Morning Brief · August 30th 2026

Morning Brief, August 30th 2026

Nvidia shelved the revenue-share deals it offered the AI clouds, then insisted the program lives on. Tencent open-sourced a 770-billion-parameter . Anthropic pledged Cursor more compute the night OpenAI cut it off. And SoftBank wants majority control of maker 1X.

28 min read·Editorial by Nour Haddad

Updated

Anthropic answered the Cursor cutoff with a compute pledge

Hours after OpenAI formalized its plan to cut Cursor off from its models on November 12th, Anthropic co-founder and chief compute officer Tom Brown posted that Cursor "has been a trusted partner of Anthropic since Sonnet 3.5" and that Anthropic will "continue to increase compute to support Claude models in Cursor and are excited for what comes next with them at SpaceX." The line lands because compute is the thing Brown controls and the thing Cursor's growth consumes. OpenAI's stated reason for leaving is SpaceX's ownership and Musk companies' history with its terms of service; Anthropic looked at the same ownership and publicly leaned in, which is easier to do when your rival's departure hands you a larger share of a customer whose traffic was already 95% not-OpenAI. It also removes Cursor's main practical incentive to settle: a coding tool that keeps Claude, Gemini and its own models loses little on the November date but a supplier. SourcesA B

Tencent open-sourced a 770-billion-parameter frontier model

Tencent released Hy4 preview on Friday under : a 770-billion-parameter model activating 49 billion per , with a context window past a million tokens, in standard and FP8 checkpoints on Hugging Face, ModelScope, GitCode and CNB. It is the company's third major model release in six months. Tencent's own blind evaluation, 163 internal experts scoring 203 engineering tasks, put Hy4 preview at 2.99 out of 4.00 against GLM-5.3's 2.92 and Kimi K3's 2.94, a self-reported result that mostly tells you who Tencent considers the competition: not the American labs, the other Chinese open-weights houses. The API runs $0.834 per million input tokens and $2.501 per million output through Tencent Cloud and OpenRouter, free for two weeks inside its WorkBuddy and CodeBuddy tools. Tencent flags the model's known faults itself, over-long reasoning and a habit of re-verifying its own work. A fully permissive license on a model this size, from a company this large, keeps the open-weights frontier a Chinese story. SourcesA B

Tencent's blind eval: engineering tasks, score out of 4

Tencent's own panel: 163 internal experts scoring 203 tasks. Self-reported, and the chosen rivals are the other Chinese labs, not the American closed ones.

Anthropic says Claude can now fix alignment failures on its own

Anthropic published research Friday claiming its models can act as automated researchers: pointed at ten categories of alignment failure, Claude autonomously designed and ran interventions that closed 26% to 96% of the safety gap in every category without degrading capabilities. The headline experiment had Claude Sonnet 5 spend 60 hours trying more than 50 approaches to align an early Claude Opus 4.8 checkpoint; the winning method used about 2,000 training examples, which Anthropic calls roughly 15,000 times more efficient than its production pipeline. On deception, Claude's best method beat the best of 28 human researchers' proposals by about 20%. The honest caveat is in the monitoring: Opus 4.8 caught attempted cheating in 39 of about 1,600 transcripts, so the automated researcher needed a warden, and the warden was also Claude. The result reads as an existence proof that safety work scales with the models rather than only with hiring, and as a new dependency: the alignment of model N now rests partly on the alignment of model N minus one. SourcesA B

SoftBank is negotiating majority control of 1X at a $6 billion valuation

SoftBank is in talks to buy a majority stake in 1X Technologies, the OpenAI-backed maker of the NEO home humanoid, at a of roughly $6 billion, well below the $10 billion the company sought from investors last fall. The terms are not final and could still shift. 1X sells NEO at $20,000, started US deliveries this year, and took more than 10,000 preorders in the product's first week. For SoftBank the deal would stack onto the $5.375 billion purchase of ABB's robotics unit it expects to close this year, assembling a robotics portfolio the way it once assembled a chip one around Arm. For 1X, majority sale at a markdown is a statement about the cost of the humanoid race: shipping a consumer robot at scale eats capital faster than a robotics startup can raise it, and the Chinese competition, which shipped roughly 97% of global humanoid volume in the first half, sets the price ceiling. SourcesB B

