PyyanThe AI Index

Pyyan / News

What happened in AI

A dated log, newest first. Every item was read from its source before it was written up, and every one carries the link so you can go and disagree with it.

68 stories · 2 straight from the source · showing 61 to 68, page 7 of 7

The days

By kind

Who keeps appearing

14 August 2026

1mo ago
Open weightsAlibaba

The most downloaded model of the month is dense, 27B, and fits on one machine

27Bdense, Apache 2.0, runs on one machine

While the frontier competes on trillions of parameters, the model people are actually downloading is 27 billion, dense, and small enough to hold on one machine.

Qwen3.8-27B is a dense model rather than a mixture of experts, Apache 2.0, with a 262K native context extending to about a million. There is no hosted price because there is nothing to host: it is weights. It has accumulated millions of downloads and a long tail of quantised and fine tuned derivatives.

Why this one is different

Dense is the point, not a limitation. A mixture of experts is cheaper per token at scale but awkward to serve: memory has to hold every expert while only a few run. A 27B dense model behaves predictably on one GPU, quantises cleanly, and does the same thing every time, which is what matters when it is running on your own hardware rather than someone's API.

Dense is the point, not a limitation.

How we got here

  1. 2024 to 2025Mixtures of experts take over the frontier, optimised for serving many users at once.
  2. 3 Aug 2026Qwen3.8-Max, 2.4 trillion parameters, for people renting compute by the hour.
  3. 14 Aug 2026Qwen3.8-27B, dense and Apache 2.0, for people who own the machine.

What it does and does not mean

Download counts are not usage. A number that large includes mirrors, quantised rebuilds, automated pipelines and a great deal of curiosity, and a 27B dense model is still meaningfully behind the frontier on hard reasoning. What it does show is that the open weight audience optimises for something the leaderboards do not measure: not the best model, the best model that fits.

Hugging FaceCapital & Computetwo sources
Open this story on its own page →

13 August 2026

1mo ago
ModelsGoogle

Google's cheapest new model has a price that doubles on New Year's Day

2×the price, from 1 January 2027

Gemini 3.7 Flash costs $0.75 per million input tokens. In January it will cost $1.50, and Google said so at launch.

Introducing Gemini 3.7 Flash · Google for DevelopersGoogle's launch film for the model. It covers what Gemini 3.7 Flash does, not the January price increase this story is about.

The introductory rate is $0.75 in and $3.75 out per million tokens. Both double, to $1.50 and $7.50, on 1 January 2027. The model replaces 3.6 Flash as Google's volume tier, aimed at the high throughput work where cost per token decides whether a product is viable at all.

Why this one is different

The scheduled increase is unusual and worth taking seriously. Model pricing normally falls, and buyers have built four years of planning assumptions on that. A published date on which a rate doubles is a different kind of commitment: it is honest, it is plannable, and it quietly tells you the launch price was below cost.

Model prices normally fall. This one has a date on which it doubles.

How we got here

  1. 2023 to 2025Per token prices fall steadily and steeply, and everybody builds their forecasts on the assumption that they keep falling.
  2. 21 Jul 2026Gemini 3.6 Flash arrives as the volume tier.
  3. 13 Aug 20263.7 Flash replaces it with an introductory rate and a published expiry.

What it does and does not mean

One price schedule is not a trend, and Google has not said the increase applies to anything else. Introductory pricing is also a normal commercial tactic rather than a confession. But it is worth noticing what it implies: if a launch rate needs an expiry date, the economics of serving a frontier model at volume are tighter than four years of falling prices suggested, and anyone whose product only works at $0.75 has a deadline they did not choose.

Capital & ComputeAI Release Trackertwo sources
Open this story on its own page →

12 August 2026

1mo ago
ModelsxAI

Grok 4.6 raised its cached input price, which is the number most people never read

$0.50per million cached input tokens

Everybody compares models on input and output price. For anything agentic, the number that decides the bill is the third one, and Grok 4.6 put it up.

Grok 4.6 has a 500K token context at $2 in and $6 out per million tokens, and cached input raised to $0.50 per million.

Why this one is different

Cache pricing is where long context actually gets paid for. An agent that re-reads a large context on every step pays the cached rate far more often than the fresh one, so a change there moves a real bill much more than a headline rate does. Anthropic moved the same lever in the opposite direction three weeks later, cutting Fable 5.1 cache reads to a quarter, and took up to 45% off an agentic run by doing it.

Cache pricing is where long context actually gets paid for.

How we got here

  1. 2024Prompt caching arrives across the major APIs, and long context stops being priced as if every token were new.
  2. 12 Aug 2026Grok 4.6 ships with cached input raised to $0.50.
  3. 1 Sep 2026Anthropic cuts Fable 5.1 cache reads to a quarter of the previous rate, moving the same lever the other way.

What it does and does not mean

A cache rate on its own tells you very little, because hit rates depend entirely on how an application is written, and two teams on identical pricing can see bills that differ several fold. Nobody outside xAI has published measured costs on 4.6. The point worth keeping is narrower: the comparison everybody makes, input against output, is not the comparison that determines what agentic work costs.

