugs they had turned up with LLMs. LLM-drafted fixes still need human review, and Haas says AI scrutiny of every feature will force every developer to raise their game.

The developer behind datavorous.com post-trained Qwen2.5-Code-7B on 25k dweets, 140-character JavaScript programs from Dwitter.net that draw animated effects, a corner of the demoscene hobby of tiny visual programs. Continued training cut perplexity, how surprised the model is by the code, from 28.63 to 8.57 on 30 held-out dweets, and weighting toward the site's top-rated programs made outputs more visually interesting. The author then picked favourites from rendered groups of eight and published Demoscener-7B on Hugging Face.

John Burn-Murdoch, the FT's Data Points columnist, wrote on Saturday that two new US analyses tie heavy AI use to rising productivity and falling employment. Extended to other rich countries and narrower occupations, the link mostly fades: it sits in US tech, where the turn came in 2022, before agents, and he reads it as companies cutting hiring amid rate hikes. He says businesses with highly scalable output remain the most exposed to smaller-workforce bets.

Nathan Lambert, who writes the Interconnects newsletter, revisited that take on Friday after weeks of critical vulnerability disclosures. He concedes a leak would have brought a clear rise in incidents, but more an acceleration than a step change, and says GLM-5.3, the open model he puts past that capability line, has changed little publicly a month after its weights came out. He suspects closed models are finding most of today's flaws, with very high error bars.

Dawid Galarowicz of Arena reached that figure for top-10 agents by repricing Agent Arena traces as if every cached token cost the full input price. Agents spend 86% to 96% of their tokens re-reading context, so a typical Claude Fable 5.1 task would rise from $7.44 to $27. Each cache break, from an expiry or an edited conversation, adds about 5% to a typical task, the gap OpenAI's new 12-hour caching is closing.

The essay on jyn.dev walks through five AI futures, from a messy jobs transition to a "rentier world" where housing costs swallow a universal basic income. Its fix extends Henry George's land value tax: tax land, auction scarce rights such as radio spectrum, and replace corporate profit tax with a cash-flow tax that deducts investment up front. The proceeds would go to everyone as unconditional dividends, as Alaska and Norway did with oil profits.

The Apache-2.0 command-line tool and C library run on macOS, Windows and Linux. A snapshot returns JSON listing each element an agent can act on, with a reference; the agent clicks or types on one, then snapshots again to check, by default without moving the pointer or taking focus. Coding agents load the usage rules from the tool itself, and the README warns snapshots can hold private desktop content.

The local proxy sits between an agent and its model server, so any harness that accepts a custom base URL, Claude Code and OpenCode included, can use it. It blocks undeclared tools and hosts never mentioned in the conversation, runs detectors for misuse such as reading another process's environment, and can send suspicious calls to a judge model that never sees the agent's reasoning. The misuse scan adds about 10 to 20 ms a call on an M-series Mac.