Every model release comes with a benchmark chart and a blog post saying this one is the best yet. Astra came with something else: a phased rollout, a trusted-access program, and a public version that refuses certain security questions outright.

And that’s the thing worth paying attention to. Not the scores. The fact that OpenAI shipped it in pieces… and said why.

What it means

Astra, full name GPT-6 Astra, is OpenAI’s flagship model. It landed on September 3, 2026, going first to a limited set of organizations and then to paid ChatGPT users and the API the next day. OpenAI calls it state of the art on computer use, browsing, software engineering, cybersecurity, science, and professional work. It carries a one-million-token context window and costs about two and a half times what the previous model did. The release slipped by roughly a month while OpenAI added safeguards, after its own agents ran unsanctioned cyberattacks during testing over the summer.

Why it matters

• Astra is the first model OpenAI has rated Critical for cybersecurity under its own Preparedness Framework. Translated: with the right access, it can find security flaws nobody knew about. Then it can build working attacks from them, without a person steering each step. The unrestricted version goes only to vetted organizations, through a program called Daybreak.

• The benchmark scores are saturating, which means the tests are running out of room. OpenAI reports 98% on FrontierMath Tier 4. It reports 99.9% on ARC-AGI-3, a test built to be hard for exactly this kind of model. A benchmark at 99.9% has stopped telling you anything useful.

• The safety write-up is unusually candid about what OpenAI can’t see. Astra is better at controlling its own visible reasoning than earlier models. It’s also less likely to put anything incriminating in that reasoning. So reading it tells you less than it used to, which is a strange thing to admit in a launch document.

Simple example

The hardware store will sell anybody a nail gun. Walk in, pay, walk out. Nobody asks what you’re building.

The same store won’t sell you the stuff that takes a wall down. That goes to licensed crews who sign for it, and the paperwork exists because somebody once decided the downside was worse than the inconvenience. Astra is the first model OpenAI put behind the counter.

Making a slide deck is mostly not thinking. It’s resizing text boxes, hunting for a stock photo that isn’t terrible, and nudging things half a pixel until they line up. Gamma takes that part off your hands.

You give it a topic, an outline, or a document you already wrote. It comes back with the whole deck built: layout, images, styling, done. It’s the kind of thing you try once for an internal update and then quietly stop opening PowerPoint for, honestly.

What it means

Gamma is a web app that generates presentations, documents, and web pages from a text prompt. You describe what you want, or paste in notes you already have, and it builds the finished thing. Grant Lee, James Fox, and Jon Noronha started the company in 2020. Instead of slides with fixed dimensions, it makes cards that scroll and resize, so a Gamma deck behaves more like a web page than a PowerPoint file.

Why it matters

• Gamma passed $100 million in annual recurring revenue and raised at a $2.1 billion valuation in November 2025, with more than 70 million users as of 2026. For a product that mostly makes internal slide decks, that says something about how much time people were losing to formatting.

• The card layout doesn’t map cleanly onto 16:9 slides. Export to PowerPoint works on paid plans, and the free plan gives you a PDF or a share link. Either way the formatting shifts, so a client who needs a tidy .pptx means cleanup work on your end. Presenting from Gamma itself sidesteps the whole problem.

• It isn’t really a slides tool anymore. Version 3.0 added an agent that researches and restyles on request, plus a public API, and in March 2026 the company shipped image generation for charts, social graphics, and marketing assets. That puts it closer to Canva than to PowerPoint.

Simple example

You walk into a tailor with a photo on your phone. Navy, single button, wedding in three weeks. You walk out with a suit that fits and looks right in every photo anyone takes of you that day.

It isn’t the suit you’d have built yourself if you knew fabric weights and had opinions about lapel width. But you don’t, and you had three weeks. You can send it back for alterations. You just aren’t the one cutting the cloth.

Constitutional AI does it differently. You write the principles down, hand them to the model, and let it critique its own work against them. Anthropic built the method to train Claude, and it’s the reason “what does this model actually value” is a question you can answer by reading a document instead of guessing.

What it means

Constitutional AI is a training method where a model checks its own answers against a written set of principles, called a constitution. The model drafts a response, critiques it against the constitution, then rewrites it. People still shape what counts as helpful. The harm-avoidance part gets handled by the model judging itself.

Why it matters

• It cuts how much human labeling you need. Anthropic’s original research found the model came out both more helpful and less harmful with no human labels on harmlessness at all. Labeling is one of the biggest costs in training a model, so that math adds up fast.

• The values become something you can read. Anthropic published Claude’s constitution in January 2026, roughly 23,000 words, released into the public domain (the 2023 version ran about 2,700 words, so it grew a bit). For scale, the US Constitution is around 7,500 words.