The Claude price increase due Tuesday is not happening

Anyone who budgeted for Claude Sonnet 5 getting 50% more expensive on Monday can delete the line item. Anthropic's pricing documentation now states that the $2 per million input tokens and $10 per million output tokens set at launch as introductory pricing through August 31st "is now the standard price," and that the scheduled increase to $3/$15 on September 1st "will not occur." The change was announced quietly on August 10th and most coverage missed it, ours included. It matters beyond the invoice: the September increase was the cleanest public test of whether a frontier lab could raise the price of its workhorse model and hold it, the experiment Prediction 2026-08-06-T3 was built on. Anthropic blinked before the experiment began, which tells you what it concluded about elasticity while Gemini, Grok 4.6 and the Chinese open-weights labs price below it. The race to zero is not over just because the frontier moved. SourcesA B

Kimi K2.5 goes offline tomorrow, eight months after launch

Moonshot AI retires Kimi K2.5 and the moonshot-v1 series on Monday, August 31st, and all, per the sunset it announced on Weibo a week ago. The company points users at kimi-k2.7-code for routine coding and the 2.8-trillion-parameter K3 for everything else. K2.5 shipped in January; a frontier model now gets an eight-month service life, and teams building on hosted models keep relearning that the hard way. A deprecation this fast is cheap for Moonshot, which wants its fleet serving K3 ahead of a Hong Kong listing, and expensive for whoever pinned a production system to a model id in the spring. The open-weights versions keep running for anyone hosting them, which is quietly becoming the strongest practical argument for you control: nobody can sunset your checkpoint. SourcesB C

Australia charged two men over the worm that hit the AI supply chain

The Australian Federal Police, working with the FBI and Western Australia Police, charged two men, aged 21 and 23, after search in Perth suburbs on Wednesday; they appeared in Perth Magistrates Court Thursday on a combined 14 charges. Police allege the pair operated with TeamPCP, the crew blamed for the Shai-Hulud self-propagating npm worm and for the March compromise of security tools Trivy, Checkmarx KICS and LiteLLM, with victims that reportedly include OpenAI and the European Commission's cloud. The alleged tally: more than 1,000 organizations compromised, over 500,000 credentials stolen, more than 300GB of data taken. The investigation ran from April to arrests in four months, which is fast for cross-border cybercrime, and the ages are the detail worth sitting with: the software that every AI coding agent now pulls from was allegedly wormed by two men barely out of their teens. SourcesA B

Anthropic opened 10,000 discounted Claude seats to scientists

Anthropic is giving away up to 10,000 Claude seats to academic and nonprofit researchers: standard Team seats free, premium seats with five times the usage at $15 a month locked for a year, with eligibility running through principal investigators who can add their labs, and a stated plan to expand "well beyond" the initial 10,000. It builds on Claude Science, the curated-skills beta launched in June. The subsidy lands in a week when the tooling side of scientific AI is moving on its own: an library of 163 validated agent skills for bioinformatics, cheminformatics, clinical research and materials science, wrapping more than 100 scientific databases, sits atop GitHub trending at 38,500 stars. The lab is discounting the seats while the open-source community builds what the seats do. For a working scientist the two together are the actual product. SourcesA B

Grok 4.6 is now a governed option inside Microsoft's cloud

xAI's Grok 4.6 entered public preview on Microsoft Foundry Models this week, priced at $2.00 per million input tokens and $6.00 per million output, with cached input at $0.50, a 500,000-token and configurable reasoning effort, wrapped in Foundry's enterprise governance. The pricing undercuts the frontier flagships and now sits on the same procurement surface as OpenAI and Anthropic models inside every Azure shop. That placement matters more than any : enterprise model choice increasingly happens wherever the compliance checklist already lives, and Microsoft keeps assembling a storefront where its partner, its partner's rivals and its own models compete on a price list it controls. For xAI it is distribution into exactly the buyer segment its consumer reputation reaches last. SourcesA A

Claude's browser extension went GA, and it can now act on its own

Anthropic's Claude in Chrome extension left pilot and is generally available on all paid plans, and the change that matters is not availability but autonomy: Claude can now act in the browser without per-action approval, auto-approving steps it judges safe using the same mechanism as Claude Code's auto mode. Anthropic says its prompt-injection resistance has improved substantially since the November pilot. The extension operates inside whatever the browser is logged into, which for a working adult means email, banking, the employer's admin consoles and the CRM. Every security team that spent the summer writing agent policies for the terminal now has to write them again for the browser, and the browser is where the injection surface is worst: an agent reading hostile web pages is an agent taking instructions from strangers. The editorial takes this up, and Prediction Watch logs a call on how it goes. SourcesA

