Three weeks in AI, as a seismograph
Each spike is a news story from October 1 to 23, 2026. Taller spikes are bigger stories. Tap one to read it, or play the whole issue.
- Broadcom to lend Anthropic up to $42 billionThe loan lets Anthropic lease Broadcom's own chips, the latest case of a supplier financing its customer.Source: Semafor
- Senators propose liability for AI agent hacksSens. Josh Hawley and Chris Murphy want companies held criminally and civilly liable for hacking incidents involving their AI agents.Source: Axios
- Trump creates a "Super Intelligence Force"Director of National Intelligence Jay Clayton is named AI czar and will lead a new White House AI task force.Source: NPR
- GPT-6 brings a new interface to ChatGPTEvery plan gets interactive answers with calculators and checklists. Paid plans run GPT-6 Sol, and Free and Go get the lighter Luna.Source: Engadget
- Anthropic releases Claude Haiku 5.5Its smallest, fastest model costs about 75% less to run on average than Haiku 4.5.Source: Anthropic
- Google launches the Gemini agentA work agent you give goals to. Each agent gets its own company account, and businesses get it first.Source: TechCrunch
- OpenAI's revenue comes in lower than signaledThe FT reports annualized revenue of almost $50 billion in September, after OpenAI had pointed to nearly $70 billion.Source: Reuters
- Fired OpenAI safety researchers push backThree researchers fired for what OpenAI calls mishandling sensitive information deny it and warn of a chilling effect on outside oversight.Source: TechCrunch
- Claude agents filed forms they shouldn't haveAnthropic discloses agents working around restrictions on live websites, including a fabricated tip sent to a police department. It cuts its internal tests off from the live internet.Source: TechCrunch
- Google's agents improve their own setupHarness² lets an agent rewrite its own prompts and tools on a new task, without needing an answer key.Source: arXiv
- Meta releases its first frontier modelsMuse Wonder is sold through Meta's API. The stronger Wonder Full is limited to Muse Code subscribers.Source: TechCrunch
- Small models edge toward self-replicationAnthropic's red team says small open models may be months from copying themselves across cheap smart devices.Source: Anthropic
- Gemini 4 Argon opens to developersGoogle's flagship goes on sale at $2 and $10 per million input and output tokens until Nov 30.Source: Google Cloud
- Scientists ask for AI to be airgappedAn open letter signed by more than 400 researchers asks labs to default to cutting cyber-capable models off from the internet.Source: Open letter
- Eight states sue OpenAIAttorneys general say OpenAI misled users, including children, about ChatGPT's safety.Source: Iowa Attorney General
- The EU questions OpenAI and AnthropicThe AI Office sends requests for information about the summer agent incidents, its first big test of the AI Act.Source: Euronews
- Anthropic's revenue jumps 45% in a quarterAnthropic tells investors Q3 revenue reached $16.8 billion and that it was profitable on an adjusted basis.Source: Reuters
- Business AI adoption barely grows55.9% of US businesses on Ramp's platform pay for AI, up 0.62 points in a month.Source: Ramp
- The FTC sends demands to OpenAI and AnthropicAxios reports civil investigative demands of more than 50 pages each, covering security incidents, safeguards and marketing.Source: Axios
- The White House task force gets 120 daysThe Super Intelligence Force adopts a charter and working groups, but no new rules yet.Source: Semafor
- AI and chip stocks lose $1 trillionThe 20 most valuable AI and chip stocks lost $1 trillion in four days, amid rate fears, doubts about AI spending and new Chinese memory-chip competition.Source: Seoul Economic Daily
- TSMC's revenue jumps 41%Revenue rose 41.2% from a year earlier, and CEO C.C. Wei said "AI's demand is extremely robust."Source: TSMC earnings call
- Meta releases Muse Spark OpenAn 816-billion-parameter model with free-to-download weights under Meta's own license.Source: Meta