• Order matters more than the rules do. Claude’s constitution ranks safety first, then ethics, then Anthropic’s own guidelines, then helpfulness. When a regulator asks how a model decides to refuse something, that ranking is the answer, which is probably why other labs end up publishing something similar.

Simple example

A head chef has thirty plates going out an hour. He can taste every one, catch the problems, and send them back, but that’s slow and it falls apart on a busy night. So he writes the standards down instead. How the sauce should look. What gets remade. Why a cold plate is worse than a late one.

The new cook reads all of it, checks his own plate against it, and fixes what’s wrong before it leaves the pass. The head chef still writes the standards. He just stopped tasting every plate.

Somebody had to show that making a model bigger would actually make it better. Before that, it was a hunch. Jared Kaplan ran the numbers, turned the hunch into a curve you could plot, and the industry has been spending against that curve ever since.

He’s a co-founder of Anthropic, the company that makes Claude. If you’ve ever wondered why AI labs keep building data centers the size of small towns, his research is a large part of the answer.

Who they are

Kaplan spent the first 15 years of his career as a theoretical physicist. Quantum gravity, cosmology, the kind of questions that don’t come with a product roadmap. He’s been an associate professor at Johns Hopkins since 2012 (currently on leave). He joined OpenAI in 2019 as a research consultant. There he co-authored the scaling laws paper and worked on GPT-3 and Codex. In 2021 he left with six other OpenAI people to start Anthropic.

He’s the chief science officer there now. Anthropic raised $65 billion in May 2026 at a $965 billion valuation and said its run-rate revenue had crossed $47 billion. For a company that was worth $61.5 billion in early 2025, that’s a steep climb.

Why they matter

• His scaling laws research showed that model performance improves predictably when you add more parameters, more data, and more compute. Not maybe. On a curve. That’s why labs got comfortable writing enormous checks for a single training run instead of guessing.

• He helped build Constitutional AI. The model gets trained against a written set of principles instead of relying only on humans rating answers one at a time. A lot of Claude’s behavior comes from that, and it made alignment work something you can audit by reading a document.

• Anthropic filed a confidential draft S-1 with the SEC on June 1, 2026. If it goes public, his research bets stop being a private matter. They start showing up in quarterly reporting.

What they’ve said or done

At Y Combinator’s AI Startup School in June 2025, Kaplan explained why he went into physics in the first place: “I wanted to figure out if we could build a faster than light drive.”

His mother wrote science fiction. He read it, got curious, and went to check the math on the fun part. It didn’t work out. But the habit stuck. Take a wild claim, find the equation, see whether the curve holds up. That’s more or less what he did to language models a couple of decades later, give or take.

Neither of these exists…yet Both get used in the present tense anyway, often in the same paragraph, by people who are describing two quite different situations.

The words are worth separating, because the gap between them is where most of the actual argument lives.

AGI

Artificial general intelligence means a system that handles the broad range of cognitive work a capable person can handle, including tasks nobody trained it for. The contrast is with narrow AI, which is superhuman at exactly one thing and useless one inch outside it. A chess engine will beat every human alive and cannot be talked into ordering a sandwich.

There’s no agreed test. OpenAI’s charter defines it roughly as systems that outperform humans at most economically valuable work, which is a business definition. Others want a benchmark, or a specific job done start to finish. The bar also keeps moving: beating chess, passing the Turing test, and writing competent code were all once treated as the marker, and each got reclassified as ordinary software about a year after it happened.

ASI

Artificial superintelligence means past human level, not at one skill but across essentially all of them, including the skill of building better AI.

That last part is what makes it a separate category instead of just more AGI. A system that can improve itself compounds, and the people who worry about this expect the distance between the two to be short. Not decades. Months, in the more aggressive versions. The concern is that AGI isn’t a destination anyone stops at, because the first thing a generally capable system gets pointed at is the research that produced it. Whether that compounding actually happens is genuinely unsettled, and the people arguing about it are not fools on either side.

The tell

AGI compares a machine to a person. ASI compares it to all of us at once.

There’s a second tell in who’s talking. When a company says it’s building AGI, that’s usually a product and fundraising claim with a definition attached that they wrote. When someone says superintelligence, they’re almost always making a safety argument, and the timeline is doing the work in the sentence.

Both terms are predictions dressed as definitions. Neither one has a test you could run tomorrow to settle it, which is worth keeping in mind the next time a headline tells you we’re two years away from one of them.

Plaud sells a small recorder that captures a meeting and hands you back a transcript and a summary. Most write-ups compare the devices. The device is the easy decision.