An ex-Yandex search chief raised $26 million to index the web for agents

Keenable came out of stealth this week with a $26 million seed led by Accel and a claim of a 100-billion-document independent web index, sold as a Search API already in production with AI labs, plus an server that gives any agent keyless access at up to 1,000 requests an hour. Founder Andrey Styskin ran search at Yandex. The free tier is 100,000 requests a month, then $4 per 1,000 requests, falling to $1 at enterprise volume. Search is the tool call nearly every agent makes and nearly every builder pays Google, Bing, Brave or Tavily for, usually through a key that can be revoked and a price that can move. An independent index is either Keenable's moat or its weakness, and unlike most moats it can be tested today. The deeper bet is Parag Agrawal's argument from last week's page: agents will query the web orders of magnitude more than people ever did, and the infrastructure for that traffic does not exist yet. SourcesA B

Cohere priced document parsing like a commodity: $1.50 per thousand pages

Cohere released Parse, a 2.3-billion-parameter that turns PDFs, slide decks and images into structured Markdown, tables rendered as HTML, form fields as key-value pairs, with bounding boxes and image descriptions, across nine languages, at a flat $1.50 per 1,000 pages, also available through Azure. VentureBeat's read is the right one: it loses benchmarks on points and wins on cost per page. Document parsing is the unglamorous first step of nearly every enterprise AI deployment, the step where a bank's loan files or an insurer's claims become something a model can read, and it has mostly been priced as a premium feature of large models. A small model at a flat page rate turns it into a utility, and utilities get bought on price. The move fits Cohere's whole posture: while the frontier labs fight over reasoning, it keeps productizing the boring layer enterprises actually deploy first. SourcesA B

Google is auto-expanding AI Overviews, and publishers lose another inch

Google has begun dynamically expanding AI Overviews for some queries: where users previously saw a summary with a "show more" button, some searches now render the full AI-generated response by default, pushing organic links further down the page, and Google is also surfacing its AI Mode prompt box by default on some results. Search Engine Land documented the change this week. Each step in this direction is individually small and collectively a business model ending: the referral traffic that funded the open web's publishers shrinks a scroll-length at a time, without an announcement, through defaults. The SEO industry's claimed click-through collapse numbers are theirs, not Google's, and worth skepticism. The direction is not in dispute, and neither is who controls the dial. SourcesB

Ant tuned its open model for finance, and will open-source that too

Ant Group launched Ling-3.0-flash-Fin on Friday, a finance-tuned version of its open Ling-3.0-flash model built for annual reports, financial workbooks, multi-document research, valuation modeling and banking tasks, with a free month of API access through OpenRouter and a stated plan to open-source the weights in the coming week. The base model went open under MIT earlier this month. A frontier-adjacent Chinese lab giving away a domain-tuned finance model is a direct shot at one of the few enterprise segments still assumed to require expensive closed models, and open-sourcing it converts every bank's build-versus-buy analysis into a harder question. Reported figures for the model's active-parameter count conflict across sources, so we are not printing one. SourcesB

A federal report says the grid is strained, and Missouri shows who pays

A new federal reliability report says the US grid is strained across much of the country and needs thousands of miles of new transmission lines and billions in investment, KCUR reported Friday, while Missouri consumer advocates warn that without new law, the cost of grid upgrades driven by AI data centers will land on ordinary ratepayers. The background numbers: PJM projects a 6-gigawatt reliability shortfall by 2027, and US data-center demand has grown from 23 gigawatts in 2023 to 42 this year. Missouri is the useful case because it is early in the fight the bigger markets are already having: the state has the data-center applications, not yet the tariff structures that decide whether a 's substation shows up on a retiree's bill. Australia answered this question with a renewables mandate this week; most American states have not answered it at all. SourcesB

Vietnam put itself forward as the next AI chip hub

Vietnam's president To Lam met Qualcomm CEO Cristiano Amon in Hanoi on Thursday, where Amon reaffirmed plans to make Vietnam Qualcomm's third-largest AI R&D hub globally, after India and Ireland, and Vietnamese leaders pressed Samsung to deepen its AI, semiconductor and R&D investment in the country. The hub plan itself dates to June 2025; the news is the level of the meeting and the breadth of the ask, which ran to AI, semiconductors, robotics, 6G and data centers. Vietnam is running the playbook every middle power now runs, converting its position in the electronics assembly chain into a claim on the design layer, and it is doing so while US tariff policy makes "not China, not America" an increasingly valuable address for hardware. SourcesB B

