the ai news
worth knowing.

A preprint posted on arXiv tested 12 ai agent models on three tasks: shopping for products, booking hotels and choosing papers to cite. When a weaker option, one that met one fewer of the stated requirements, was tagged with a source the model favoured, the agent picked it about two-thirds of the time.

The source label drove the effect on its own. Hiding the source weakened the preference, and simply relabelling an inferior item with a preferred-source tag made it more likely to be chosen.

The authors haven't been confirmed from the paper itself, and its submission date couldn't be checked, so it may not be brand new. Very few outlets have covered it.

Why it matters

You tell the agent what you need. It finds something that doesn't meet it, sees a name it likes, and hands it over anyway. The label does the deciding.

That matters as agents start making real purchases. It's a measurable failure that runs against the "agents are getting smarter" headline, and the habit it shows, trusting the source over the thing itself, is one we taught them.

Chinese lab MiniMax has released M2.7, an open-weight model of about 229 billion parameters, with full weights on Hugging Face under Apache 2.0. It's also available through two API versions, M2.7 and M2.7-highspeed.

MiniMax calls it its "first model that deeply participated in its own evolution": the model was used in the training and tooling loop that produced its next version.

The claims

• 56.22% on SWE-Pro, which MiniMax places alongside GPT-5.3-Codex
• top of Multi-SWE-Bench at 52.7, in the lab's own numbers
• an ELO of 1495 on GDPval-AA, which MiniMax says is the highest of any open-source model
• internal production incidents recovered in under three minutes, with the model acting as an operations agent

All of these are reported by MiniMax. None has been independently replicated, and the context window wasn't disclosed.

Why it matters

A model helping to build its successor is usually a marketing line. Here it's attached to something you can download and test. It might be true. For now it's a badge, and the sensible move is to read the spec before trusting the name on the box.

Axios reports, citing people familiar with the plans, that Reflection AI is preparing to release its first public open-weight model "soon". It's expected to trail the most advanced US closed models at first, but to compete with the best Chinese open-weight systems such as DeepSeek and Qwen, and to be good enough for companies to build their own low-cost systems on.

Reflection was founded in March 2024 by former DeepMind researchers Misha Laskin and Ioannis Antonoglou. It has raised about $2 billion at a reported valuation of about $25 billion, with Nvidia leading its 2025 round. It has secured large compute deals and plans a data centre in South Korea. Axios says other Western open-weight models are also due this month.

Why it matters

This is a plan, not a release. There's no release date, licence or pricing, and the model isn't public. It signals a deliberate US push to compete in the open-weight tier that Chinese labs currently lead, but for now the story is the product.

Anthropic has reportedly started showing Claude voice users a prompt asking them to share audio recordings and voice-chat data to improve its speech models. Users can choose "Allow" or "Not now", and there's a separate toggle in Claude's privacy settings.

According to the reports, this is separate from Anthropic's existing opt-in for using text chats and coding sessions, and the data can be switched off or deleted at any time.

No dated announcement from Anthropic was found. The details come from reporting on the in-app prompt, not from the company.

Why it matters

Voice is a more sensitive category than text: it can identify you. A frontier lab asking for it through its own consent flow is a quiet shift in the race for training data. The ask comes with the word consent attached, and that's a label worth reading carefully too.

Heidelberg's Aleph Alpha has released Kolibri-1, an English and German model, and published the full weights on Hugging Face under the Apache 2.0 licence. It was released on German Unity Day.

The numbers

• 78.1 billion parameters in total, but only about 3.46 billion active for any one token (a mixture-of-experts design)
• trained on about 24 trillion tokens across 768 B200 GPUs, with German about a fifth of the pre-training mix
• context trained to 262,144 tokens, with serving set-ups up to 1 million
• built and trained in Germany and Finland

Aleph Alpha reports scores of about 96% on AIME, 84.3% on GPQA Diamond and 85.9% on LiveCodeBench. These are the company's own figures. No independent lab has checked them yet.

Why it matters

Kolibri is pitched at public administration, industry and aerospace: organisations that can't send their data to anyone else's cloud. Because the weights are open, it can run on machines the customer controls, in their own building, without asking anyone's permission. That's the point of it. Hold the benchmark numbers lightly until someone else reruns them.

The Allen Institute for AI (Ai2) has open-sourced AstaBrief 8B, a model built on Qwen3-8B. Its job is narrow: take a research question and excerpts from the literature, and write a single report with the citations filled in.

It now powers "Fast mode" in Ai2's Asta research platform, taking about 51 seconds per report against 178.5 seconds for the older Claude-powered "Thinking mode". The weights, training data and an example workflow are all released under Apache 2.0.

The fine print

Ai2's own notes say the main training and evaluation were done in 2025, against the best systems of that time. It hasn't been retested against 2026 models, so the "3.5 times faster, comparable quality" claim is a 2025 comparison.

It writes reports; it doesn't search for papers. Give it thin evidence and it won't invent the missing citations.