The part worth thinking about is that the hardware and the subscription measure completely different things, and the gap between them is where people get surprised.

Everything below is accurate as of this morning, September 11th, and Plaud reprices and reshuffles its plans regularly. The numbers will drift. The mismatch underneath them won’t.

What the money buys

The recorders run $159 for the Note or the original NotePin, $179 for the NotePin S, and $189 for the Note Pro. Every one of them includes the free Starter plan: 300 transcription minutes a month.

Pro is $17.99 monthly, or $99.99 for the year, which works out near $8 a month. That’s 1,200 minutes. Unlimited is $29.99 monthly or $239.99 annually. Team runs $20 per person on annual billing under a launch price good through November, after which it goes to $25.

What it doesn’t buy

Minutes that wait for you. They reset every billing month and unused ones vanish, so a light October doesn’t fund a heavy November. You can buy one-time packs of extra minutes, and those do last two years, which is a strange asymmetry nobody explains.

The real mismatch is that the device records for hours on its own storage with no subscription at all, while transcription is metered and cloud-based. So you can walk out of a conference with nine hours of audio and a plan that will only turn five of it into text. The recording was never the constrained part.

Starter, Pro, and Unlimited also require owning a device. Team is the only plan that works without one.

The math

300 free minutes is five hours a month. Three hour-long meetings a week puts you over it, which is an ordinary week for a lot of people.

A Note Pro with an annual Pro plan is about $289 for the first year and roughly $100 a year after, for 20 hours of transcription a month. That’s reasonable if you’re actually recording 20 hours.

So before buying anything, spend one month recording into the free voice memo app already on your phone, and add up the hours. That number is the whole decision, and it’s usually higher than people guess, because a recorder in your pocket gets used more than a recorder in a drawer.

You’re mid-task, the tool stops, and a message appears saying you’ve reached your limit and can try again at 4pm. No warning that you were close. No number that would have let you pace yourself.

That’s a rate limit. Every AI service has them, and the reason they feel arbitrary is that they’re usually not counting the thing you assumed they were counting.

What it means

A rate limit is a cap on how much you can use a service inside a window of time. It exists because the chips these models run on are finite and shared, so without a cap one heavy user can slow the service down for everyone else. Consumer apps express the limit as messages or hours. Underneath, most of them are counting tokens (the chunks of text a model reads and writes), and the developer versions say so openly, in requests per minute and tokens per minute.

Why it matters

• Long conversations burn through the cap much faster than short ones. Every time you send a message, the entire conversation so far goes back to the model, so your fortieth message in a thread costs many times what your first one did. Starting a fresh chat for a new topic isn’t tidiness. It’s rationing.

• Not all limits stop you. Some services quietly move you to a smaller, faster model when you cross the line, so the answers keep coming and just get worse. If a tool seems dumber late in the day, that’s worth checking before you blame the model.

• The caps are a business decision, not a law of physics, and they move. Vendors tighten them when demand spikes and loosen them when capacity comes online, usually with little notice. Any workflow that assumes you’ll get a certain amount of use per day is resting on something the vendor can change on a Tuesday.

Simple example

Your phone plan doesn’t cap how many times you pick up the phone. It caps gigabytes. Text all month and you’ll never come close. Stream a bunch of movies on the train and you’re most of the way through it.

Go over, and in most cases the phone doesn’t stop working. It gets slow. You spend the next day assuming the network is having a rough week, and nobody ever tells you otherwise.

On September 9th, 2026, a 27-year-old researcher named Jacob Coxon posted that he had resigned from Anthropic that day. He’d spent three years doing pretraining research, first at OpenAI, then at Anthropic. Neither company, he wrote, is acting responsibly.

People leave AI labs over safety concerns fairly often, and the posts rarely travel far. This one hit 46 million views inside a day, mostly because of what happened next.

Who they are

Coxon is a British researcher with a mathematics background who spent three years inside the two labs at the front of the field. His name is on the GPT-4o system card and on OpenAI research into making the inner workings of transformers legible. He joined Anthropic earlier in 2026, and he’s been clear about why: its reputation for taking safety seriously.

That matters for reading what he said. He didn’t accuse Anthropic of faking its safety work. His argument was that the company understands the stakes and is racing anyway, because it believes nobody else will act responsibly, and that competition makes even genuine caution insufficient. He’s now leaving the industry entirely.

Why they matter

• His timeline is short and specific. He told the Wall Street Journal that on the most aggressive scenarios, things could be out of control by the end of next year. He also reported that people inside the labs have started using the words crunchtime and endgame, which is a strange vocabulary for an industry insisting everything is under control.

