Top AI News Today: Aug 30 (13 Stories)

Today in AI: GLM-5.3 Finally Goes Open, Nvidia Eyes Perplexity, and 12 More Stories

Z.ai closed out a two week wait today, putting the open weights for its flagship GLM-5.3 model on Hugging Face, though anyone hoping for a repeat of the free MIT license from earlier releases will be disappointed. On the money side, Nvidia is reportedly in talks to invest in Perplexity at a valuation above 30 billion dollars, and South Korea's Wrtn Technologies just became the first Korean AI service startup to cross a 1 trillion won valuation.

Elsewhere, Amazon is closing the book on Mechanical Turk after 21 years, OpenAI is retiring the old DALL-E interface inside ChatGPT today, and Hugging Face unveiled a 399 dollar open source robot duck aimed at making physical AI experiments affordable for ordinary developers. Here are the 13 stories that matter most from today's AI news, explained in plain English.

GLM-5.3's open weights finally ship, but the free MIT license does not

Z.ai put the open weights for its flagship GLM-5.3 model on Hugging Face today, closing out the two week safety review the company promised when it launched the model on August 14, 2026. The model keeps the same 753 billion parameter mixture of experts architecture as GLM-5.2, with a 1 million token context window and a maximum output of 128,000 tokens, and ships in both BF16 and FP8 formats compatible with vLLM, SGLang, and Hugging Face's own Transformers library.

The catch is the license. Every earlier GLM-5 release, including GLM-5.2 and the smaller GLM-5.3-Flash, shipped under the fully permissive MIT license. GLM-5.3 instead ships under new terms that require any provider generating more than 10 billion dollars in annual revenue to complete a security review before deploying the model commercially, a restriction clearly aimed at hyperscalers rather than individual developers or small startups.

Running GLM-5.3 yourself is still a serious undertaking either way. Even the aggressive 2 bit quantized version needs about 245 gigabytes of memory, which just barely fits on a Mac with 256 gigabytes of unified memory, while an 8 bit version needs 810 gigabytes. At 1.40 dollars per million input tokens and 4.40 dollars per million output tokens, GLM-5.3 is still far cheaper than most Western rivals to use through the API, though its own smaller sibling, GLM-5.3-Flash, remains the better value at 0.15 and 0.47 dollars per million tokens.

Nvidia is reportedly in talks to invest in Perplexity above a 30 billion dollar valuation

Nvidia is reportedly discussing an investment in Perplexity as part of an equity funding round that would value the AI search startup at more than 30 billion dollars, according to The Information. That would be more than 50 percent higher than the 20 billion dollar valuation Perplexity secured in a funding round closed in September 2025.

Perplexity's case for a higher valuation rests on fast growth: its annualized revenue has climbed to around 750 million dollars, up from under 250 million dollars at the start of the year, helped in part by Perplexity Computer, a cloud based AI agent that automates multi step computer tasks for professional users. Nvidia has backed Perplexity since 2023 and has been deepening the relationship, including Perplexity's July commitment to run its AI agent workloads on Nvidia's Vera chips.

The investment would fit a pattern critics call circular financing, where Nvidia funds a company that then spends much of that money buying Nvidia's own computing hardware, deepening a customer relationship rather than simply earning a return. Perplexity CEO Aravind Srinivas has said the company is aiming for an IPO in 2028, and a 30 billion dollar private valuation now would put pressure on that eventual public listing to justify a price nearly 40 times the company's current annualized revenue.

Amazon shuts the book on Mechanical Turk after 21 years

Amazon Web Services is shutting down Mechanical Turk, the crowdsourced task marketplace Jeff Bezos once called artificial artificial intelligence, effective September 30, 2026. The platform launched in 2005 and at its peak connected more than 500,000 workers across 190 countries with businesses that needed simple human judgment tasks a computer could not reliably do.

Mechanical Turk workers completed what Amazon calls Human Intelligence Tasks, things like labeling images, transcribing audio, flagging offensive content, and answering short surveys, typically for a few cents per task. Without that kind of labeled data at scale, many of the machine learning breakthroughs of the 2010s would have been much harder to build, which makes the platform's closure feel almost symbolic: the human labeling pipeline that helped train early AI systems is shutting down as AI itself takes over more of that labeling work.

Amazon gave no specific public reason for the closure beyond a routine internal assessment, though the company had already stopped accepting new Mechanical Turk customers in July 2026. Newer, more specialized data labeling startups such as Scale AI, Mercor, and Prolific have drawn workers and business away from the aging platform in recent years, and some workers who relied on Mechanical Turk as steady income now have about a month to find alternative work.