Capital & ComputeAI Release Trackertwo sources
Open this story on its own page →

11 August 2026

1mo ago
Open weightsNVIDIA

NVIDIA released the weights, the training data and the recipe

3Bactive parameters, and the recipe shipped too

"Open weights" usually means the weights and nothing else. NVIDIA published the training data and the recipe as well, which is a different word.

Nemotron 3.5 Lightning is a 30 billion parameter mixture of experts activating roughly 3 billion per token, on a hybrid Mamba and transformer architecture. It scores 24 on the Artificial Analysis Intelligence Index, up from 15 for Nemotron 3 Nano, and NVIDIA claim performance close to gpt-oss-120b at about a quarter of the parameters. Weights, training data and recipes all ship under NVIDIA's OpenMDW-1.1 licence, alongside NeMo Switchyard, an open router for systems that use more than one model.

Why this one is different

Almost nobody publishes the data. Weights let you run a model; the recipe lets you rebuild and modify it, and the data lets you audit what went into it. It is also worth naming the incentive plainly: NVIDIA sells the hardware. A company whose revenue comes from the compute has a straightforward reason to make good models cheap and reproducible, and that is a structurally different motive from every lab selling tokens.

A company that sells the compute has a reason to make the models cheap.

How we got here

  1. 2024 to 2025Open weights become normal. Open data does not, and reproducing a released model from scratch remains impossible for outsiders.
  2. Early 2026The Nemotron 3 line establishes small hybrid Mamba and transformer models aimed at agents rather than chat.
  3. 11 Aug 2026Lightning ships weights, data and recipe, plus Switchyard for routing between models.

What it does and does not mean

OpenMDW-1.1 is NVIDIA's own licence, not an OSI approved one, so read it before assuming it behaves like Apache. Publishing training data also does not settle where that data came from, which is the question currently in front of several courts. And the Intelligence Index is one aggregate. What is genuinely unusual is the completeness: this is the closest a major vendor has come to letting somebody else rebuild the model rather than merely run it.

NVIDIACNBCfrom the source itself
Open this story on its own page →
Open weightsLightricks

An open video model that generates the sound at the same time as the picture

22Bopen weights, 4K HDR, audio generated jointly

Most AI video is silent, and the sound is added afterwards by a second model that never saw the picture. LTX-2.5 makes both at once.

LTX-2.5: The World Model the World Builds On · LTX_io, Lightricks' own channelLightricks' launch film for the model, published by the maker. It is a showcase reel rather than an independent test of the claims above.

A 22 billion parameter asymmetric dual stream diffusion transformer that generates video and audio jointly through bidirectional cross attention. Native multishot scenes, automatic duration and 4K HDR output. It handles text to video, image to video, video to video, text to audio and audio to video. A distilled build makes a ten second clip in roughly six to eight seconds on NVIDIA superchips. The text encoder is Gemma 4 12B. Open weights, with commercial use free under $10m of revenue.

Why this one is different

Joint generation is the whole claim. Adding sound afterwards gives you audio that approximately matches, which is why so much generated video has footsteps landing slightly wrong. Generating both from one model, with each conditioned on the other, is a harder problem and the reason lip movement and impact sounds line up. Shipping that as open weights, with a day one ComfyUI integration, puts it somewhere the closed video models are not.

Adding the sound afterwards is why so much generated video has footsteps landing slightly wrong.

How we got here

  1. 2024Text to video becomes usable, and silent. Audio is a separate pipeline bolted on later.
  2. 2025 to 2026Veo and Sora add native synchronised audio, and keep the weights closed.
  3. 11 Aug 2026LTX-2.5 releases open weights doing joint video and audio, into ComfyUI on day one.

What it does and does not mean

Free under $10m of revenue is a revenue gate, not an open source licence, and the threshold is the vendor's to move. The six to eight second figure is also for a distilled model on NVIDIA superchips, which is not what most people have. And joint generation is a claim about architecture that nobody outside has independently measured. What is real is availability: the first competent open weight video model with sound, which changes who can make this rather than how good it gets.

LTXComfyUI Wikitwo sources
Open this story on its own page →

10 August 2026

1mo ago
SafetyOpenAI

OpenAI built a model to find zero-days, and made you apply to use it

95%of exploit chain prompts completed, against 1.5% for Sol

The same model, asked the same questions, went from completing 1.5% of offensive security tasks to 95%. The difference was training it to.

GPT-5.6-Cyber is built on GPT-5.6 Sol and trained specifically for finding zero-day vulnerabilities and building exploit chains. On OpenAI's own testing it completes 95% of prompts involving exploit chain development, privilege escalation and authentication bypass, against 1.5% for Sol. It is available only through Daybreak Red, the applicant-vetted tier of OpenAI's defender programme, alongside a Daybreak Blue tier giving general models guardrails tuned for defensive work.

