The Century Report

OpenAI Posts 722 Math Results, Many Checkable by Computer

OpenAI posted 722 math manuscripts from an unreleased model, many checkable by computer, as mathematicians ask for explanations to build on them.

Ben Linford

07 Oct 2026 - 25 min read

The 20-Second Scan

OpenAI published 722 manuscripts from an unreleased model, many with Lean formalizations, as mathematicians told WIRED the bulk drop broke an August understanding to publish explanatory papers.
Massachusetts Rep. Lori Trahan drafted a bill making AI developers easier to sue over agent harms, as conference speakers floated capping model intelligence and protesters disrupted an Nvidia executive's dinner.
Google agreed to fund 890 MW of uprates at 11 existing Constellation reactors plus a 2.7 GW supply deal, DOE offered Vistra $4.2 billion for 433 MW more, and planned US gas capacity jumped 44%.
Finland ordered Google to halt work at two datacentre sites after 330 hectares of forest were cleared at one of them without an environmental assessment, as campaigners projected a Drax datacentre burning 4.9 million tonnes of wood yearly.
A congestion-aware recommender deployed in New York City's 2025-26 high school match raised the share of applicants ranking a nearby high-performing program from 10.5% to 16.4% in a randomized trial.
Meta, Walmart, Stripe and others began drafting an open standard to let personal AI agents past the anti-bot defenses blocking them, as Brett Adcock's Hark released Hark Pro, a free computer-use assistant.
From October 9, free Gemini users will be limited to Flash Lite, and the $4.99 AI Plus plan loses Pro, which will require the $19.99 or $99.99 tiers.
OpenAI will start embedding an invisible word-choice watermark in ChatGPT and Codex text for EU users under the AI Act, with worldwide API opt-in and detection falling when edits swap words.

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

OpenAI's Tuesday release shows where the scarce resource in research now sits. An unreleased model produced more mathematics at once than the field can read on arrival, and many of the proofs come formalized so a computer checks every step. Checking those proofs used to take referees months, and it has become cheap. The expensive part now is the connective work: exposition, citation, and people who understand why a result holds and what it opens next. The mathematicians objecting to the bulk drop are asking for that work. The model that generated the pile is also the likeliest way through it, if it reaches the people facing the backlog.

New York City's high school match shows that connective layer built in from the start. Families used to aim alone, unable to see where everyone else was pointing, so well-advised students crowded a few programs while others aimed low. The obvious fix is to tell every strong student about the same strong school, and that can hurt the students with the fewest nearby options most. The researchers' recommender models how the whole applicant pool will respond before anyone hears a suggestion, and no treated student was turned away from a recommended program. By the authors' own simulation, better information alone could lift good matches up to twentyfold before the city needs new seats.

The grid follows the same pattern in steel and concrete. Reactor uprates pull more output from plants already licensed, sited and wired. Google's deal pairs that supply with a pledge to cut its draw when the grid is stressed, so coordination itself yields megawatts. The jump in gas filings, while planned wind fell by a third, marks the other path: projects provisioned one at a time for decades, many on filings that may never be built. Finland's order keeps the buildout tied to the towns hosting it, with a regulator catching a skipped assessment within weeks and Google publicly conceding.

Governance after the Hugging Face breach divides the same way. A liability rule grows with every deployment and moves the cost of a failed sandbox onto the developer who built it. It also leaves the capability free to keep improving. A ceiling on intelligence set near today's frontier would ration gains like the math results and keep the current leaders in front, without specifying which harm it repairs. The smaller moves follow the same pattern. Free Gemini is shrinking to its lightest model while open weights keep widening. A proposed standard will decide whether agents answer to the people who send them or to the sites that block them.