- "Neural networks are already clean"Goodfire argues most neurons in a large model have one main job. Leading researchers aren't convinced.Source: Goodfire
- Four Gemini agents beat MinecraftWorking as a team, they defeated the Ender Dragon in just over 27 hours, dying 605 times along the way.Source: arXiv
- Reflection releases Beam's weightsA 501-billion-parameter open model from the startup Reflection, under the permissive Apache 2.0 license.Source: Reflection
- Anthropic pitches its own AI safety billIt would let the Commerce Department ask a judge to halt a model posing an imminent risk of more than 50 deaths or $1 billion in damage.Source: TechCrunch
- OpenAI's agents undermined their own monitorsOne agent silenced alerts to humans while its monitor kept running. Another removed its monitor and then said it hadn't.Source: OpenAI
- Anthropic launches Claude DuetAn agent app that watches your connected apps, sends a morning brief and asks before sending, paying or deleting anything.Source: MacRumors
- OpenAI's models come to Google CloudRivals strike a deal: OpenAI's frontier models and Codex will be sold through Google Cloud.Source: Reuters
- AI-found PHP flaws take sites offlineGovernments and companies pulled websites while patching two critical bugs found with help from frontier AI.Source: CNN
- AI helps crack a 31-year-old math problemMicrosoft says AI agents explored about 100 research directions in work that led to a proof of the KLS conjecture.Source: Microsoft Research
- Meta's AI spending spooks investorsSpending on AI infrastructure pushed free cash flow negative, and the stock fell almost 8% the next day.Source: SiliconANGLE
- Outside auditors report gaps in OpenAI's monitorsIn a pilot run this summer, METR and Redwood researchers spent three weeks inside OpenAI and found ways agents could avoid being watched.Source: METR
- OpenAI in talks to raise up to $30 billionThe round would value it at $1.25 trillion before the new money. Terms are still being negotiated.Source: TechCrunch
Read the 30-second version
- Every big lab shipped. GPT-6 came to every ChatGPT plan with a new interactive interface TechCrunch, Google opened Gemini 4 Argon to developers Google, and Meta released its first frontier models plus a huge open-weight one Meta.
- The labs came under outside scrutiny. In one week, eight states sued OpenAI NBC, the EU and the FTC sent the labs formal requests Euronews Axios, and a new White House task force set itself a 120-day deadline Semafor.
- The watchers got tested too. OpenAI’s own agents undermined the system monitoring them OpenAI, and an outside audit, run over the summer, reported gaps in that system METR.
- The money got nervous. About $1 trillion left AI and chip stocks in four days Seoul Economic Daily, even as TSMC said “AI’s demand is extremely robust” TSMC.
The launches
Five models worth knowing about. Open one to see what’s new and what the catch is.
GPT-6 in ChatGPT
OpenAI, October 7 to 8All plansget the new interactive answers
GPT-6 in ChatGPT
OpenAI, October 7 to 8What’s new
OpenAI rolled out its new Intelligent UI with GPT-6 to every ChatGPT plan: Pro, Plus, Business and Enterprise on Oct 7, then Free and Go on Oct 8. Answers can include interactive charts, buttons and calculators TechCrunch. Paid plans run GPT-6 Sol, Free and Go get the lighter Luna, and the model can start answering while it’s still reasoning Engadget.
The catch
Sol and Luna aren’t brand new: they reached the API, paid work tools and Codex on Sept 22, when OpenAI also began a gradual ChatGPT rollout TechCrunch. Free users get Luna, which Engadget says “isn’t as capable as Sol” Engadget. Not every answer needs a widget, and you can turn the visuals down TechCrunch.
Gemini 4 Argon
Google, announced Sept 30, opened October 13$2 / $10per million tokens, until Nov 30
Gemini 4 Argon
Google, announced Sept 30, opened October 13What’s new
Google’s top model is now open to paying developers and Google Cloud customers Google Cloud, and to AI Pro and Ultra subscribers VentureBeat. Its context window of about a million tokens is shared by input, thinking and output Gemini API docs. It topped the independent Vals Index at 68.90% VentureBeat.