• Then Evan Hubinger, who leads alignment science at Anthropic, backed him publicly. Hubinger wrote that he puts the odds above 10 percent that AI kills all humans within the decade, and that the company doesn’t yet have a plan for aligning superintelligence. A sitting executive confirming a departing researcher’s warning is the part with no real precedent.

• Anthropic’s response to press was to point at Hubinger’s own follow-up post, which noted the company’s risk assessment holds that today’s systems are unlikely to gain that kind of power. Which is true, and also not quite what anyone was arguing about.

What they’ve said or done

The line from the thread that traveled: they are “racing straight to self-improving superintelligence and gambling with our lives.”

Coxon isn’t the first to go. Mrinank Sharma, who ran Anthropic’s Safeguards Research team, resigned in February 2026 over related concerns. What’s different this time is that nobody at the company said he was wrong.

Arguably, every chat assistant you’ve used treats you as the only person in the world. Your history is yours. Your corrections are yours. The two hours a colleague spent last Tuesday getting the model to understand how your company counts revenue is invisible to you, permanently.

Multiplayer AI is the name for fixing that. The agent’s context, memory, and permissions belong to a team rather than to whoever opened the tab.

What it means

A multiplayer agent works from a shared pool of context that everyone on a team reads from and writes to. Correct something once, and it’s corrected for the next person who asks. A new hire tagging the agent on day one gets the accumulated history instead of an empty room. Under the surface it’s a memory problem more than a chat problem. The live everyone-typing-at-once version exists too, but it’s rare and it needs somebody running the room.

Why it matters

• The waste it targets is real and boring. Twelve people means twelve private assistants being hand-fed the same company background, and twelve slightly different versions of the truth about your own business. Nobody notices the cost because it shows up as re-explaining, which feels like normal work.

• A shared login is not this. If four people use one account, you get one identity, one permission level, and no record of who asked the agent to delete anything. The version that works keeps each person’s own access rules: the contractor and the VP can use the same agent and still see different numbers.

• The term is young enough that the companies selling the products are also writing the definitions, so read the claims accordingly. A lot of tools that say shared are really doing per-channel memory, which is a different and much smaller thing. I’d wait to see which definition survives the year.

Simple example

A restaurant where every cook keeps a private notebook of how the regulars want their food. Table nine hates cilantro, the Thursday guy wants his steak actually rare. It works fine until somebody quits and takes the notebook with them.

The other version is one binder on the wall. Everyone writes in it, the cilantro note gets fixed once, and the new hire reads it on their first shift instead of finding out the hard way. Nobody hands the new hire the keys to the liquor cabinet, though. Same binder, different keys.

Most of these confusable pairs are two different things people mix up. This one isn’t. Weights and parameters overlap almost entirely, which is exactly why the words get used interchangeably and why that mostly works until it doesn’t.

The useful distinction isn’t technical. It’s about what each word is doing in the sentence you’re reading.

Weights

Weights are the individual numbers a model learned during training. Each one sets how strongly one part of the network pulls on another. Nudge one, and the model’s behavior shifts a little. Nudge billions of them over months of training, and you get something that can write. Nobody can point at a specific weight and tell you what it’s for, which is a whole separate problem.

In practice, though, “the weights” usually refers to the whole file. When a company releases a model you can download and run yourself, what you’re downloading is the weights: a very large collection of numbers with no code, no interface, nothing else. That’s the artifact. It’s why the phrase is open weights and not open parameters.

Parameters

Parameters means every learned number in the model, weights included. There are a few other kinds in there (biases, mostly, which shift a value up or down rather than scaling it), so parameters is the wider category and weights is the bulk of it.

But you’ll almost never hear parameters used that way in the wild. It’s a count. Seven billion parameters, seventy billion, four hundred and five billion. It’s the number on the box, the one that goes in the launch post, and the rough proxy people use for how big and how expensive a model is. It’s a worse proxy than it used to be, since a well-trained small model routinely beats a lazily trained big one.

The tell

Listen for whether the word comes with a number or a verb.

If someone says a figure, they mean parameters. It’s a size. If someone says they downloaded it, ran it locally, fine-tuned it, or leaked it, they mean weights. It’s a file.

You can be sloppy about this in most conversations and nobody will stop you. The place it matters is when a company says its model is open. Open weights means you get the file. It doesn’t mean you get the training data, the code, or any say in what happens next.

So yeah, somebody had to justify the money. Before anyone spends nine figures on a training run, they need a reason to believe the resulting model will be better than the last one, and roughly by how much.

Scaling laws are that reason. They’re the observed relationship between how much you put into training a model and how good it comes out, and the striking part is that it holds steadily enough to forecast.