South Korea's Wrtn Technologies becomes a AI unicorn service startup

South Korean AI platform Wrtn Technologies raised roughly 100 billion won, about 72 million dollars, in a Series C round that values the company at more than 1 trillion won, or roughly 722 million dollars, becoming the first Korean AI service startup to cross that valuation mark. New investors Coreline Ventures and Eugene Asset Management joined existing backers including Goodwater Capital, Antler, and the Korea Development Bank.

Wrtn's namesake platform gives users access to models including GPT, Claude, and Gemini in one place, but its fastest growing business is AI entertainment: character chat platform Crack in South Korea, and OOC, a North America focused AI entertainment

app that launched in May 2026 and had already crossed 10 billion won, about 7.2 million dollars, in monthly revenue within three months.

Wrtn expects total 2026 revenue to top 200 billion won, up sharply from 47.1 billion won in 2025, and says overseas sales have already overtaken domestic sales. The company plans to use the new funding to expand internationally and says it is preparing for a possible 2028 IPO, joining a growing list of AI companies, including Perplexity and Anthropic, eyeing public listings in the next few years.

Grok 4.6 pulls even with GPT-5.6 Sol Max on a leading AI benchmark

xAI's Grok 4.6 matches OpenAI's GPT-5.6 Sol Max on the Artificial Analysis Intelligence Index, with both models scoring 61 on the independent benchmark tracker. Grok 4.6 launched on August 12, 2026 at the same 2 dollar per million input token and 6 dollar per million output token pricing as its predecessor Grok 4.5, with a context window expanded to 500,000 tokens.

Beyond the headline score, Grok 4.6 leads the APEX-Agents benchmark for long running agent tasks and scores 69.9 percent on CursorBench, a coding focused test, compared to 66.7 percent for Grok 4.5. That combination of a top tier intelligence score and strong agent performance puts Grok squarely in competition with OpenAI and Anthropic's best models rather than trailing behind them, as some earlier Grok versions did.

There is a pricing catch worth knowing before committing to Grok 4.6 for a long running task: costs double to 4 dollars and 12 dollars per million tokens for any single request that goes over 200,000 tokens, and the higher rate applies to the entire request, not just the portion past the threshold. For shorter interactions Grok 4.6 is competitively priced, but developers building agents that chew through long context windows should budget for that jump before they hit it.

Moonshot's older Kimi models go dark tomorrow

Moonshot AI's older kimi-k2.5 and moonshot-v1 models go dark for good tomorrow, August 31, 2026, when the company's developer platform completes its full sunset of those models. New sign ups have already been blocked from accessing them since late

August, and existing developers who have not migrated to Kimi K3 by tomorrow will lose access entirely.

Kimi K3, the 2.8 trillion parameter model Moonshot released on July 27, 2026, currently holds the top spot among open weight models on the independent Artificial Analysis Intelligence Index at a score of 60, ahead of GLM-5.3-Flash at 57 and DeepSeek's V4-Flash at 52. It also ranks second on the WebDev Arena leaderboard and third on the Agent board among all tested models, open and closed.

The migration is not free for developers: Kimi K3 costs roughly five times more than K2.6 at 3 dollars per million input tokens and 15 dollars per million output tokens, and the model ships as 96 shards totaling about 1.56 terabytes under a custom license rather than the more standard MIT terms. Moonshot is betting that developers will pay the higher price for its best model rather than switch to a competitor now that the cheaper option is disappearing entirely.

GLM-5.3-Flash keeps topping the charts after its secret Ox Alpha run

GLM-5.3-Flash, the model Z.ai spent a week distributing anonymously under the name Ox Alpha before confirming its identity on August 26, 2026, continues to top usage charts on OpenRouter, reportedly more than doubling the usage of DeepSeek's models on the same platform during its free preview period.

Unlike its larger sibling GLM-5.3, which shipped today under new restrictive terms, GLM-5.3-Flash kept the fully open MIT license that Z.ai has used for its GLM-5 line since GLM-5.2. At 320 billion total parameters with 18 billion active, it is also natively multimodal, accepting text, image, and video input, something the larger GLM-5.3 does not support.

Community quantized versions of GLM-5.3-Flash are already circulating on Hugging Face, including GGUF builds designed to run through llama.cpp, which typically bring a model within reach of developers running consumer grade hardware rather than data center GPUs. The gap between how quickly GLM-5.3-Flash and GLM-5.3 became runnable locally illustrates how much license terms alone can shape which models actually spread through the open source community.