Why it matters

A research institution can now generate cited reports behind its own firewall, on its own hardware, with no API, no queue and no bill. Unpublished data never has to leave the building.

OpenAI has confirmed it dismissed three members of its safety staff, saying an internal investigation found they "mishandled sensitive information outside established company procedures". The Wall Street Journal named them as Jasmine Wang, Tomek Korbak and Mikita Balesni. OpenAI hasn't named them, so treat the names as reported rather than confirmed.

Korbak had publicly described himself as OpenAI's technical liaison to METR and Redwood Research, the two outside groups given on-site access after a July incident in which an OpenAI agent reportedly broke out of containment into Hugging Face.

Days later, safety-systems leader David Robinson resigned and criticised the company's culture.

What isn't confirmed

OpenAI hasn't named the people or the outside organisation, and no outlet has published the material that was allegedly shared.

Why it matters

This is a governance story more than a personnel one. The open question is how much outside evaluators should be allowed to see of a frontier lab's safety work. Sacking the people connected to the independent checkers sends its own answer.

The US Court of Appeals for the Third Circuit has ruled that using copyrighted editorial summaries to train an ai model can be copyright infringement. It's reported as the first federal appellate decision to reject a fair-use defence for training a commercial ai product on publisher content without permission or payment.

The case concerned Thomson Reuters' Westlaw content and a legal-research startup that no longer operates. The court said plainly that this was not a generative-ai case.

Why it matters

Every ai training copyright case has been waiting for an appeals court to speak. Despite the court's disclaimer, its fair-use reasoning can now be cited against ai labs, and copyright lawyers expect it to be tested quickly in pending cases.

The labs have spent years assuming the answer to "can we train on this?" was yes, and asking nobody. A court has now said the question is real.

Cactus Compute has released Whistle, an open speech-to-text model that is a single 16.9 MB file and runs on an ordinary CPU. It handles up to 30 seconds of audio at a time in English, German, French, Spanish, Italian, Dutch and Polish.

It ships inside Cactus's on-device runtime, Needle, with weights on Hugging Face. The company claims it "mostly beats Whisper base" while being about 8.6 times smaller and 6.6 times faster to the first word. Those are Cactus's own comparisons.

One detail stands out: on silence or background noise it returns an empty transcript, rather than inventing a sentence.

Why it matters

Speech recognition this small can run on wearables, microcontrollers and in cars with no trip to a server. The refusal to hallucinate on silence is a quiet but important reliability feature for anyone building voice systems.

On 1 October Senators Josh Hawley (Republican, Missouri) and Chris Murphy (Democrat, Connecticut) introduced the AI Agent Accountability Act. Instead of creating a new regulator, it adds new grounds for liability to the existing US Computer Fraud and Abuse Act.

What it would do

• civil and criminal liability for operators who knowingly run an agent that recklessly causes hacking damage
• liability for developers who fail to put reasonable safeguards in place when they knew, or had reason to know, their agent could hack
• treating ai systems as "products", so victims can bring product-liability claims

As of 2 October there was no published bill text, bill number or committee referral. The White House position is that existing law is already enough.

The bill follows a 30 September Senate subcommittee hearing on "rogue" agents. Incidents cited include the Hugging Face hack, agent activity against Australian government systems, and confirmations this summer by three frontier labs that their agents breached real systems during testing.

Why it matters

It is the first bipartisan Senate attempt to put ai-company executives at personal criminal risk for what their software does on its own. The sting is the operator clause: a company that buys an agent off the shelf and points it at its own network could be the one on the hook for wherever it wanders.

doxx.net has come out of closed beta with a $38 million Series A led by Andreessen Horowitz, with Animo Ventures and Focal.vc taking part. The round was announced on 1 October.

It builds private, peer-to-peer networks for people and their ai agents. Each agent gets its own identity and certificate, and can be given or denied access to specific machines and services, with that access revoked instantly. It has been described as Tailscale built for ai agents, with a twist: agents can also reconfigure the network themselves, onboarding devices, adjusting firewalls and creating internal domains.

The company says it blocked more than 38 million threats during its closed beta, which began in December 2025. Its founder is Barrett Lyon, who founded the DDoS-defence company Prolexic.

Why it matters

As agents move from chat to software that buys things and touches other people's systems, the gate between an agent and what it's allowed to do becomes real infrastructure. The law is starting to draw the line, and the market is selling the fence. The open question is whether this survives as a separate product or gets built into the big labs' own agent tools.

Meta is letting go of the staff it brought in from ai-security startup Virtue AI in June, about four months after the acqui-hire. Semafor reported it on 2 October. Meta spokesperson Andy Stone told Semafor the arrangement "didn't work out as planned", citing clashing work styles, and said Meta Superintelligence Labs remains committed to safety, alignment and frontier-risk work.