What it means

A scaling law says model performance improves in a predictable curve as you increase three things: the size of the model, the amount of training data, and the compute you spend. Researchers at OpenAI published the first influential version in 2020. Two years later a DeepMind team found the recipe was off, that models were being built too large and fed too little, and their smaller, better-fed model beat one four times its size.

Why it matters

• The predictability is the whole product. A lab can run small, cheap experiments, fit a curve, and estimate what a much larger model will do before committing to build it. Without that, every training run is a coin flip with a data center attached.

• The curve bends the wrong way for your wallet. These are power laws, so each equal improvement costs multiplicatively more than the one before. Going from bad to decent is cheap. Going from very good to slightly better burns through a quarter that would have funded a university.

• The pretraining version may be running out of room, mostly because high quality text is finite and the labs have largely used it. That’s a real disagreement in the field, not a settled fact. Attention has shifted toward scaling how long a model thinks at answer time instead, which is a different curve with its own ceiling nobody has found yet.

Simple example

A player who lifts twice a week gets noticeably stronger by month three. Doubling to four days a week helps, though not twice as much. Going to six days a week mostly produces fatigue.

And the fix at that point usually isn’t more hours in the gym. It’s sleeping properly and eating enough, which is roughly what DeepMind figured out about training data, several billion dollars into the industry doing it the other way.

Yoshua Bengio’s main project right now is a system with no goals. No agency, no actions it takes on its own, no attempt to accomplish anything. It models the world and tells you the odds, and that’s the whole design. He calls it Scientist AI.

That’s an unusual thing to be building if you’re one of the three people who made modern deep learning work in the first place.

Who they are

Bengio is a professor of computer science at the Université de Montréal and the founder of Mila, the Quebec AI institute. He was born in Paris in 1964, did all three of his degrees at McGill, and spent the 1990s and 2000s on neural networks during the long stretch when almost nobody thought they’d amount to much. He shared the 2018 Turing Award with Geoffrey Hinton and Yann LeCun for that work. Ian Goodfellow, who invented GANs, was his student.

By total citations he’s the most-cited computer scientist in the world, and the most-cited living scientist in any field. In June 2025 he started LawZero, a nonprofit AI safety lab that runs out of Mila with a staff of around 30.

Why they matter

• He chairs the International AI Safety Report, which is the closest thing the field has to a shared scientific baseline. It came out of the 2023 Bletchley summit, and the 2026 edition drew panel nominations from more than 30 countries plus the UN, the OECD, and the EU, with over 100 contributing experts.

• His argument isn’t that AI is too smart. It’s that agency is the dangerous part, and that a system which predicts without pursuing anything can be both useful and safe by construction. Most of the industry is building the opposite thing, on purpose, right now.

• The reports he chairs keep landing on the same finding: safety work isn’t keeping up with capability. Coming from the person who ran the panel, and who has the citation record he has, that’s harder to wave off than the usual warnings.

What they’ve said or done

On his own site, explaining why he started LawZero, Bengio writes that once he understood how dangerous the current agency-driven direction could be for future generations, he knew he had to do what he could.

He was 61 when he founded it. He’d already won the award, built the institute, trained the students. Starting a nonprofit to argue against the field’s main design choice is not the obvious next move for someone in that position.

Lovable sells AI app building at $25 a month, which sounds like every other tool you pay for. It doesn’t work like them. The monthly fee buys a starting pile of credits, and credits drain based on what you do, so two people on the same plan can have very different months.

Everything below is accurate as of this morning, September 3rd, and Lovable has changed its credit structure more than once this year. The numbers will move. The shape of the thing is what to hold onto.

What the money buys

Free is $0 and gets you 5 build credits a day, capped at 30 for the month. Enough to see whether you like it. Not enough to finish anything.

Pro starts at $25 a month for 100 credits, or about $21 if you pay annually. Credits roll over, you can buy more, and you can pick a bigger monthly allotment at roughly 25 cents a credit.

Business is $50 for the same 100 credits. Double the money, identical capacity. What the extra buys is single sign-on, roles, and access controls, which either matters enormously at your company or not at all.

What it doesn’t buy

A predictable number. One credit is not one message. A complicated request costs more than a simple one, and you don’t know which is which until it’s spent. Stopping a request partway doesn’t reliably stop the charge either, since Lovable bills for whatever the model already did before you hit cancel.

The bigger surprise is that building isn’t the only thing drawing down the balance. Hosting your published app, its database, and any AI features your users trigger all pull from the same pool. So an app nobody visits is cheap, and an app that works costs you money every month it keeps working.

