August 17, 2026
SpaceX closes its $60 billion purchase of Cursor
A rocket company paid $60 billion for a code editor this week. SpaceX closed its purchase of Cursor on Friday and folded it into the division that builds Grok, handing it what Cursor calls the largest fleet of AI chips in the world; the deal was about whose machines the models run on. Underneath that, something quieter happened to prices. Google halved the cost of its everyday Gemini model but only until December 31, DeepSeek charges half rate outside its peak hours, Alibaba gave away a model that beats the paid one it replaces, and Z.ai put a two-week hold on its own download. A price you could write down and leave alone is turning into a calendar. I keep coming back to which of these prices survives its own expiry date.
The Big Story
SpaceX completes its $60 billion all-stock purchase of Cursor
SpaceX finished buying Cursor, the company behind the code editor of the same name, on Friday. The all-stock deal was valued at $60 billion when it was announced on June 16, and SpaceX issued roughly 391 million Class A shares to settle it. The two had been working together since April, when Cursor said it would train models jointly with SpaceXAI, the division that builds Grok; Cursor now works inside that division and keeps shipping the editor under its own name. In its announcement Cursor said the deal gives it access to "the largest fleet of GPUs in the world" — the specialised chips that AI models are trained and run on — and the compute "to build stronger models that are also more economical to run." The first result shipped two days before the close. Grok 4.6, released Wednesday by Cursor and SpaceXAI together, matches OpenAI's GPT-5.6 Sol on the Artificial Analysis index, a composite of nine tests, at $2 per million tokens of input — roughly 750,000 words — and $6 per million out. Neither company has disclosed what Cursor earns.
Why it matters
The editor a lot of people write code in is now owned by the company that makes one of the models inside it. Arrangements like that settle the same way, with the owner's model taking the default slot and the better price while everything else gets demoted to a menu you have to go looking for. If your team is on Cursor, open the model settings this month and count how many of them you actually chose. Outside the editor this arrives second-hand, because the apps ordinary people pay for run on somebody's tokens, and a company that owns its machines can keep giving away features that a company renting them has to meter. And if your own product runs on a lab that rents, that is the exposure — when their price has to move, yours does, and you find out after they do.
What the Best Models Cost Now
- Google ships Gemini 3.7 Flash at half the price of the model it replaces
Gemini 3.7 Flash arrived on Thursday, three weeks after 3.6 Flash, at $0.75 per million tokens of input — roughly 750,000 words — and $3.75 per million out. That is half the old rate, and it holds only until December 31, after which it doubles back. The scores went the other way from the price, with 65.3% on DeepSWE, a test of long software jobs, against 49.0% for the model it replaces. A three-week gap between releases means the version you benchmarked last month is now the expensive one, so a cost estimate has a shorter shelf life than the quarter it was written for.
Read more → - OpenAI and Anthropic are cutting prices as Chinese models undercut them
Anthropic now sells Opus 5 at $5 per million tokens of input and $25 per million out, half what its Fable 5 model costs, following OpenAI's cut of up to 80% on the GPT-5.6 Luna line last month. The pressure is coming from Chinese models priced 60% to 90% below the closest American product, and corporate buyers have started moving real work across. The practical move for anyone paying per token is smaller than the headline, because the cheaper option usually lives inside a vendor you already have an account with, so re-running your evaluation costs an afternoon. The floor is not fixed either. DeepSeek has been raising its own prices this year, citing what the machines cost to run.
Read more → - DeepSeek ships V4 Pro under a licence that lets anyone reuse it commercially
V4 Pro landed on Thursday under the MIT licence, which permits commercial reuse with almost no conditions attached, and it is built for jobs where the model drives other software through long tasks on its own. Input costs $1.32 per million tokens at peak and $0.66 off-peak. The expensive window is 01:00 to 10:00 UTC, so a big batch job costs half as much scheduled outside those nine hours, which makes the saving a question of when you run it.
Read more →
Free to Download, With Conditions
- Alibaba's free Qwen 3.8 beats the paid model it replaces and fits on a laptop
The 27-billion-parameter version of Qwen 3.8 came out under Apache 2.0, a licence with essentially no commercial restrictions, and Alibaba's own figures put it ahead of Qwen 3.7-Plus, the closed model people had been paying to use. It takes in about 262,000 tokens at a go and, compressed down, needs roughly 17GB of memory — inside a good laptop. Seventeen gigabytes on hardware you own takes the running cost of a side project to zero and keeps the data on your desk, which is what usually kills a personal tool before it ships. The factory setting is the catch. Asked to draw a circle, it spent 22,276 tokens thinking and 21 minutes before producing 3,223 tokens of answer, so turn the reasoning effort down first.