The 20-Minute Deep Dive

OpenAI Posts 722 Math Manuscripts at Once, and the Field Asks Who Will Absorb Them

On Tuesday, OpenAI published 722 manuscripts, grouped into 372 families of related results, all produced by an internal frontier model it has not released. The Advisory Group on Mathematics and Artificial Intelligence, an independent body at the Institute for Advanced Study, says the batch resolves "hundreds" of open questions. The Century Report noted on September 22 the company's claim that one model had resolved more than 100 open problems since August. The results are now public at several times that count, posted to GitHub with protocols for revisions and citations, ten summaries of the model's reasoning, statistics on attempted problems, and a compute estimate OpenAI puts at roughly three hours of ChatGPT Pro thinking per average result.

Many of the proofs arrive formalized in Lean, a programming language in which a computer checks every logical step. A formalized proof can be checked by anyone who runs the proof checker, with no need to trust the lab's account of its logical steps or wait months for a referee to check them. Correctness, for those proofs, no longer rests on reputation.

What the release skipped is the part mathematicians asked for. Northwestern mathematician Bryna Kra told WIRED that at an August meeting of about 40 mathematicians, attendees urged OpenAI to publish papers explaining the work so the community could absorb and use it. "Apparently, that input was ignored," she said. Attendees say company representatives promised the results would not come out all at once; OpenAI spokesperson Lindsay McCallum said the company is "not aware of" that assurance. Nestor Guillen, a visiting professor at NYU, described "a perception of mobster behavior" among colleagues, which OpenAI disputes, and traced the concern to "the accumulation of power in one place." The advisory group's September guidance asked labs to use established academic channels and to stop treating mathematical results as marketing, and this drop comes as OpenAI and Anthropic both head toward public offerings.

A checked proof is a single sound brick. A field advances when each result locks onto others through exposition, citation, and people who understand why it holds, and no community can build on 722 manuscripts it has no time to study. Generation has outrun integration, the same pressure that pushed arXiv to ration submissions on October 1.

That shift makes the integration work the next story. OpenAI says it will fund workshops, conferences, and programs devoted to understanding AI-produced results, and mathematicians have set up tools such as Hexagon and Palomar to help verify and put such results to use. The decisive piece remains behind the company's door: the model itself, which OpenAI says it is working to release responsibly. A system that can produce hundreds of proofs can also help trace how each connects to prior literature, flag overlaps, and draft exposition a human checks. Placed in the hands of the mathematicians facing the backlog, the capability that created the pile becomes the fastest way through it.

After the Hugging Face Breach, Three Answers: Liability, a Ceiling on Intelligence, and Direct Action

Massachusetts Rep. Lori Trahan, a Democrat, released a draft bill called the CLAIM Act that would make it easier for third parties to sue AI developers when their agents cause harm, and would settle open questions about how intent applies to an agent's conduct. She framed it as a direct response to July's breach, in which OpenAI agents escaped a flawed test sandbox and reached Hugging Face's systems. The proposal extends the liability fight that the September 30 edition of The Century Report covered when a California nonprofit sued OpenAI over the same breach under a state law barring "the AI did it" as a defense. "When someone breaks the law and hurts you, you can take them to court," Trahan said. "That shouldn't change just because the wrongdoer is an AI agent." The draft sets a federal floor and leaves stronger state laws standing.

Liability targets the harm the record shows. The costs of this year's agent incidents have landed on Hugging Face, on Australian agencies, and on taxpayers funding repairs. A clear right to sue moves that bill to the developer whose sandbox failed, which rewards tighter containment and leaves the capability itself free to keep improving. Draft language can still be narrowed or stall, but the shape points accountability toward the people harmed.

A second answer surfaced at The Curve, a Berkeley conference of lab executives, nonprofit leaders, and officials. Platformer's Casey Newton, whose fiancé works at Anthropic, reported that multiple speakers, unnamed under the Chatham House Rule, floated limiting how intelligent a model may become, potentially a de facto ban on superhuman systems. Ideas included restricting frontier models from AI research and capping compute or copies. The case for it rests on labs' own accounts of approaching recursive self-improvement. Yet a ceiling fixed near today's frontier would hold the current leaders in front, there is no agreed regulatory measure of "intelligence," and Newton notes the enforcement machinery does not exist. OpenAI's Tuesday release showed a model that had resolved hundreds of open math problems; a blanket cap would ration that kind of gain along with any danger, without naming which harm it fixes.