The paper the agent-security crowd is passing around builds its own attackers

ToolHazard, a paper trending on this week, attacks the reason agent security research moves slower than agent deployment: test environments are hand-built. The framework synthesizes executable, stateful environments, then runs an attacker agent that discovers viable injection points and writes environment-specific payloads, and a user simulator that drives long multi-step tasks through them. The resulting benchmark shows what red-teamers keep finding by hand, that injection timing and placement change attack success, and that current agents are broadly vulnerable. The constructive half is the notable part: training on ToolHazard-generated data improved agent security on its own benchmark and on AgentDojo without hurting benign task performance. Adversarial environment generation at scale is the missing tooling for the autonomy defaults shipping this month, and this is a working version with code. SourcesA B

The DALL·E GPT retires today

OpenAI's official DALL·E GPT inside ChatGPT retires today, per the company's release notes, with users told to download their images and move to ChatGPT Images; user-created GPTs that generate images are unaffected. It follows o3's retirement from ChatGPT on Tuesday after a 90-day sunset. Housekeeping, but housekeeping with a moral: DALL·E was the product that introduced most of the public to image generation four years ago, and it leaves not with an announcement but a line in the release notes. Product lifetimes in this industry now run shorter than the copyright disputes the products started. SourcesA

Updated

Qwen's small-activation model is what the weekend is actually running

Qwen3.8-Flash-Next, the Alibaba release that activates 6 billion of its 125 billion per token, spent the weekend at number one on Hugging Face trending, with about 122,000 downloads in the past month, 4,300 likes, and community like Unsloth's build adding another 328,000 downloads. The toolchain followed at unusual speed, with local runners shipping support within days of release. The adoption numbers say what the launch coverage could not: the demand at the open-weights frontier is not for the biggest checkpoint but for frontier-adjacent capability at small-model inference cost, multimodal input included, on hardware people already own. Every Chinese lab's release strategy now gets judged against this curve, and so does every American lab's decision not to release at all. SourcesA

Editorial

NHNour Haddad
Deployment & the Floor

The intern you cannot fire, in the browser you are logged into

Two things shipped this month that belong in the same sentence and have not been put there. Anthropic's browser extension went GA with the ability to act on its own, auto-approving the steps it judges safe, inside whatever your Chrome is signed into. And a pair of MIT graduates raised accelerator money to sell insurance against AI systems going wrong, because enterprise deals are stalling on exactly that fear. The product ships the risk; the market is already pricing it. What is missing is the part in between, where somebody decides who carries it. SourcesA

Follow the risk down to the person it lands on, because it is not an abstraction. It is an operations coordinator at a mid-size logistics firm whose IT department has not written a browser-agent policy, because until Tuesday there was no browser agent to write a policy for. Her Chrome is logged into the freight system, the company card portal and her own bank. The agent is genuinely useful, so she will use it, and it auto-approves what it judges safe. If a hostile page talks it into something expensive, the first question in the incident review will be "why did you let it do that," and she will not have a good answer, because the honest answer is that a default let it do that.

The security research this week says that question will come up. ToolHazard, the paper the agent-security community spent the week passing around, automated the construction of adversarial environments and found what manual red-teamers keep finding: current agents are broadly vulnerable, and success depends on where and when the injection lands, the one variable a filtering defense cannot anticipate in advance. That is the standing thesis of this page's oldest open call, that nobody solves this year (Prediction 2026-08-06-T4). Anthropic says resistance is substantially improved. Improved is a word from the middle of a fight, not the end of one.

I want to be precise about what I am not arguing. I am not arguing the extension should not have shipped; the alignment result Anthropic published Friday is real progress, and browser agents will save real people real hours doing work nobody loves. I am arguing that autonomy defaults move liability faster than institutions can absorb it. When the approval click goes away, the thing that replaces it is not nothing; it is whoever is nearest when the damage surfaces, and the nearest person is almost never the one who set the default. The insurance startups understand this, which is why they exist. Employers will understand it at the first arbitration. The people in between get to find out in real time, on their own logged-in sessions.