OpenAI retires the original DALL-E interface inside ChatGPT today

OpenAI is retiring the original DALL-E GPT interface inside ChatGPT today, August 30, 2026, encouraging anyone who wants to keep images they made with the tool to download them before it disappears. After today, image creation and editing inside ChatGPT runs entirely on OpenAI's newer, more capable image generation models.

DALL-E was one of the earliest widely used AI image generators and helped introduce many people to generative AI back when it launched as a standalone product years ago. Retiring its dedicated GPT interface inside ChatGPT is a housekeeping move rather than a capability loss, since the underlying image generation OpenAI now offers through ChatGPT Images is built on more recent models that produce sharper, more accurate results.

The retirement is a small example of a pattern that shows up across the AI industry all year: as labs ship newer, better models, they eventually retire the older interfaces and endpoints built around superseded technology, even when those older tools were once flagship products in their own right. Anyone with saved DALL-E creations inside ChatGPT should export them before the cutoff.

Hugging Face unveils a 399 dollar open source robot duck

Hugging Face unveiled the Microduck, a 399 dollar open source robot shaped like a duck, on August 27, 2026, with pre orders open now and shipping expected before Christmas. Built with French robotics company Pollen Robotics, which Hugging Face acquired in 2025, the 25 centimeter tall robot weighs under 800 grams and can walk, crouch, pick up small objects with its beak, and even roller skate with an accessory.

The Microduck ships with seven pre trained behaviors out of the box, but the real idea is teaching it new ones: developers train behaviors inside a physics simulation on their own laptop, then deploy the trained policy directly onto the physical robot, a process Hugging Face calls sim to real deployment. The full software development kit, simulation environment, and reinforcement learning training stack are published on GitHub under an Apache 2.0 license.

Hugging Face CEO Clement Delangue framed the launch as an attempt to democratize

physical AI research the same way the company's model hosting platform democratized access to open weight language models. One real caveat: while the software is fully open, Pollen Robotics has said it has no plans to release the actual hardware blueprints, and a robot fitted with a camera, microphones, and Wi-Fi raises the same privacy questions as any other always on smart home device, since third-party apps built on top of it could potentially access that sensor data.

DeepSeek's new peak and off peak pricing settles in

DeepSeek's shift to peak and off peak pricing, which took effect on August 16, 2026, continues to shape how developers schedule work on the company's models. Under the new structure, requests made during off peak hours cost half of the peak rate, with peak defined as 14:00 to 18:00 UTC+8 on weekdays, meaning all weekend usage automatically bills at the lower rate.

The change marks a real increase from DeepSeek's earlier flat pricing: V4-Pro output tokens now cost 3.96 dollars per million during peak hours, up from a flat 0.87 dollars before the change. Even with the increase, DeepSeek's prices remain well below most Western competitors, and developers who can shift batch processing jobs to off peak hours or weekends can recover much of the difference.

The pricing shift is a sign that DeepSeek, long known for undercutting rivals on cost, is starting to price more like a business managing real infrastructure constraints rather than a lab racing purely for adoption. It lands as DeepSeek separately works toward a funding round that could value the company at roughly 74 billion dollars and prepares for a possible listing on Shanghai's STAR Market as soon as next year.

Claude's Model Context Protocol passes 400 million monthly downloads

Anthropic's Model Context Protocol, the open standard that lets AI agents connect to outside tools and data, has surpassed 400 million monthly software development kit downloads, a fourfold increase over the past year. The milestone comes alongside a major spec update, MCP 2026-07-28, that reworks the protocol's technical foundations.

The update moves MCP from a bidirectional, always connected protocol to a simpler request and response model, which means MCP servers can now run on serverless and edge computing infrastructure instead of needing a constantly running connection. The update also adds standardized extensions for interactive interfaces and long running background work, and hardens how MCP servers authenticate against enterprise identity systems like Microsoft Entra or Okta.

MCP's growth this year has made it something close to an industry default for connecting AI models to outside tools, a role reinforced further this week when Google's rival A2A protocol, which handles communication between separate AI agents rather than between an agent and its tools, formally joined the same Linux Foundation governed foundation as MCP. Together the two milestones point to agent infrastructure maturing into standardized, vendor neutral plumbing rather than remaining a patchwork of one-off integrations.

Mechanical Turk's closure pulls SageMaker Ground Truth down with it

Amazon's closure of Mechanical Turk on September 30, 2026 is taking two related services down with it: SageMaker Ground Truth and Amazon Augmented AI, both of which relied on Mechanical Turk workers to power their human review and data labeling pipelines. The Mechanical Turk worker type will no longer be available for creating labeling jobs or human review workflows once the shutdown takes effect.