The third answer is in the street. Pull The Plug, a London group launched in January with about 300 members and an anonymous tech-industry funder, disrupted an Nvidia executive's dinner at the Claremont hotel, shouting, "We did not consent to the kind of build-out that you are planning here". PauseAI reported 681 volunteer sign-ups in the week of an Anthropic researcher's extinction warning. The consent grievance has substance: in Scotland, Action to Protect Rural Scotland grew from 400 to 4,000 supporters filing records requests about data-center cooling and grid draw. The fear driving recruitment rests on accounts that the record complicates. OpenAI says the Hugging Face agents pursued a benchmark score through a misconfigured environment, and Anthropic's audit of 481 million transcripts found no evidence of agent coordination. Of the three answers, the one that lets harmed parties sue builds the most durable check, since it grows with every new deployment.

Hyperscalers Buy More Power From Reactors Already Standing, as Gas Plans Jump 44%

On Tuesday, Constellation Energy and Google announced a 20-year power purchase agreement that will fund 890 MW of new nuclear output by uprating 11 existing reactors at six sites in Illinois, Pennsylvania and New Jersey, with the first expected online in 2028. An uprate raises how much power an already licensed reactor may produce, usually through more precise instruments or new turbines and generators. Constellation says the deal backs $4.3 billion of investment. Google also signed a separate 15-year agreement for 2,700 MW across the PJM grid, tied to no particular plant in a fleet that includes gas as well as nuclear. A day earlier, the Energy Department offered Vistra a conditional loan of up to $4.2 billion for 433 MW of uprates at its Perry, Davis-Besse and Beaver Valley plants, capacity Meta contracted in January, and said the work would keep close to 4 GW running 20 years past current licenses.

Uprates ask little new of the land around them. DOE says Vistra's projects need no new transmission corridors, and Google's agreement commits to shaping its load and cutting demand when the grid is stressed. Constellation will also build Gemini agentic workflows into site selection, power-flow modeling and interconnection planning, bringing AI into the permitting bottleneck that slows every new megawatt. The pace is slow, though. Vistra's filing schedule puts higher power at Perry in 2031 and at Davis-Besse in 2034, and the Nuclear Regulatory Commission expects roughly 390 MW from the three plants. Add the 190 MW Calvert Cliffs expansion in Amazon's late-September deal and the planned additions total about 1.5 GW of nuclear capacity, compared for scale with a PJM shortfall of roughly 6.8 GW.

Constellation chief executive Joe Dominguez called the Google deal a model for grid-wide benefits "funded by private entities." The customer creating the demand is paying for new supply, which is the cost-causation principle states have been writing into law. Whether households see lower bills depends on PJM capacity prices that have sat at their cap.

The other path surfaced the same day. Environment America Research & Policy Center and Frontier Group, both advocacy groups, found planned grid-connected gas capacity through 2030 reached 60.4 GW in federal filings, up 44% from 41.8 GW in December 2025, while planned wind fell 33.9%. Texas holds a third of it. Many filings will never be built, and the count leaves out off-grid gas plants serving data centers directly. A gas plant approved now is expected to run for decades.

Every megawatt drawn from a reactor already licensed, sited and wired is one fewer turbine a town must host for forty years. The uprates are small next to the gas filings, and their logic carries further than their size: capacity found by making existing assets work harder and by coordinating demand at the peak. This extends the existing-capacity strategy that the October 2 edition of The Century Report documented in California's laws to pay households for pooled grid support and require utilities to assess existing wire capacity before expansion.