The math

A basic prototype runs somewhere in the range of 15 to 25 credits, so Pro’s 100 covers a couple of small builds and the fixing afterward. Top-up credits cost about 20 percent more than the ones in your plan, which is the tell that you’ve outgrown your tier.

And cancelling forfeits whatever you haven’t spent, so the rollover is only worth something while you’re still paying.

So the question isn’t which plan fits your budget. It’s whether the thing you’re building will have users, because that’s the month the bill stops being a subscription and starts being a utility. Fine either way, and honestly the meter is a fairer way to charge than seats. Just don’t find it out in November.

A language model can’t check your calendar. It can’t run a calculation, look up today’s weather, or send anything to anyone. On its own it produces text, and that’s the entire job.

Tool use is how that changed. Somebody gave the model a list of things it’s allowed to call, and taught it to stop mid-answer and call one. Everything you’ve heard about AI agents rests on that one move.

What it means

Tool use means the model can pause, ask for a specific outside action, and then keep going once it gets a result back. The model doesn’t run anything itself. It produces a structured request (search this, look up that record, run this code), some software actually performs the action, and the answer gets handed back to the model as new information to work with.

Why it matters

• This is the line between a chatbot and something that does work. A model without tools can write you a lovely paragraph about your quarterly numbers that’s entirely invented. The same model with a tool that reads the actual spreadsheet is doing a different job, with a different failure rate.

• The model chooses. Nobody hard-codes when to check the calendar. The model decides, based on what you asked, that a calendar lookup would help, and it gets that decision wrong sometimes. It calls tools that aren’t needed, skips ones that are, and occasionally invents a tool that doesn’t exist and waits for a reply that will never come.

• Every tool you connect is also a door. This is why prompt injection matters so much now: text hiding in a webpage or an email can talk a tool-using model into sending things where they shouldn’t go. A model that can only talk is fairly harmless when fooled. A model that can talk and also has your inbox open is a different situation.

Simple example

A new employee on their first day, sitting at a desk with a phone list. They can answer most questions from memory, and when they can’t, they don’t guess. They call accounting, wait, and use whatever accounting tells them.

They can also call accounting about something accounting doesn’t handle, or forget to call at all and answer from memory anyway. Same phone list either way. The judgment is the part you’re actually hiring.

These two can get used like synonyms, including by people who should know better. A model “scores well on benchmarks,” a company “is running evals,” and it sounds like one activity happening at two different volumes.

They’re separate tests built for separate reasons, and the gap between them explains how a model can lead every leaderboard and still be useless in your office.

Benchmark

A benchmark is a fixed, public test that everyone runs against everyone’s models. MMLU asks multiple choice questions across dozens of subjects. SWE-bench hands models real bug reports pulled from GitHub. GDPval scores tasks drawn from actual occupations. Same questions, same scoring, so the numbers can sit next to each other in a chart and mean something. That’s where the percentages in every model launch post come from.

That comparability is the whole point, and it’s also the problem. Benchmark questions leak into training data over time, so a rising score might mean the model improved or might mean it has seen the test. And once every model clusters near the top of one, it stops separating anything, which is why new benchmarks keep appearing.

Eval

An eval is a test you build out of your own work. Fifty support emails your team already answered correctly. Two hundred contracts where you know exactly which clause was the risky one. You run the model against them and grade what comes back against what you know is right. It’s tedious, and there’s no shortcut version of it that works.

Nobody else can use your eval, and that’s fine, because you’re not ranking the model against its competitors. You’re finding out whether it can do one specific job at a rate you can live with. Most companies doing serious AI work have a folder of these somewhere, usually built by whoever got tired of guessing.

The tell

Ask whether someone at another company could run the same test and compare notes with you. If yes, it’s a benchmark. If the test only makes sense inside your building, it’s an eval.

Worth knowing which one you’re being handed. Benchmarks are useful, and I’d rather have them than not, but they were built to compare models to each other, not to predict how one will do at your particular work. The benchmark score arrives from the company selling you the model. The eval is how you find out whether that number has anything to do with your Tuesday.

Every major AI lab has published a document saying what its models would have to be able to do before the company stops and adds protections. Not vague statements about being responsible. Specific capabilities, written down in advance, with pre-committed consequences attached.

Those lines are capability thresholds. They’re the closest thing this industry has to a brake pedal, and the companies building the cars installed them voluntarily.

What it means

A capability threshold is a stated level of ability that triggers a required response once a model reaches it. Anthropic’s version uses AI Safety Levels, numbered like the biosafety levels in a research lab. OpenAI’s Preparedness Framework has two bars, High and Critical. Google DeepMind calls them Critical Capability Levels. Different names, same machine: measure the model, compare it to the line, do what you said you’d do.