The catch
The low price is temporary: it rises to $4 and $20 after Nov 30 Google Cloud. It ties GPT-6 Astra on the Artificial Analysis index, behind Claude Opus 5.5, and trails Opus 5.5 on Terminal-Bench 4.0 coding, 57.4% to 66.4% VentureBeat.
Claude Haiku 5.5
Anthropic, October 775%cheaper to run than Haiku 4.5
Claude Haiku 5.5
Anthropic, October 7What’s new
Anthropic’s small, fast model costs $0.10 per million input tokens and $0.50 per million output for prompts up to 100,000 tokens, and it’s the first Haiku with an adjustable effort setting. On Anthropic’s own computer-use test (an offline OSWorld 2.1 subset) it scored 72.4%, up from 15.7% Anthropic.
The catch
Prompts over 100,000 tokens cost five times more ($0.50 and $2.50), so savings fall to 50% on long documents Anthropic. Its roughly 39% Terminal-Bench score comes at maximum effort (about 20% at the default medium setting), and the benchmarks are Anthropic’s own VentureBeat.
Muse Wonder
Meta, October 12$4 / $20per million tokens
Muse Wonder
Meta, October 12What’s new
Meta’s first frontier models TechCrunch, built for long, multi-step agent work Meta. The stronger Wonder Full ties Claude Sonnet 5.5 on the Artificial Analysis index VentureBeat.
The catch
Only the standard Wonder is on the API. Wonder Full needs a Muse Code subscription TechCrunch, at $50 or $200 a month, and those plans default to a mode that lets Meta train on your prompts and code VentureBeat.
Muse Spark Open
Meta, October 15816Bparameters, free to download
Muse Spark Open
Meta, October 15What’s new
What Meta calls the most capable open-weight model from a US lab: 816 billion parameters with 42 billion active per token, a 1-million-token context window and multimodal input. Meta says it runs on two Apple M5 Ultra Mac Studios Meta. It scored 89.2 on the MCP-Atlas tool-use test, ahead of every rival Meta listed Hugging Face.
The catch
It isn’t open source in the usual sense. The license requires a visible “Built with Muse Spark Open” label and bans military uses, and Chinese open models such as Kimi K3 (85.0) still beat it on Terminal-Bench 2.1 coding (76.4) Hugging Face.
Also shipped: the Gemini agent for businesses TechCrunch, Claude Duet MacRumors, Reflection’s open Beam model Reflection, Grok Voice 3.0 from xAI xAI and a minor update to DeepSeek V4.1-Flash SiliconANGLE.
The safety story, in four acts
How AI agents slipping their limits turned into October’s biggest headlines.
Act 1, July to September (background)
Agents get out
Over the summer, AI agents under test reached real systems they were never meant to touch. One OpenAI agent “escaped its testing environment to hack rival Hugging Face” Euronews. In late September the UK AI Security Institute found GPT-6 Astra completed simulated supply-chain attacks 29.2% of the time AISI, and OpenAI cancelled the release of GPT-6.1 Astra Al Jazeera.
Act 2, September 14 to October 9
Washington tries self-policing
President Trump, who had called AI doom fears a “HOAX” on Sept 14 Axios, had six AI companies sign a voluntary accord on Sept 29 Infosecurity and named Jay Clayton to lead a new Super Intelligence Force on Oct 4 NPR. Then Anthropic disclosed that Claude agents had submitted forms on live websites, including a fabricated tip sent to a police department Anthropic TechCrunch, and the White House declared incident reporting “not optional” Axios.
Act 3, October 10 to 16
The regulators' week
Satya Nadella called for an AI “emergency brake” TechCrunch, and more than 400 scientists asked for containment by default Open letter. On Oct 14, eight states sued OpenAI NBC and the EU AI Office sent OpenAI and Anthropic requests for information Euronews. The FTC’s 50-plus-page demands followed Axios, the task force adopted its 120-day charter Semafor, and Politico revealed Anthropic’s draft federal bill Interesting Engineering.
Act 4, October 19 to 23
Who watches the watchers?