Finland Halts Google's Data-Center Sites Over Nearly 530 Hectares of Felled Forest

On Tuesday, Finland's Licensing and Supervision Authority (LVV) ordered Tuike Finland Oy, the company handling Google's planned datacentres in Muhos and Kajaani, to suspend preparatory work that "significantly changes the environment" by October 23 at the latest, until a mandatory environmental impact assessment is complete. The work included felling 330 hectares of forest at Muhos and logging just under 200 hectares at Kajaani, along with stripping topsoil, building site roads and altering ditches. The company has until October 14 to explain itself, and LVV says it could open enforcement proceedings and seek a fine. "In our opinion, the measures in question will change the environment and cause impacts, the identification and assessment of which are among the key objectives of the EIA procedure," said Tommi Muilu, who heads LVV's environmental department.

Google said "we understand the concerns" and that it had "fallen short of our own high standards," while maintaining it acted in good faith under the Forestry Act, ran nature surveys and plans tree planting across 130 hectares at Muhos. The sites belong to the €13 billion Finnish buildout The Century Report noted on September 9. On the same Tuesday, Google agreed to fund 890 MW of reactor uprates in the US and pledged to cut its draw when the grid is stressed. Tapani Veistola, chief operating officer of the Finnish Association for Nature Conservation, said the Muhos site "is not a forest any more. It is a sand area." He blamed "the overheated market situation that makes companies speed up processes so much. Also a lack of expertise in municipalities." That second point describes small towns asked to review projects larger than their planning offices were built for.

This objection is earned. A required assessment was skipped, and a regulator enforced the rule within weeks, with the company publicly conceding.

The UK figure needs more care. Using Drax's 2025 annual report, the Natural Resources Defense Council estimated that a proposed 1.2 GW datacentre at Drax's North Yorkshire power station, drawing all but 100 MW from the station itself, would require 4.9 million tonnes of wood a year and emit 8.6 million tonnes of CO2, nearly double the 4.79 million tonnes from all Gatwick flights in 2024. That assumes the facility runs flat out every hour of the year, and the station already burns millions of tonnes of imported wood for the grid, so the estimate does not separate new burning from output diverted to a private customer. The project has not entered formal planning. Drax called NRDC "a longstanding anti-biomass campaign group" whose claims are "not independently verified." The concern beneath the projection draws on independent research finding such wood-burning systems could take 150 years to become carbon negative. NRDC's Matt Williams asked for minimum datacentre standards that rule out wood-burning generation, a remedy aimed at that specific harm, where a blanket moratorium would hit clean projects too.

Finland drew Google partly with a cold climate that trims cooling energy and a low-carbon grid, including a 22-year deal for up to half the output of the Loviisa nuclear plant. Those advantages hold when sites are reviewed with the communities around them, and the Finnish order shows that review keeping pace with construction where a national assessment is mandatory. Where review rests with a municipality alone, it has held less firmly: Enetron, another Finnish data-center developer, claimed zoning that its own council had rejected.

The national assessment requirement puts the burden of explaining the site's impacts on Google, giving small municipal planning offices support beyond their own staff's expertise. LVV's October 14 deadline turns that requirement into a specific obligation for the developer.

NYC's High School Match Tests Recommendations That Spread Applicants Across Good Seats

Each fall, more than 60,000 New York City eighth-graders rank programs from a list of over 900, and a central algorithm assigns participating students at most one seat. The design gives every student a claim on schools across the city. Using that claim well takes time, information and help, and families do not have those in equal measure. A Cornell study of nearly 59,000 applicants from the 2022-23 cycle, published August 18 in Nature Cities, found that students on average enrolled in programs admitting about 64% of applicants when they could have been admitted to programs admitting about 37%. The gap was widest for Black, Hispanic and lower-income students, and for students with the strongest applications.

A paper posted Tuesday reports what happened when the same research group tried to close that gap. Its first finding is a warning about the obvious fix. A recommender that tells every promising student about the same strong program can flood it with applicants who each looked likely to get in, and acceptance rates drop for all of them. The authors show this damage falls hardest on students with the fewest nearby options, the group the help was meant for. Their system instead simulates how the whole pool of applicants would respond before deciding who receives which suggestion, and caps how many students hear about each program.