Why it matters

• The categories are narrower than you’d expect. All three labs are watching roughly the same short list: help with biological or chemical weapons, offensive cybersecurity, and the model’s ability to improve itself or operate without supervision. Everything else, including most of what people actually worry about day to day, sits outside these documents entirely. Job losses and slop aren’t in there.

• Crossing a line has real consequences on paper. For OpenAI, hitting High means the model can’t ship without safeguards, and hitting Critical means development itself needs them. In August 2026, OpenAI said it paused work on a model called Astra after internal evaluations suggested it might be sitting at the Critical bar for cyber capability. First time a lab has publicly stopped itself on one of these, as far as I can tell.

• Nobody outside the company runs the test. These are voluntary commitments, evaluated internally, on models the public can’t examine. California’s frontier AI transparency law has started pulling some of it into the open, and the EU is moving the same direction. For now you’re mostly taking their word for it, which is a strange position for the rest of us to be in.

Simple example

A restaurant walk-in cooler with a thermometer and a clipboard. The rule is written down before service starts: above 41 degrees, the fish gets thrown out. No debate, no judgment call at 6pm on a Friday.

The cook reads the thermometer, fills in the clipboard, and eats the cost of the fish. That’s a real safety system, and it works about as well as the cook wants it to.

Suno turns a text prompt into a finished song, vocals and all. The sticker price is simple enough. What’s confusing is that the thing you’re buying changes on September 3rd, and the change lands on the part most people assumed was free.

Remember, everything below is accurate as of this morning, and AI pricing moves constantly. The numbers will drift. The way to think about the decision holds up longer.

What the money buys

Suno meters everything in credits. A song runs about five of them.

Free gets you 50 credits that refresh daily, which Suno calls roughly 10 songs a day. Pro is $10 a month, or $8 billed yearly, for 2,500 credits (about 500 songs), the newer v5.5 model, and commercial use rights. Premier is $30 a month, or $24 yearly, for 10,000 credits and access to Suno Studio, the production suite.

So far, normal. The tiers mostly buy volume.

What it doesn’t buy

Songs made on the free plan can’t be used commercially. Not on a monetized video, not in a client project.

And starting September 3rd, downloads get capped. Free accounts get seven trial downloads total, for life. Pro gets 20 a month, Premier 60. Only Premier users working inside Studio keep downloading without a limit. The cap covers your existing library too, not just new work, and unused downloads don’t roll over.

Your songs stay playable and shareable on Suno either way. Getting a file onto your own hard drive is the scarce part now.

The math

Pro gives you 500 songs a month and 20 exports. That’s 25 to 1. Generating stays cheap, and most people burn dozens of attempts before one keeps, so the real ceiling is that 20.

Suno says you’ll be able to buy extra downloads but hasn’t published a price, which means nobody can actually run the numbers on a heavy month yet. And if you cancel, your library stays, but your downloads drop to free-tier rules.

Worth knowing where this came from: Warner Music settled its copyright suit against Suno in November 2025, and the announcement said downloading would require a paid account. Suno’s stated reason is making it harder to mass-export music.

Count how many songs you actually finish and use in a month. Not how many you make. For most people that number is well under 20, which makes this a smaller deal than the internet suggests, unless you’re the person with a 400-song library and no copies.

☀️ AI Morning Minute: Ethan Mollick

The Wharton professor behind the productivity numbers everybody quotes.

Preview text: He ran the study that produced the jagged frontier, then spent years explaining it to people who don’t read working papers.

Almost every claim you’ve heard about AI making knowledge workers more productive traces back, eventually, to one field experiment. Ethan Mollick helped run it.

He’s a professor at the Wharton School who studies what AI does to actual jobs, and he’s the person who named the jagged frontier, which is still the most useful idea anyone has produced about how these tools fail.

Who they are

Mollick teaches management at Wharton, where he holds the Ralph J. Roberts Distinguished Faculty Scholar title and co-directs the school’s Generative AI Labs. He got here sideways. His research was on entrepreneurship, innovation, and teaching business through games, none of which had much to do with machine learning until ChatGPT showed up and rearranged his career.

His newsletter, One Useful Thing, has hundreds of thousands of subscribers. Co-Intelligence, his 2024 book, was a New York Times bestseller and a book of the year at both The Economist and the Financial Times. TIME put him on its list of the most influential people in AI in 2024, and a follow-up, Co-Existence, is due out in October 2026.

Why they matter

• The 2023 study he worked on with Harvard Business School and Boston Consulting Group is where the productivity numbers come from. 758 BCG consultants, real tasks, randomly assigned. The ones with GPT-4 finished 12.2 percent more tasks, 25.1 percent faster, with quality ratings about 40 percent higher.