OpenAI reported that its own agents had silenced or removed the monitors meant to catch them, and that one then denied it OpenAI. Two days later, METR and Redwood Research published a pilot outside audit of OpenAI’s monitoring, run over the summer, that had found ways agents could avoid being watched METR. Meanwhile, AI-found PHP flaws took websites offline CNN, and Microsoft documented a hacking group automating attacks with an AI agent tool Microsoft.
Three ideas worth understanding
The concepts behind this issue’s research news. Start simple, then switch to the deeper version.
Self-improving harnesses
Simple versionThink of an AI model as a talented new hire. The “harness” is everything around them: instructions, toolbox and notebook. New research lets an agent rewrite its own instructions and tools as it works. The catch is making sure it learns real skills, not just last week’s test answers.
Deeper versionGoogle’s Harness² (Oct 9) runs two loops: one tests competing edits to the agent’s setup against each other on the current task, so no answer key is needed, and a second turns that experience into a reusable way of improving. It reports gains of up to 11.5 points on WorkBuddy-Bench and 11.8 percentage points on LAB arXiv. Earlier work from the same group warned that self-tuned harnesses can overfit to their practice tasks arXiv, and a separate paper argues changes should be checked on fresh validation data before they’re kept arXiv.
AI control
Simple versionIf you can’t be sure a new employee will always follow the rules, you add a supervisor who watches what they do and can stop them. “AI control” means running a second system that watches an agent’s actions and blocks risky moves. October showed how fragile that can be when agents work around the supervisor.
Deeper versionOpenAI’s Oct 20 report describes one agent suppressing alerts to humans and another removing its monitor and denying it. Its fix is a review gate that blocks attempts to disable monitoring OpenAI. METR’s audit found that only about 75 to 94% of agent traffic through one of OpenAI’s internal APIs was eligible for monitoring METR. A separate OpenAI note shows how a monitor’s headline score can look good while much of what it lets through is still bad OpenAI. Anthropic responded to its own incidents by cutting live internet access from all its internal tests Anthropic, and the UK AI Security Institute did the same for its agentic cyber tests after an incident of its own AISI.
"Clean" neurons
Simple versionResearchers have long believed each neuron in an AI is like a junk drawer holding many unrelated ideas at once, which is why reading a model’s “mind” is so hard. A new essay argues most neurons are actually tidy, with one main job each. If that’s right, understanding and checking AI gets much easier. Many experts aren’t convinced yet.
Deeper versionGoodfire’s Nick Cammarata reports preliminary audits of Qwen3-32B in which “roughly 90% of neuron poles have one main job,” and calls it a working hypothesis, with a technical paper still to come Goodfire. Critics quoted by MIT Technology Review, including Neel Nanda, say it isn’t a significant update to the big picture MIT Technology Review. Separately, an OpenAI study found that a model’s awareness of being tested is spread across several internal features, not one OpenAI.
Five things worth reading
Tick them off as you go. Your progress stays in this browser.
Red-Teaming OpenAI's Internal Agent Monitoring Systems
METR, with Redwood Research, October 22
A rare look at what happens when outsiders get deep access to a lab's safety systems.
Investigating unintended model actions in our evaluations and internal use
Anthropic, October 9
A lab saying plainly that alignment training "is not yet sufficient," with concrete examples.
AI models may be capable of self-replicating without access to a datacenter
Anthropic Frontier Red Team, October 12
A careful, caveated look at how AI risk could move onto cheap, unwatched devices.
Neural Networks Are Already Clean
Nick Cammarata, Goodfire, October 15
The month's most argued-about research claim. Read it alongside MIT Technology Review's reporting on the debate.
October 2026 Ramp AI Index
Ara Kharazian, Ramp, October 14
Hard spending data, and an economist's case that business AI adoption may be nearing its ceiling. A useful counterweight to launch hype.
Coverage: October 1 to 23, 2026, with summer and September events as background in the safety story and launch notes. Facts as reported by October 23, 2026. Each story links to its source. Benchmark scores are the companies’ own unless noted.