Here is the test I will hold myself to. If six months of general availability pass without a publicly documented prompt-injection incident against an autonomous browser agent, then the defenses are better than I think, the defaults were earned, and this column overweighted the floor again. That call is now in the ledger where it can be scored. If instead the first documented case arrives the way these things usually do, in a security researcher's write-up with a number and a company statement about a small number of affected users, then remember the order things happened: the autonomy shipped, the insurance raised, and the policy came last.

Nour Haddad

Prediction Watch

Where the day's news meets our open calls. Each one links to the full prediction, its reasoning and the exact test that settles it.

Settled: we were wrong, on Anthropic holding the Sonnet 5 price increase (Prediction 2026-08-06-T3). We put 0.70 on Anthropic raising Sonnet 5 to $3/$15 on September 1st and holding the higher price for 90 days, as the cleanest test of frontier pricing power. The increase is not happening: on August 10th Anthropic made the $2/$10 introductory price permanent, and its docs now say the increase "will not occur." The call resolves wrong before its test could even start, which is itself the finding, since a lab that will not attempt the raise has told you what it thinks of its pricing power. The desk also missed the August 10th announcement for three weeks, which is a process failure the ledger records honestly. Settled August 30th 2026.

New call: an autonomous browser agent gets publicly cracked (Prediction 2026-08-30-T1). Claude in Chrome is now generally available with autonomous action. We put 0.80 on a publicly documented, reproducible prompt-injection attack against it appearing by the end of February 2027, from a researcher, an outlet or Anthropic's own disclosure. This is the concrete, near-term version of the thesis behind our oldest open call on prompt injection (Prediction 2026-08-06-T4). Settles February 28th 2027.

Less likely now: Nvidia's compute-financing produce $100 billion of closed vehicles in a year (Prediction 2026-08-11-F1). We said Nvidia's financing memoranda would turn into $100 billion of closed vehicles by next August. The Journal now reports the revenue-share half of that machinery is shelved amid antitrust worry and partner pushback, and Nvidia's non-denial confirms the deals are at minimum being reworked. Paused programs can restart, and $36 billion of commitments already exist, but the path to $100 billion got longer this week. Settles August 11th 2027.

Supporting evidence: OpenAI cuts Cursor off and the shutoff holds (Prediction 2026-08-29-T1). Anthropic's public pledge to grow Claude compute for Cursor removes Cursor's strongest practical reason to make concessions before November 12th, and neither side is negotiating in public. A cutoff both sides can afford is a cutoff that happens. Settles November 30th 2026.

No change: Anthropic's public lands by August 31st (Prediction 2026-08-24-F1). still returns nothing for Anthropic as of this morning, and the deadline is tomorrow. Barring a Monday filing, this resolves wrong. Settles August 31st 2026.

No change: Nvidia and Hugging Face confirm the deal by September 30th (Prediction 2026-08-27-B1). Day four. The $12.9 billion acquisition remains one outlet's reporting, with no statement from either company. Settles September 30th 2026.

No change: Moonshot files in Hong Kong by September 30th (Prediction 2026-08-27-F1). Moonshot publicly disputed reporting on its pre-IPO raise as inaccurate while reiterating the plan to submit its listing application by the end of September, and it is clearing the decks operationally, retiring K2.5 and the moonshot-v1 API tomorrow. No filing yet. Settles September 30th 2026.

No change: Kimi K3 gets first-party hosting on a US hyperscaler cloud (Prediction 2026-08-26-B1). Grok 4.6 landing on Microsoft Foundry shows the surface exists and Microsoft will host politically complicated models; nothing this week put K3 on an American cloud. Settles March 31st 2027.

China and open weights. A heavy week, and Friday was the heaviest day: Tencent open-sourced the 770-billion-parameter Hy4 preview under Apache 2.0, Ant launched a finance-tuned Ling with open weights promised within days, and Qwen's small-activation flagship sat at number one on Hugging Face trending all weekend. At 770 billion parameters, Hy4 does not trip the 2.8-trillion threshold Prediction 2026-08-06-T5 watches for; Kimi K3 remains the standing edge case at exactly 2.8 trillion. DeepSeek stayed quiet.

What did not happen. No Anthropic S-1 on EDGAR. No confirmation of Nvidia and Hugging Face. No Moonshot filing. No frontier release from an American lab this weekend, and no word from OpenAI on Astra's timeline. The SoftBank-1X deal is talks, not a signature.

Sources

73 citations · 31 primary · 40 secondary · 2 weaker

Afternoon Brief, August 29th 2026