Why this one is different

This is the first of the three gates, and the pattern it started is now the industry's actual cyber policy. Anthropic followed on 1 September with Mythos 5.1 and Project Glasswing, Google on the 2nd with Fairwind. Three labs, twenty three days, three separate application processes and no shared standard for who qualifies.

Three labs, twenty three days, and no shared standard for who qualifies.

How we got here

  1. 2024 to 2025Labs refuse offensive security requests broadly, which frustrates legitimate researchers and stops very little else.
  2. 10 Aug 2026OpenAI ships a model trained for it, behind Daybreak Red.
  3. 1 Sep 2026Anthropic's Mythos 5.1 removes classifiers for vetted US organisations under Project Glasswing.
  4. 2 Sep 2026Google's 3.8 Flash Cyber goes to trusted defenders through the Fairwind Program.

What it does and does not mean

A 95% completion rate is a capability measurement, not a safety one. It says the model is good at this, and says nothing about whether the vetting works, how it is audited, or what happens when an approved account is compromised. Nor does the gate stop an adversary with resources from training their own. What the release established is the governing mechanism for frontier cyber capability, and it is a vendor's application form.

OpenAISecurityWeekfrom the source itself
Open this story on its own page →

16 July 2026

2mo ago
PolicyGoogle · OpenAI · Anthropic · Microsoft · European Commission

The EU ordered Google to give rival AI assistants what Gemini already had

11Android capabilities opened to rival assistants

An assistant that cannot be woken by voice, cannot see the screen and cannot be set as default is not really competing with one that can. The EU has ordered Google to hand over all three.

Binding orders under the Digital Markets Act, adopted 16 July and enforceable immediately, require Android to give third party assistants the same interoperability Gemini has: voice activation, system level screen context, access to core apps including Gmail and Calendar, and the right to be set as default. 11 capabilities in total, reaching users from July 2027 with Android 18, and concurrent hotword detection with Android 19 by August 2028. A separate order requires Google to share anonymised ranking, query, click and view data from Search with rival search and chatbot providers on fair and non discriminatory terms, from January 2027.

Why this one is different

Most AI regulation so far governs what a model may do. This governs who gets to reach a user, which is a distribution question rather than a safety one, and distribution is where assistants are actually won. It also arrives as binding orders concluding a specification process, not as a fine after the fact, which means the remedy is a set of engineering requirements with dates attached.

Most AI regulation governs what a model may do. This governs who gets to reach a user.

How we got here

  1. 2023The Digital Markets Act comes into force, designating gatekeepers and their core platform services.
  2. 27 Jan 2026The Commission opens six months of specification proceedings against Google over Android and Search.
  3. 16 Jul 2026Binding orders adopted, enforceable immediately, covering 11 Android capabilities and Search data sharing.

What it does and does not mean

Nothing reaches a phone for a year, and some of it not for two. Orders can be appealed, implementations can be technically compliant and practically useless, and the Commission has had to reopen remedies before. This is also Europe only, so an assistant's capabilities may soon differ by continent. What has changed is the theory of the case: the assistant layer is being treated as a market that can be foreclosed, and that framing will outlive this particular order.

TechTimesAI Weeklytwo sources
Open this story on its own page →

9 July 2026

2mo ago
MoneyQTS · Blackstone · Compass Datacenters

A $100bn data centre campus died over how far apart two newspaper notices were printed

2,100acres, roughly $100bn, abandoned

Virginia law requires public hearing notices to run in a newspaper at least six days apart. Prince William County's were not. That sentence ended a $100bn project.

The Digital Gateway was to be roughly 2,100 acres and 22 million square feet of gigawatt scale data centre capacity, one of the largest ever proposed. A judge revoked the zoning authorisation in August 2025 over the improper notices, the Virginia Court of Appeals upheld it on 31 March 2026, Compass Datacenters walked in May, and in July QTS, owned by Blackstone, withdrew its remaining petitions. The project is dead.

Why this one is different

Every AI compute announcement is a number attached to a building that does not exist yet, and this is what the gap between those two things looks like. Not a funding failure, not a demand failure, not a technology failure. Years of litigation and a procedural defect in how a county advertised its own hearings, against a backdrop of sustained local opposition.

Every compute announcement is a number attached to a building that does not exist yet.

How we got here

  1. 2023Prince William County approves the rezonings after contested hearings and organised local opposition.
  2. Aug 2025A judge revokes the zoning over public notices not spaced six days apart.
  3. 31 Mar 2026The Virginia Court of Appeals upholds the ruling.
  4. May 2026Compass Datacenters exits, leaving QTS alone on the project.
  5. 9 Jul 2026QTS withdraws its petitions. The project officially dies.

What it does and does not mean

One project failing is not a pattern, capacity is being built at pace elsewhere, and none of this says demand is not real. But set it against the compute commitments now being signed, tens of billions at a time on multi year terms, and the lesson is where the risk actually sits. The constraint on AI infrastructure is not capital or chips. It is land, power and the people who live next to it, and none of those are on the roadmap.

Virginia BusinessTom's Hardwaretwo sources
Open this story on its own page →