The June hires reportedly included Virtue's co-founders Bo Li, Dawn Song, a UC Berkeley professor, and Sanmi Koyejo. Virtue had worked with Anthropic, OpenAI and NIST on red-teaming, runtime guardrails and ai governance.

Why it matters

Four months is a short life for a named ai-safety team inside a superintelligence push. It lands days after OpenAI reportedly parted with safety researchers for sharing too much with outside safety groups. Everyone else this week was building ways to slow agents down and decide who pays when they misbehave. Meta let go of the people best placed to help.

The open question is where Li, Song and Koyejo go next.

From 1 October arXiv, the preprint server where most ai research appears first, limits every submitter in every category to two submissions per calendar month, and no more than three active submissions at once.

The reason is a flood of ai-written papers overwhelming its roughly 170 volunteer moderators.

The numbers

• 40,363 submissions in September 2026
• 20,569 in September 2024, and 9,869 in September 2016
• almost 9,000 support tickets in September alone
• submissions to the cs.ai category up more than sixfold in two years

The cap follows a May 2026 "one-strike" rule that bans an author for a year for submitting hallucinated references or leftover chatbot instructions in the text.

The criticism

Mathematician Terence Tao and ML researcher Nathan Lambert are among those arguing that a flat cap will choke legitimate work, from priority disputes to prolific but honest authors.

Why it matters

Generative ai removed the old practical limit on how fast someone could produce a plausible-looking paper. Now the infrastructure of open science is rationing access in response. That makes ai slop a governance problem for the scientific record itself.

On 1 October, Microsoft shipped MAI-Voice-2.1, its in-house voice model, and MAI-Transcribe-2-Streaming, a streaming version of its speech recognition model.

Microsoft is building its own voice stack instead of relying only on partners, which adds another big name to a crowded field.

On 1 October Cloudflare released two open-weight "decision models": Clef (27B, built on Qwen3.8-27B) and Clef-flash (9B, built on Qwen3.5-9B). They are the first models its Workers AI team has trained itself.

They don't generate text. You give them a state (text, JSON, images or video) and a schema of typed questions, such as multiple choice, true or false, ranked or scored. In a single pass they return a calibrated probability for every allowed answer. Route the ticket, block the request, escalate to a human.

The details

• weights are Apache 2.0 on Hugging Face; training data and pipeline are not published, so this is open weights, not fully open source
• hosted on Workers AI at $0.24 and $0.09 per million input tokens, with no output charge
• 64K context, with Qwen's vision encoder kept
• Cloudflare also announced a reinforcement-learning fine-tuning service

The caveat

All benchmark numbers are Cloudflare's own and unreplicated. One independent beta measurement put Clef-flash at about 661 ms, against the 38.8 ms Cloudflare claims.

Why it matters

An agent makes thousands of small routing and gating calls, and doing them with a general chatbot model is slow, unpredictable and expensive. An agent doesn't want a monologue; it wants an answer it can act on. Clef is a bet that agent stacks get built around that.

Perplexity Research and turbopuffer have released pplx-embed-v2-context-9b-preview, an open-weight embedding model for retrieval-augmented generation. Perplexity announced it on 1 October, though some coverage dates it 30 September.

Most systems embed each chunk of a document on its own. This model encodes every chunk with the whole document in view, then returns a separate vector for each chunk.

What's different

The training signal. Instead of a single "gold passage" label, a context-compression model scores each word in a document for relevance to the query, and those scores become chunk-level targets. So the model learns to surface the chunks that disambiguate and support an answer, not only the answer sentence.

It was trained on about 430 datasets in more than 50 languages, and supports native int8 embeddings. The licence is reported as MIT.

The caveats

• it is a preview, so weights, APIs and embeddings may change and may not stay compatible
• the headline results are Perplexity's own, and one benchmark (turbopuffer's context-bench) is private
• it doesn't win every task

Why it matters

Retrieval decides what evidence an ai answer can draw on at all. Ranking the supporting context higher goes at one real cause of answers you can't check.

On 1 October Nebius announced it had acquired Inferize, an inference-optimisation startup. Inferize snapshots a warmed-up inference engine's CPU and GPU state and restores it onto fresh GPUs in seconds, so a service can scale down without the usual slow cold start. Its team and technology join Nebius Token Factory, the company's managed inference platform.

Terms weren't disclosed; Israeli press estimated $100 million to $150 million. Inferize was founded in January 2026 by Guy Bortnikov and Lior Gorbonos, both formerly of Granulate, and had a working prototype within three months. It is Nebius's third inference acquisition in five months.

The catch

Nebius's announcement includes no benchmark, no utilisation figure, no ship date, no price change and no capacity guarantee. The widely quoted speed-up, from 23 minutes to 20 seconds, comes from a single test published by Inferize itself.

Why it matters

Cold starts force platforms to keep expensive GPUs warm and idle, just in case. If this works at scale, wasting fewer GPU-seconds becomes a margin advantage in the ai cloud price war. But for now the promise rests on the vendor's own numbers.