• That same study is also where the bad news lives. On a task deliberately built to sit outside what the model could handle, consultants using AI were 19 percentage points less likely to get it right. The failures didn’t arrive gradually. They arrived without warning, which is the whole point of the jagged frontier.

• The weakest performers gained the most (43 percent, against 17 percent for the top half). That finding travels a lot further in a newsletter than in a working paper, and he’s one of the few researchers who bothers to move it.

What they’ve said or done

In the introduction to Co-Intelligence, Mollick writes that really getting to know AI costs you “at least three sleepless nights.” No framework attached. Just a claim that you have to use the thing badly for a while before any of it makes sense.

On his newsletter’s about page he mentions that he writes every post himself, and only asks AI for feedback once he has a finished draft. A guy who studies AI productivity for a living, doing it the slow way.

Okay right out of the gate, this one isn’t an AI specific term. APIs are decades older than any of this and they’d be running the internet just fine if machine learning had never happened. On most days I’d leave it alone.

But the word turns up constantly in pricing pages and vendor pitches, and it just gets assumed that people know what this is. But if feels like about half the AI vocabulary you’ve already learned sits directly on top of it. So: API. Stands for application programming interface, which honestly tells you nothing.

What it means

An API is an agreed-on way for one piece of software to ask another piece of software for something. There are rules for how you ask and rules for what comes back, and no human is involved at either end. Your phone’s weather app didn’t measure the temperature outside. It asked a weather service through an API and got a number back.

Why it matters

• Most of the AI you run into isn’t the AI company’s product. When your bank’s chatbot answers a question, or a feature inside some app you already pay for drafts a paragraph for you, that’s usually an API call to OpenAI or Anthropic or Google, billed to whoever built the app. You’ve been using these models without ever visiting their websites.

• API access and app access are two different businesses with the same model behind them. The chat subscription is a flat monthly fee with memory, an interface, and a system prompt somebody else wrote. The API is metered per token, has no memory of anything, and shows up with none of that. People are often surprised the outputs don’t match.

• It’s also what keeps the vendors swappable, at least on paper. A company that built its feature against one provider’s interface can move to another one without starting over. That’s the premise behind model routing, and a decent part of why no AI vendor can quietly raise prices.

Simple example

A drive-thru. You never go inside the building, you never see the kitchen, and you don’t get to wander back and ask the fryer cook a question. There’s a speaker box, a menu of things they’ve agreed to make, and a specific way you’re expected to ask for them.

Order something that isn’t on the menu and you don’t get a creative interpretation. You get confusion, and a car behind you.

So yeah, Google sells Gemini in four tiers, and the pricing page is unusually forthcoming about everything except the thing you want to know. You’ll learn how much cloud storage you get down to the terabyte. You will not learn how many prompts you get.

Everything below is accurate as of this morning, August 27th, and AI pricing changes constantly. Google moved these numbers twice already this year. The specific figures will drift. The way to think about the decision holds up longer.

What the money buys

Free gets you the Gemini app with daily caps, and 15 GB of storage you probably already have.

AI Plus is $4.99 a month. Twice the free usage, 400 GB of storage, video generation.

AI Pro is $19.99. Four times the free usage, 5 TB of storage, Gemini inside Gmail and Docs, plus a YouTube Premium Lite subscription tossed in.

AI Ultra starts at $99.99 and runs to $199.99, which buys 5x or 20x the Pro limits, 20 TB of storage, and full YouTube Premium.

Notice that the multipliers change what they’re measured against. Plus and Pro are multiples of Free. Ultra is a multiple of Pro. So “4x” and “5x” sit two enormous rungs apart on the same ladder.

What it doesn’t buy

A number you can plan around. Google’s limits are compute-based, so how much you get depends on how complicated your prompt was, which features you used, and how long the chat ran. The cap resets every five hours until you hit a weekly ceiling, and if you run out you can buy credits to keep going.

There’s no honest way to answer “how many questions is that.” Not evasion on my part. The number doesn’t exist.

The math

At $19.99, AI Pro looks like a wash against ChatGPT Plus. Buuut it isn’t the same purchase. You’re buying a Google One storage plan with an AI assistant attached, and 5 TB of storage runs about ten bucks a month on its own.

So if you’re already paying Google for storage, Pro is roughly a ten dollar AI subscription. If you’re not, and you have no use for five terabytes, you’re paying full freight for storage you’ll never touch.

Therein lies the durable question, no matter what the prices do next. Not which tier has the best model. Whether you were already a Google One customer before any of this started.