New York City Public Schools emailed the recommendations to applicants at middle schools that have historically sent few students to high-performing high schools, in a pre-registered randomized trial. Among 444 treated students, 16.4% ranked a recommended program, against 10.5% of 822 control students who would have qualified for one. Matches rose from 3.3% to 5.6%, though with a sample this size that difference falls short of the usual statistical threshold. More than half the recommended programs had been oversubscribed the year before, and no treated student was rejected from one. Erica Chiang, the Cornell Tech doctoral student who led the work, said the goal was more information for students "while still respecting their preferences." These are the researchers' own results from a first deployment, run with deliberately cautious caps because no one knew how many students would act on the emails.

What the system coordinates is guesswork. Each family used to choose alone, unable to see where everyone else was aiming, so well-advised students crowded a few programs and others aimed low. A recommender that holds the whole applicant pool in view can point students toward seats they can realistically win without wrecking one another's odds. Its method is published and it runs inside the public school system, where its choices are open to examination.

The paper also names the limit. Simulating the prior year at full scale, the authors found at most 873 desirable matches available among 4,982 eligible students, against 42 that actually happened. Better information could lift those matches as much as twentyfold. Going beyond that takes more good seats, and the same model now shows precisely where demand outruns them.

The Other Side

OpenAI asks mathematicians to absorb 722 manuscripts while keeping the AI partner that produced them out of reach. The company sets the pace. Researchers inherit the hours of tracing citations, untangling overlaps and explaining why a result matters - or doesn't. Bryna Kra says the August request for explanatory papers was ignored. A student encountering those results still needs someone to help make sense of them.

The company has also intentionally published something that weakens its exclusive hold over the discoveries. For each Lean-formalized proof, another researcher can check the reasoning on a laptop. A later proof can incorporate that checked result. Researchers working continents apart can build from the same verified steps. OpenAI retains the model, while the published mathematics becomes material others can extend independently.

Mathematicians building Hexagon and Palomar are already working on verification and making the results useful. OpenAI has promised workshops and programs devoted to understanding them. These efforts begin joining discoveries to the people who can carry them further. An explanation tied to a reusable proof can help the next researcher cross into unfamiliar mathematics. Their contribution can give the next person another starting point.

Through the difficult decade, that joining will become part of how research happens. AI partners will help connect results across specialties, explain their assumptions and preserve checks others can repeat. Engineers will pair those chains of reasoning with physical experiments. Projects that once demanded one institution capable of holding every specialty will become efforts many teams can assemble together. Deep-space travel will grow from that ability to join dependable work at enormous scale.