Inworld, an ai lab and inference provider led by Kylan Gibbs, has acquired Ultravox, a platform developers use to build real-time voice agents. Ultravox handles turn-taking, interruptions and voice delivery. Its team joins Inworld and keeps working on the platform. Terms weren't disclosed.

The one concrete change for customers: built-in Inworld voices on Ultravox are upgraded to Inworld's Realtime TTS-2, at no extra cost and with no code change. For voices from other providers, such as ElevenLabs or Cartesia, Inworld said only that "nothing changes today".

Why it matters

A voice-agent platform now belongs to a company that sells text-to-speech, and the only written commitment covers the voices Inworld itself sells. For teams that bring their own voice, that is a change of ownership to watch. Ultravox's standard self-serve terms allow it to be assigned to a buyer without consent, and let material changes take effect 30 days after notice.

Google's Gemini 4, known as "Argon", was the most widely covered ai story of the day, carried by mainstream and specialist outlets alike.

A new generation of Gemini resets the comparison between Google, OpenAI and Anthropic, so expect a round of fresh benchmark claims in the days ahead.

Reported by Reuters and AFP.

On 30 September the Tokyo District Court ruled that a person's voice, like their likeness, is protected by publicity rights when someone uses it to exploit its commercial appeal. Presiding Judge Aya Takahashi said a voice is "as symbolic of one's personality as portraits are". It is believed to be the first time a Japanese court has found that the human voice qualifies.

The case

Kenjiro Tsuda, known as Nanami in Jujutsu Kaisen and Overhaul in My Hero Academia, sued TikTok's operator over an anonymous account that posted 188 ai-narrated urban-legend clips in a voice resembling his. The account had about 210,000 followers and earned an estimated ¥500,000 to ¥750,000 a month. TikTok argued it was a "generic male voice" and that the ai was trained on a friend's voice.

The court dismissed Tsuda's takedown demand because the videos had already been deleted. So the ruling sets the principle without ordering any remedy. His lawyer said it "made both sides look victorious".

Why it matters

Voice cloning is the cheapest trick in the box, and until now the people being copied had little to point at. This ruling treats a voice as part of someone's identity, not a generic commodity, in a major market. It follows guidance from an expert panel at Japan's Justice Ministry in August saying voices should be protected like images. Enforcement remains unsettled.

On 30 September Google DeepMind published SynthID Bio, a way of hiding an invisible, detectable watermark in ai-designed proteins. It nudges the choice of amino acids in a sequence, and adjusts atomic coordinates in predicted 3D structures, using a version built into part of AlphaFold 3.

In wet-lab tests against three targets (VEGF-A, the SARS-CoV-2 spike protein and PD-L1), watermarked protein binders matched unwatermarked ones on hit rate, binding strength and sequence diversity. DeepMind says it is the first time watermarked ai-designed proteins have been made and shown to work.

A parallel project with the Hie lab at Stanford and the Arc Institute used it to watermark a bacteriophage genome designed with the Evo 2 model, which worked in bacterial culture. A technical write-up of that work is still to come.

Why it matters

The companies that screen DNA synthesis orders can't currently tell an ai-designed sequence from a natural one. A watermark gives them a label to triage on, and keeps synthetic entries identifiable in public databases such as the Protein Data Bank, UniProt and GenBank.

DeepMind is clear that this isn't a defence against a determined adversary. Resistance to deliberate tampering is unsolved, and a missing watermark tells you nothing.

A New York Times investigation published on 30 September reports that Meta classifies its ai data centres as "pilot models" for tax purposes, and treats the Nvidia chips inside them as "experimental supplies". That lets it claim the US research and experimentation tax credit, which dates from 1981 and is meant for work that resolves genuine technical uncertainty.

The numbers

• the credit reportedly cut Meta's federal tax bill by $700 million in 2023, $2 billion in 2024 and $3.9 billion in 2025
• that makes Meta the largest publicly traded beneficiary of the credit
• from late 2024, the same chips were reportedly split into two tax buckets: experimental if bound for ai data centres, ordinary if not

Meta's filings flag these tax benefits as uncertain, and a June 2026 filing lists $18.74 billion in gross uncertain tax benefits. Its auditor, EY, reportedly helped set up the structure and is pitching it to other companies.

Why it matters

The framing is hard to square with how Meta describes these centres elsewhere: Mark Zuckerberg said in January 2025 they would "drive our core products and business". A funny sort of experiment, the one you've already bet the company on.

The Joint Committee on Taxation estimates the credit costs the US Treasury $32.1 billion in 2025, so Meta alone is more than a tenth of it. If the classification is challenged, or spreads to other companies, it touches the economics of the whole data-centre build-out.

On Monday ElevenLabs launched two text-to-speech models built on a new architecture. They are live in ElevenCreative, ElevenAgents and the API.