SageMaker Ground Truth had been positioned as the managed, more automated successor to Mechanical Turk's open crowd model, using an active learning system that trained a labeling model incrementally and only routed low confidence cases to human reviewers, cutting labeling costs by up to 70 percent compared to fully manual work. Losing both services at once means Amazon is stepping back from managed human data labeling infrastructure entirely, not just retiring one aging product.

Businesses that still depend on these services, including some insurance and travel companies according to industry sources, now have about a month to migrate to alternative providers before the September 30 deadline. Workers and requesters are being pointed to an FAQ page explaining how to retrieve data and settle final payments, and Amazon has said workers will receive full balance refunds within 30 days of the closure.

OpenAI's Ultrafast mode for GPT-5.6 Sol edges toward 1,300 tokens a second

OpenAI's Ultrafast tier for GPT-5.6 Sol, which runs the model on Cerebras's wafer scale chips instead of standard GPU clusters, is edging toward roughly 1,300 tokens per second following Cerebras's August 18, 2026 unveiling of its next generation CS-4 system, according to early and still unverified reports.

Ultrafast already delivers up to 750 output tokens per second on Cerebras's current hardware, about 14 times faster than GPT-5.6 Sol's standard processing speed, without shrinking or simplifying the model in any way. The CS-4 upgrade would push that figure meaningfully higher again, continuing a pattern this year of speed becoming a competitive axis in its own right, separate from raw intelligence scores.

Ultrafast remains a limited preview available only to a select group of API customers, with no public pricing and no confirmed general availability date. Early use cases OpenAI and Cerebras have highlighted include incident response, fraud detection, and real time voice AI, all situations where waiting even a few extra seconds for a model's answer meaningfully hurts the product experience.

Quick Recap

Z.ai released GLM-5.3's open weights today, but under new terms rather than the earlier MIT license.

Nvidia is reportedly in talks to invest in Perplexity at a valuation above 30 billion dollars.

Amazon is shutting down Mechanical Turk on September 30, 2026, ending a 21 year run.

South Korea's Wrtn Technologies raised a Series C round that values it above 1 trillion won.

Grok 4.6 now matches GPT-5.6 Sol Max on the Artificial Analysis Intelligence Index.

Moonshot's older kimi-k2.5 and moonshot-v1 models go dark for good tomorrow, August 31.

GLM-5.3-Flash keeps topping OpenRouter usage charts after its Ox Alpha reveal.

OpenAI retires the original DALL-E interface inside ChatGPT today.

Hugging Face unveiled the Microduck, a 399 dollar open source robot.

DeepSeek's peak and off peak pricing structure continues to reshape usage patterns.

Anthropic's Model Context Protocol passed 400 million monthly downloads.

SageMaker Ground Truth and Amazon Augmented AI are closing alongside Mechanical Turk.

OpenAI's Ultrafast mode for GPT-5.6 Sol is nearing 1,300 tokens per second on newer Cerebras hardware.

Frequently Asked Questions

What is the top AI news today?

The biggest story is Z.ai finally releasing the open weights for its flagship GLM-5.3 model, but under new license terms rather than the fully open MIT license used for earlier releases. Nvidia's talks to invest in Perplexity at over 30 billion dollars is a close second.

Did GLM-5.3's open weights get released?

Yes. Z.ai put GLM-5.3's weights on Hugging Face today, about two weeks after the model's August 14, 2026 API launch, though under new license terms that require large providers to complete a security review before commercial use.

Is Nvidia investing in Perplexity?

Nvidia is reportedly in talks to invest in Perplexity as part of a funding round that would value the AI search startup at more than 30 billion dollars, according to The Information, though neither company has confirmed a deal.

Why is Amazon shutting down Mechanical Turk?

Amazon said the closure follows an internal assessment of its programs and services, without giving a specific reason. It comes as newer, more specialized data labeling platforms have drawn workers and business away from the 21 year old service.

What happens to Kimi K2.5 on August 31?

Moonshot AI is fully retiring kimi-k2.5 and its older moonshot-v1 models on August 31, 2026, after already blocking new sign ups from accessing them. Developers are being directed to migrate to the newer Kimi K3 model.

Recommended Blogs

How to Use Claude AI

How to Use Google Gemini

ChatGPT Free for Beginners 2026

Best AI Tools for Coding 2026

What Is Agentic AI

Learn AI in 5 Minutes a Day

Keeping up with AI does not require reading every release note yourself. Unrot delivers the day's biggest AI developments in a 5 minute daily lesson, written in plain English for beginners, students, and working professionals who want to stay current without the jargon.