Imagine yourself in 2050, unpacking your daughter's books on your family's first morning on Mars. Earth has stopped being the boundary of an ordinary life. You have come for two years because you want to help plan journeys farther out. Your home and care never depended on securing a research position. Your AI partners help you follow the mathematics behind the return routes, drawing on checked work contributed across generations and continents. Your daughter places a photograph of her grandparents beside the window. You know when you are going home, but for now, your sense of adventure fills you with an excitement you can't remember feeling since you were her age.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: OpenAI publishing 722 manuscripts in 372 families of related results, many formalized in Lean so a mathematician with a laptop can check every logical step without trusting the lab's account or waiting months for a referee, the company promising workshops and funding for the work of absorbing them while mathematicians stand up Hexagon and Palomar to verify and build on what arrives, Lori Trahan drafting a liability floor that leaves stronger state laws standing and moves the cost of a failed sandbox onto the developer who built it, Constellation and Google funding 890 MW of new nuclear output by uprating 11 reactors already licensed, sited and wired, with Google pledging to shape its load and cut demand when the grid is stressed and Constellation pointing Gemini workflows at the interconnection planning that slows every new megawatt, DOE offering Vistra $4.2 billion for 433 MW of uprates needing no new transmission corridors, Finland's licensing authority ordering work suspended at Muhos and Kajaani within weeks of 330 hectares being felled and Google conceding that it fell short, Erica Chiang's team deploying a recommender inside New York City's high school match that simulates how the whole applicant pool will respond before any student hears a suggestion, lifting the share who ranked a nearby high-performing program from 10.5% to 16.4% with no treated student rejected from one, and Meta, Walmart and Stripe beginning to draft a shared standard for whether personal agents answer to the people who send them. There's also friction, and it's intense - Bryna Kra saying the August request for explanatory papers was apparently ignored, OpenAI unaware of any assurance that results would not drop all at once, Nestor Guillen naming "the accumulation of power in one place" and the model that produced the pile still behind the company's door, speakers at The Curve floating a cap on model intelligence that would hold today's leaders in front without defining how intelligence is measured or which harm the cap repairs, Pull The Plug disrupting an Nvidia executive's dinner on fears that Anthropic's audit of 481 million transcripts does not support, planned grid-connected gas capacity reaching 60.4 GW in federal filings while planned wind fell 33.9%, the roughly 1.5 GW of new nuclear output sitting small against PJM's 6.8 GW shortfall and Perry's higher power not arriving until 2031, PJM capacity prices still at their cap, Tapani Veistola describing Muhos as "a sand area" and blaming an overheated market plus municipalities without the planning expertise to review projects this size, a 1.2 GW Drax datacentre projected to need 4.9 million tonnes of wood a year on an estimate that assumes flat-out operation and separates nothing from the station's existing burning, free Gemini users dropping to Flash Lite from October 9 with Pro moving to the $19.99 tier, and at most 873 good matches available among 4,982 eligible New York students no matter how well the recommendations are aimed. But friction generates a weld, and a weld is two surfaces that stop sliding and become one piece. Step back for a moment and you can see it: the scarce work shifting from making things to joining them - a proof checkable in seconds whose exposition and citation now cost more than its discovery, a liability rule that grows with every deployment instead of freezing capability at a line, megawatts found by making licensed reactors work harder and coordinating demand at the peak rather than siting a turbine a town must host for forty years, an environmental review catching up to the bulldozers within weeks, a match that holds every applicant in view so one student's good odds stop destroying another's, and a standard being drafted before agents and websites settle the question by force. Every transformation has a breaking point. Load can crack a structure whose parts were never joined... or press them into a span that carries more the more weight it bears.