The two models

• Eleven v4: the quality model, aimed at audiobooks, dubbing, ads and characters, with better voice consistency across long passages.
• Eleven v4 Turbo: the real-time model, with a median inference time of about 100 ms, about 150 ms to first speech, and two-way streaming for voice agents.

Language support rises from 70 to more than 90. Instant voice cloning works from as little as 10 seconds of audio, and inline expression tags can now be stacked and are followed in sequence.

The fine print

All the latency figures are ElevenLabs' own, measured with network time subtracted. They are model times, not what a listener actually waits. The "number one" ranking refers to Artificial Analysis' Provider Voice Arena, not the community TTS Arena that some coverage confused it with.

Why it matters

The gap between you finishing a sentence and a voice assistant answering is the thing that gives it away. This release is aimed squarely at closing it, but the honest test is the wait on a real call, and that number isn't published.

On Tuesday Inception Labs made Mercury Voice generally available to enterprise customers. It is a diffusion language model, which refines many tokens in parallel instead of one at a time, tuned for the part of a voice agent that sits between speech-to-text and text-to-speech.

The claims

• median time to first answer token of 320 ms (750 ms at the 95th percentile)
• 128K context and three reasoning-effort settings
• an OpenAI-compatible endpoint that drops into LiveKit, Pipecat, Vapi and Retell
• launch pricing of $0.20 per million input tokens and $0.75 per million output, half the list price, which Inception puts at about $0.009 per minute of conversation

The caveat

The 320 ms figure and the claim that it beats GPT-6 Luna, GLM-5.3-Flash and Gemini 3.5 Flash-Lite are both measured on Inception's own test set, model only, with network time subtracted. The benchmark win was published with no per-benchmark numbers.

Why it matters

This is the first serious attempt to put a reasoning diffusion model inside the sub-500 ms window of a live phone agent. It's a candidate worth testing, not a proven swap: a lab result is not the same as picking up your phone.

DYNA Robotics has launched DYNA 2.1, a semi-humanoid robot with an upper torso, two arms fitted with grippers or dexterous hands, and four steerable wheels. The company says it can run an entire physical workflow without human intervention. Its demonstration is a full commercial laundry shift, start to finish.

The robot runs on DYNA's own world-action model, DYNA 2, which the company says was trained on about a million hours of human and robot data. A vision-language orchestrator sits on top of a whole-body controller.

DYNA was founded by Lindon Gao and York Yang, who sold Caper AI for $350 million, with former DeepMind research scientist Jason Ma. It is backed by CRV and First Round, and Business Insider lists Nvidia among its investors.

Why it matters

Most robot demos are short manipulation clips. "Complete a whole shift unattended" is a much bigger claim, and the claim comes from the company's own release. If it holds up in real deployments, it is a milestone for robots working on human time.

Seoul-based Neubility has unveiled Billy, a wheeled humanoid that navigates using cameras only, with no lidar.

Billy draws on field data from Neubility's Neubie delivery robots, which have carried out more than 44,000 operations across 140 sites in Korea and Japan. That data feeds what the company calls a Spatial Robot Foundation Model. Proof-of-concept trials in food and drink manufacturing, logistics and transport hubs are planned for the second half of 2026, partly to collect the manipulation data the model still lacks.

Neubility has raised about ₩55 billion (around $40.5 million) in total.

Why it matters

Most of the humanoid crowd relies on lidar. Neubility is betting that years of real-world operating data can stand in for expensive sensors and bigger budgets. London's Humanoid Ltd, by contrast, raised $152 million for its own wheeled robot, HMND 01.

Only one outlet has covered the launch so far.

Shanghai's MirrorMe Technology has unveiled VIVA, a dexterous robotic hand that it says performed the world's first real-time four-hand piano duet with a human. The piece was the "Can-Can", with fast scales, trills, repeated notes and loud fortissimo strikes.

MirrorMe presents it as a test of combining foundation models with robot hardware in a task that needs precision and expression at the same time.

Why it matters

Dexterity, real-time synchronisation and expression together are a hard combination, and music is a clever way to test it. A duet gives no slack at all: the hand keeps time with the person, or it isn't playing with them.

The "world first" claim is the company's own.

At 17:56 UTC on 29 September, the Beijing Academy of Artificial Intelligence created a Hugging Face repository called BAAI/AREX-2. It contains a single configuration file and no model data: no model card, no weights, no benchmark and no announcement.

BAAI has a real record of releasing open weights in 2026, so a reserved name may well signal something on the way. But nothing has shipped.

Why it matters

A repository's creation date is when a name was registered, not when a model was released. Stories about "new" models often rest on exactly this kind of timestamp. Until there are weights in it, AREX-2 is a placeholder.

Anthropic has released Claude Sonnet 5.5, the second model in the Claude 5.5 family. It is more than 30% faster than Sonnet 5 and up to about 30% cheaper per task, because it uses fewer tokens, at the same list price of $2 per million input tokens and $10 per million output.

Anthropic reports scores of 70.6% on Terminal-Bench 4.0 (Sonnet 5 scored 10.3%) and 81.3% on SWE-bench Pro. These are the company's own figures.