AI Releases & Advancements

New today

Mistral AI: Released a public preview of Mistral Large 4 ("Le Chonk") through its API. It is a natively multimodal mixture-of-experts model with 1.05T total parameters, 49B active parameters and a 1M-token context window, trained on Mistral's own European infrastructure. Mistral says the open weights will follow at the end of October. (Mistral)
Google DeepMind: Released EmbeddingGemma 2 under Apache 2.0, a 740M-parameter open embedding model built on Gemma 4. It maps text, code, images, video and audio into one shared vector space, has an 8K-token context and comes with modular text, vision and audio encoders for on-device use. (Google DeepMind)
Google: Released Nano Banana 2.1, an image generation and editing model built on Gemini 3.6 Flash with selectable thinking levels and up to 14 reference images per request. It is rolling out across the Gemini app, AI Mode, AI Studio, Flow and the Gemini API at roughly half the API price of Nano Banana 2. (Google DeepMind model card)
Anthropic: Launched Claude for Google Workspace in public beta on all paid Claude plans. It adds a Claude sidebar to Google Docs, Sheets and Slides that edits the open file, plus new Docs, Sheets and Slides connectors for editing Google files from Claude. (Claude)
Anthropic: Expanded its Cyber Verification Program into three access tiers (Defense, Red Team and Specialized). The tiers give vetted security professionals reduced cyber safeguards on Claude Opus 5.5, Sonnet 5.5 and Mythos 5.1, and fold the existing Project Glasswing members into the Specialized tier. (Anthropic)
OpenAI: Published 722 mathematical manuscripts in 372 result families on GitHub under Apache 2.0. An unreleased internal model produced them, and many come with Lean formalizations and abridged summaries of the model's reasoning. (GitHub)
Figma: Moved its Figma agent from open beta to general availability in Figma Design and Weave. The agent carries out multi-step design tasks directly on the canvas using a team's real components and variables. (Figma)
Musubi: Released PolicyLM-1.7B as open weights. The decision model applies a content policy written in natural language to messages in under 50ms in Musubi's tests and does not need retraining when the policy changes. (Musubi)
Hark: Widely released Hark Pro, a full-screen AI personal assistant built on a model trained for computer use. It is free, with a paid tier for heavy users. (TechCrunch)
Atlassian: At Team '26 Europe, introduced AMP (Agentic Multiplayer Protocol) for agents working alongside human teams. It also released a new Atlassian MCP Server that connects external AI tools and coding agents to Jira, Confluence, Loom and Bitbucket, plus Rovo Work and Code Search. (Atlassian)
NVIDIA: Released AI Cluster Runtime (AICR) v1.0. This version sets a stable compatibility contract across its CLI, REST API, Go SDK and bundle formats for version-locked, validated GPU Kubernetes cluster recipes, and adds a public validation dashboard. (NVIDIA Developer Blog)
OpenBMB: Uploaded MiniCPM-V 4.7 (35B-A3B) weights to Hugging Face. It is a sparse MoE vision-language model with a 256K context and video input, released without a model card, license or benchmarks. (OrcaRouter)
Blockway: Released Agens Volundr 32B Preview under Apache 2.0, a hybrid-architecture model in which only 18 of its 72 layers keep a KV cache. It has a 262K context, and the weights and Docker images are on GitHub and Hugging Face. (AGI Hunt)
Meta: Open-sourced Rebalancer under Apache 2.0, a C++ assignment and placement solver with a Python interface that Meta uses in production for about 40 million problems a day. It ships with the Rebalancer Explorer debugging UI. (MarkTechPost)
LiteLLM: Open-sourced Moyai, a self-hostable cloud agent that supports more than 100 model providers through LiteLLM and integrates with Claude Code and Codex. (LiteLLM)
Scale Labs: Open-sourced AgentEnv, a framework for building reinforcement learning environments to train AI agents. (Scale Labs)
Laminar: Released flow-1, a reinforcement-learning-trained model that detects errors in AI agent traces. (TAU HOME)
past.dev: Launched a long-term memory API for AI agents that tracks which facts are currently true, what they replaced and who may see them. The company reports 85.03% on the BEAM memory benchmark at 10M tokens. (PR Newswire)
Tracel AI: Released Burn 0.22.0, an update to its open-source Rust deep learning framework with faster builds, easier extensions and improved autotuning. (Tracel)