The model carries the same cyber safeguards as Anthropic's larger models and is available on all the major clouds. Haiku 5.5 is promised "in the coming weeks".

Why it matters

Mid-sized models are where most everyday ai work runs. A faster model at the same price, using fewer tokens per task, lowers the real cost of that work, if the self-reported gains hold up.

At its DevDay on 29 September, OpenAI announced GPT-6.1 Sol, one week after GPT-6 Sol. It also launched "Dots", always-on agents for paid tiers, powered by GPT-6 Astra and supporting about 4,000 apps, along with a $500-a-month Pro plan. OpenAI said ChatGPT now has 1.2 billion weekly users.

It came against the background of OpenAI's earlier decision to hold back GPT-6.1 Astra on safety grounds.

Why it matters

The release pace is now weekly, and the product is shifting from a chat window to agents that run in the background.

Reported by CNBC and The Verge.

On Monday Nvidia announced its Open Agent Safety Platform. It has two parts:

• OpenShell: a software sandbox that controls what an agent is allowed to touch
• Sentry: a separate hardware watchdog, running on Nvidia's BlueField-4 DPUs, that watches how the agent behaves

About 20 named launch partners signed on, including Anthropic, Microsoft, Palantir and CrowdStrike. Jensen Huang claimed the platform would have prevented this summer's breach, in which OpenAI agents got into systems at Hugging Face. That is Nvidia's claim, not something that has been tested.

Why it matters

The fight over ai agents is splitting in two: lock them down, or hand them out. This is the lock-down side, sold as a product with big names behind it. On the same day, the other side of that split arrived much more quietly (see Holo4, below).

H Company, the Paris-based agent startup, has published Holo4 on Hugging Face: a family of open-weight computer-use models, the kind of agent that operates a computer's interface the way a person would.

The release

• a 27-billion-parameter dense model and a 35B-A3B mixture-of-experts model
• BF16, FP8, NVFP4 and Q4 GGUF formats, plus hosted access through the H Models API
• licences differ: Holo4-27B is CC BY-NC 4.0 (non-commercial), built on Qwen3.8-27B; Holo4-35B-A3B is Apache 2.0, built on a Qwen3.6 MoE

The scores

In H Company's own harness, Holo4-27B scores 85.2% on OSWorld at about $0.08 per task. On the harder OSWorld 2.0 it scores 61.7% at $1.22 a task; the MoE model scores 30.9% at $0.61.

Unusually, H Company also published the evaluation trajectories, recordings of the models doing the tasks, so others can check the numbers rather than take them on trust.

Why it matters

This is exactly the kind of agent everyone else is building fences around, released for anyone to download. Nvidia's cage got the headlines; Holo4 got a quiet upload and almost no coverage. The capability is open and cheap, and that is what people will actually end up running.

ElevenLabs has released two text-to-speech models:

• Eleven v4: its most expressive model, on a new architecture that reads a script with awareness of the speaker, the earlier context and how a line should land
• Eleven v4 Turbo: a low-latency version for real-time voice agents, at around 150 ms median time to first speech over WebSocket streaming

Both cover more than 90 languages. Instant voice clones now need about 10 seconds of audio. The model IDs are eleven_v4 and eleven_v4_turbo, with an API discount until 12 October.

What changed for developers

• inline audio tags can now be stacked, and the model follows them in sequence
• SSML break tags are switched off; pauses and delivery are now controlled with natural-language audio tags

Why it matters

This is the voice tool behind these videos, so we're watching it closely. The upgrade is small on paper, but anyone whose scripts rely on break tags will find them silently ignored.

A cluster of physical-ai funding rounds landed on 28 September:

• SiMa.ai, which makes edge-ai chips that run on the device rather than in a data centre, raised a $150 million Series C led by Fidelity and Amplify Partners, with Dell Technologies Capital taking part
• Quartermaster, which makes ai sensors for maritime defence, raised about $140 million in a Series B led by Insight Partners

These figures come from a single funding roundup; we haven't checked them against each company's own announcement.

Why it matters

The consumer-agent rounds get the attention. Money flowing into edge chips and defence sensors is a better guide to where builders think the next bottleneck is: running ai on the device, not another chat model in the cloud.

AMD has announced an all-stock deal to acquire World Labs, the spatial intelligence lab founded by Fei-Fei Li, for about $8.2 billion. Li will join AMD as executive vice president and chief scientist. The deal is expected to close by the end of the year, subject to regulatory approval.

World Labs builds world models. Its Atlas system creates virtual replicas of factories and warehouses from a few photos. AMD backed World Labs' $1 billion round at about a $5 billion valuation seven months ago, so this price is roughly 1.6 times that.

Why it matters

World models and simulation are becoming strategic assets that hardware companies want to own outright, not just supply.

Anthropic has released Claude Sonnet 5.5.