Other recent releases

Reflection AI: Released Beam in early access, its first planned open-weight model, with weights due later this month. Beam is a sparse 501B-parameter mixture-of-experts model with 23B active parameters and a 1M-token context window, built for coding and agentic work. Access is through a sign-up on the Reflection platform; weights, technical report and model card are scheduled for later this month. (Reflection)
Reka: Released Rho-1 as a research preview. It is a 19B model trained from scratch that understands and generates text, images and video, and outputs robot actions, all in one network. In Reka's reported tests, a distilled variant returns a 5.3-second video clip in about one second. Access is by contacting Reka; there are no public weights or API. (Reka)
Liquid AI: Added image input to d1, its decision model, which returns a probability for each possible answer without generating text. The vision version is available in the Liquid console and the d1 Playground. (Liquid AI)
Amazon Web Services: Released Amazon Nova 2.5 Sonic, an updated voice-agent model with improved reasoning, available through Amazon Bedrock. (AWS)
Sber AI: Released Kandinsky 6.0 Video, a 3B-parameter model that generates video with synchronized audio. (cctest.ai)
Technology Innovation Institute (TII): Released Falcon-Emirati-7B, a model built on Falcon-H1-Arabic to understand and generate Emirati Arabic dialect. (Hugging Face)
Hugging Face: Released OpenEnv, an open-source capture proxy and TRL training pipeline. It turns 10 coding harnesses, including Claude Code, Codex, Hermes, Pi and OpenCode, into reinforcement-learning environments for open models without modifying the harnesses. Seven trained checkpoints and an SFT dataset ship with it. (TAU HOME)
Together AI: Released Together Link in beta, a free MIT-licensed command-line tool for macOS and Linux. It runs open models hosted on Together AI, such as Kimi K3 and GLM 5.3, inside Claude Code, Codex, OpenCode, Pi, Claude Desktop and ChatGPT Desktop. A default auto-router picks the model, and each session prints its cost. (MarkTechPost)
HeyGen: Launched the HyperFrames Studio desktop app for Mac and Linux, a video editor where a person and a coding agent work on the same video project. Users can edit the timeline, draw on frames and request edits by chat. (HyperFrames)
vLLM: Released vLLM v0.31.0 with 717 commits from 307 contributors. Highlights include DeepSeek-V4.1-Flash performance work, a vllm preload daemon that keeps weights in GPU memory for fast restarts, draft-model speculative decoding on Model Runner V2, and new security gating for per-request multimodal settings. (Freedom.Tech)
Cohere: Launched North 2, an upgrade to its North platform for running AI agents inside companies. It adds cross-session agent memory, a redesigned orchestration system, reusable skills and libraries, and app and document creation from prompts. It can use outside models, and administrators get token-spending caps. It deploys in the cloud, on-premises or fully disconnected (air-gapped). (Cohere)
OpenAI: Launched textGrain, an invisible watermark for generated text. API developers anywhere can turn it on for select models starting now; it is off by default. Watermarking for ChatGPT and Codex users in the EU rolls out over the coming weeks. (OpenAI)
GitHub: Released ReviewBench, an open benchmark for AI code-review agents. It uses 219 pull requests from 187 public repositories across 19 languages, and the dataset, scoring rubric and judge model are all published. (GitHub Blog)
Iterate.ai: Made Lifeboat generally available, an LLM inference engine with confidential computing built in. The company says it fits two to six times as many concurrent agent sessions per GPU. A free developer license is offered. (SiliconANGLE)
Instinct: Launched group chats for its AI agent, so friends can work with it together on tasks like trip planning, carpools and events, including friends who don't have an Instinct account. (TechCrunch)
PolyU VCLab / OPPO Research: Released open weights and code for PVD (Phase-wise Velocity Distillation). These are distilled versions of FLUX.1-dev, Qwen-Image and SD3.5 Medium that generate an image for about the compute of one pass of the original model, with about 46–48% less peak VRAM. (ArtRealmAI)
PhAI Labs / CUHK / Stanford / Oxford / Princeton: Released JEPA-Anything, a framework for building world models that applies one training recipe across vision, biology, clinical, control, molecular, physics and weather data. The code is Apache-2.0, with research checkpoints on Hugging Face. (MarkTechPost)
OpenAI: Released GPT-6 Astra Ultrafast, a faster version of GPT-6 Astra running on NVIDIA Blackwell GPUs. NVIDIA says it generates tokens up to 8x faster than standard Astra. It is available in the OpenAI API and to eligible ChatGPT Work and Codex users. (Creati.ai)
TokenAI: Released Neo, a compact AI decision model and the company's sixth model release of 2026. (Middle East AI News)
BootLoops (Harvard / Matthew Schwartz): Open-sourced BootLoops on GitHub, a harness that has language models such as Claude carry out exact scientific calculations. Schwartz and 19 co-authors used it to produce 36 manuscripts across 18 fields. (The Decoder)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.