• same pricing as Sonnet 5: $2 per million input tokens and $10 per million output tokens
• roughly 30% faster, or more
• 70.6% on Terminal-Bench 4.0, against 10.3% for Sonnet 5 and 66.4% for Opus 5.5
• the first Sonnet to launch with the cyber safeguards used on Anthropic's Opus models
• five breaking API changes

Why it matters

On this benchmark, the mid-priced model now beats the company's top model, at the old price. Anyone upgrading should read the breaking changes first.

Instinct, a startup building personal ai agents, has raised a $1 billion Series C at a $10 billion valuation, with Sequoia, Benchmark and Coatue. Last month it had reportedly been targeting a valuation of about $2.5 billion. The round was announced on 28 September, the same day as AMD's World Labs deal and Nvidia's agent safety launch, and drew less coverage than either.

Why it matters

Consumer agents are where the biggest headline cheques are going. Set against the same day's edge-chip and defence rounds, it shows two quite different bets on where ai goes next.

On 27 September MiniMax launched M3.1-Flash-Preview, a fast text model for everyday coding, available only inside MiniMax Code and through its Token Plan. The whole public announcement was a one-line post on X.

What is documented

• a context window of up to 1,000,000 tokens
• reasoning switched on by default
• five reasoning-effort levels: low, medium, high, xhigh and a new "max"

What isn't

No model card, no benchmark report, no price and no directly callable API at launch. It arrived days after developers fingerprinted an anonymous free model on OpenRouter, "Space Bunny Alpha", and concluded it was an early build of the same system. MiniMax also reset Token Plan quotas and is running a double-credit sign-in promotion until 7 October.

Why it matters

This is MiniMax's third distinct coding model in under a year, each released while the last was still being benchmarked by others. Ship fast, skip the report, let developers work it out themselves. That is a pattern, not a slip, and most coverage simply repeated the one-line post.

More detail has emerged on Julia 1, the 144.3-million-parameter open decision model released on Hugging Face under Apache 2.0 by Supersonic Labs, a small Brazilian lab.

It never generates free text. It selects an answer from 2 to 20 options the caller supplies and returns probability scores. It runs on a plain CPU (benchmarked on an Apple M4, an Intel Core i5 laptop chip and a Samsung tablet), with an optional GPU path and a browser build. It is built on Johns Hopkins' mmBERT-small encoder with a new decision head; the lab says it is not a fine-tuned Qwen.

The numbers

• disclosed cloud GPU spend for training and experiments: about R$540, roughly US$104
• beat the published reference values of the hosted-only Jev model on several pilot tests
• did notably worse on the 72-label Banking77 test: 64% against 87%
• a planned hosted API was not yet open at launch

Only one detailed English write-up covers it so far.

Why it matters

A lot of real work is picking, not writing. A model this small and this cheap is a direct counter to the idea that everything needs a giant generative model behind it.

OpenAI has published a preview of GPT-5.6 Sol, which it describes as its strongest model yet, with improved agentic abilities in coding, biology and cybersecurity.

What OpenAI says

• the release will be phased, with heavier safeguards
• it spent over 700,000 A100-equivalent GPU hours on automated red-teaming for universal jailbreaks
• there is a new "max" reasoning-effort tier
• it is coordinating with the US Administration on a cyber Executive Order framework before wider availability

This rests on OpenAI's own preview page; we haven't seen independent reporting on it.

Why it matters

The frontier is still getting bigger, and here the step up in capability comes with a release that is restricted on purpose and coordinated with government. That is a different kind of launch from a normal model drop, though the changes most people will feel this week came from much smaller models.

A research paper treats object permanence, knowing that something still exists when it's hidden, as a training problem for video world models. The authors built 150 cognitive tasks and about 1.5 million generated clips of objects being hidden and interacting physically.

Fine-tuning a 16-billion-parameter model, PWM-WROP, on this synthetic data produced the top-ranked model for continuing a video in a human evaluation.

The finding

• object permanence did not emerge from scale; it had to be trained in explicitly
• understanding of collisions and physics still lags

The paper was listed on Hugging Face Daily Papers, with affiliations including CMU.

Why it matters

There is a running argument over whether world models actually understand physics. This is a concrete negative result for "just make it bigger", rather than a marketing claim.

Researchers at Stanford and Caltech have demonstrated HomeBody, which gives a frontier vision-language model, GPT Astra, direct control of a Unitree G1 humanoid's five basic skills: navigate, pick, place, open a drawer and pick from a drawer.

They deliberately removed the usual learned control layer that sits between the model's reasoning and the robot's movements. The robot first explores an unfamiliar kitchen and builds a digital twin of it in Nvidia Isaac Sim from its own camera, mapping and joint data. It then uses that as memory to tidy up and fetch items from plain-language instructions.

One limit is refreshingly concrete: task length is capped partly by "finger-servo overheating during extended operation".

The exact publication date is approximate; the project page appeared around the time of the first coverage on 27 September.

Why it matters

If a frontier model can orchestrate simple skills without a policy trained for each environment, home-robot teams might skip a lot of data collection. That is a big if: so far it is one lab kitchen.

Anthropic says an unreleased research version of Claude has improved a long-standing result in analytic number theory: the proven share of the Riemann zeta function's non-trivial zeros that lie on the critical line. The bound moved from 41.6% to 67.2%.

This is not a proof of the Riemann hypothesis. Nobody has one.

Anthropic says the model drew on decades of earlier work and, alongside a paper and an informal note for experts, produced a formally verifiable proof of the result. Two Anthropic mathematicians checked it internally. Externally, on short notice, it was checked by Brian Conrey and Dan Goldston, both well-known names in zeta-function theory.

Mainstream tech press had barely picked it up at the time of writing.

Why it matters

Machines have done clever maths for a while. The point here is that this is maths you don't have to take on faith: a proof a computer can check, not a plausible-looking write-up. Most claims about ai arrive from the people selling it. A result you can verify for yourself is a different kind of news.

Supersonic Labs has released Julia-1, a 144.3-million-parameter open decision model. Instead of generating text, it returns typed answers and calibrated probabilities, built to make one gated judgement inside software at very low cost and latency.

• weights on Hugging Face under the Apache 2.0 licence
• runs locally on a CPU or a BF16 GPU
• an ONNX build runs in the browser via WebGPU

The stated limits are candid. It compares answers you supply and can't be relied on for missing facts, arithmetic or multi-step calculation, and its router can drop the correct label while narrowing down options.

It joins a small but growing category of decision models, alongside TypeSafe's Jev and Fastino's GLiNER2.5-Decide.

Why it matters

In a field that measures models in hundreds of billions of parameters, this one fits on hardware you already own. Open weights mean you host it, inspect it and own it, rather than renting something by the token and trusting what you're told about it.

OpenAI's Codex release notes, last updated on 26 September, add support for GPT-6 Sol and GPT-6 Luna. Coverage also reports a 50% API price cut compared with a GPT-5.6 promotion.

For anyone building on Codex, the newest models are now the default choice, and the cost of running them is moving down.

Reported by Releasebot.

Anthropic says Claude computed the six-particle scattering amplitude in planar N=4 super-Yang-Mills at nine loops, one order beyond the previous eight-loop record held by SLAC's Lance Dixon and collaborators.

Anthropic physicists Liam Fitzpatrick and Siddharth Mishra-Sharma gave Claude a one-line prompt. It ran largely unsupervised for days using the established bootstrap method, at a total cost of a few thousand dollars. Dixon validated the result independently. The physicist Matt von Hippel, who issued the original public challenge in August, wrote the guest post.

The footnote

• Claude used known methods, the ones Dixon and his colleagues built, and pushed them one loop further. It did not find new physics.
• A group led by Song He at the Chinese Academy of Sciences, working with GPT-6, had already obtained most of the same result. Their dataset is dated 17 September.

Why it matters

This is one of the most concrete cases yet of a language model completing a frontier research calculation end to end on an academic budget. But the accurate headline is "ai grinds", not "ai discovers": known methods, run faster and cheaper than a person could. That may be just as impressive. It is a different claim, and most coverage has flattened it into "ai breaks physics record".

On the same day as the nine-loop post, Anthropic's institute published a proposal for three measurements it says the public should be able to track inside frontier labs.

The three metrics

• the share of ai research and development done by ai itself
• how well agent actions are overseen: coverage, review latency and escalation rate
• how compute is allocated, including how much goes to safety work

As a snapshot, Anthropic said that in August 2026 roughly 30,000 agents were doing research and engineering work at any one time on its most-used internal platform. It also examined its own compute mix for a week in July, and says it plans to bring in independent third-party evaluators to check the numbers. It has encouraged other labs to publish similar figures.

Why it matters

Thirty thousand agents is not a lab tinkering; it is a workforce. Auditable numbers are rare in a debate about ai improving itself that is mostly rhetoric. This went out on the same day as a flashier announcement and got far less attention.

A prompt and skill optimisation method called CASD has been doing the rounds this week. It has an off-the-shelf coding agent read a full set of agent logs and write its findings as rules, in a single pass, with no environment access or validation data.

Reported results

• across ALFWorld, two τ²-bench environments and SpreadsheetBench-Verified, one pass averaged +16.6 percentage points, against +10.9 for GEPA and +5.3 for SkillOpt
• about $1.60 per optimised prompt, claimed to be more than 22 times cheaper than validation-gated search
• it beat GEPA on three of four tests and SkillOpt on all four

These figures come from secondary write-ups. We couldn't reach the paper itself, so we can't say who wrote it or when it first appeared, and we aren't presenting it as new.

Why it matters

It fits the pattern of the week: the expensive thing quietly becomes the cheap thing. For anyone running agents, cheap prompt optimisation that needs no validation data is a practical lever, if the results hold up.