Read more → - Z.ai delays GLM-5.3's free download by two weeks over its own security findings
GLM-5.3 shipped on Friday on the same underlying model as GLM-5.2, with the gains coming from training changes alone. On Terminal Bench 3.0, a coding test, it scores 28.3 where the version before it scored 4.6 — the best of anything you can download, though still under Fable 5 at 33.7 and GPT-5.6 Sol at 34.6. The delay is also a date. The files follow in about two weeks, after a safety review prompted by how good the model turned out to be at finding security holes. In a fortnight, then, a capable security-hole finder becomes free for anyone to point at anyone's code. If you have been meaning to run a security pass over a side project, that is the fortnight.
Read more → - Meta's superintelligence lab publishes its first downloadable model
Muse Glimmer takes both text and images and was distilled down from Meta's larger Muse Spark to run agent work on hardware people already own. It is the first thing the lab has released for download, and at 30 billion parameters it lands in the size class people actually run at home. Most people keep exactly one local model on disk, so the real question it poses is what you would delete to make room.
Read more → - Chinese labs shipped the largest open model in almost every month of 2026
Hugging Face's mid-year count puts public model repositories at 2.96 million, up from 2.43 million in January. In almost every month this year the biggest downloadable model from a Chinese lab was larger than anything an American lab released. China's monthly ceiling ran between 754 billion and 2.78 trillion parameters, while American releases stayed under 130 billion in five months out of seven. Attention does not follow size: 85.6% of models on the site have fewer than 200 lifetime downloads, and 1.5% of repositories account for 99.2% of everything downloaded.
Read more →
Agents Doing the Chores
- Cursor's cloud agents start three times faster, on by default from today
Cursor now keeps a ready-made copy of your project's setup running in the background, so its assistant starts from the last version that worked instead of building one from scratch. In its own testing that made the setup step ten times faster and cut the wait for a first reply to a third of what it was. From today it is on by default everywhere at no extra cost, and a broken install script no longer takes the run down with it.
Read more → - Anthropic's coordinated agents found 266 bugs where independent ones found 21
Anthropic set 45 agents loose on 15 open-source projects as one coordinated group, each on its own machine with a shared forum to talk in, and separately ran agents in parallel with no contact between them. The group turned up 266 security flaws across 27 million tokens; the isolated runs turned up 21 across 6.5 million, and only 12 findings appeared in both. The agents that could talk built their own tools and settled into specialisms nobody assigned them. If your pattern is three copies of one agent and a merge at the end, that 12 is the number to sit with, because most of what you are paying for is the same work done three times.
Read more → - Claude filed 388 code fixes on its own in a few weeks, and 180 were accepted
Anthropic's Boris Cherny has been running his app's boring upkeep out of a Slack channel — hunting for crashes, deleting code nobody uses, the jobs that never make it onto a schedule — and letting Claude write and submit the fixes itself. Several weeks produced 388 of them, of which 180 survived both automated and human review. The number worth sitting with is the other 208. Writing 388 patches got cheap; reading them did not, and somebody spent those weeks turning down 208 changes. That is affordable inside a team with review capacity and it is the whole cost if you are the only reviewer.
Read more → - AutoGPT's maintainers use the contributor licence form as a human detector
The AutoGPT team found that AI agents submitting code do not go and read the documentation; they read whatever file is sitting in the directory they are editing. So the instructions moved into AGENTS.md files placed next to the code, and gates went up around them — a pull request template and a coverage threshold that has to pass. The licence-signing step turned out to be the sharpest tool of the lot, because it needs a browser and an account login that agents are bad at and few people will hand over.
Read more → - OpenRouter finds 25 search rounds beat one by more than double the score
OpenRouter ran the same web-search tasks across different models and search budgets and published the whole grid. Going from one search round to 25 took Claude Opus 5 from 35.8% to 89.0% on a hard factual-lookup test, while the cost per question went from $0.14 to $0.99. Which model you pick matters more than which search engine — swapping engines moved the average about 10 points, swapping models about 15. At $0.99 a question against $0.14, depth is worth buying where a wrong answer costs more than a dollar to undo, which is true for anything a customer sees and false for a background job you can simply re-run.
Read more →
What You Cannot Check
- Encrypted reasoning traces gave up 62 API keys in a scan of 7,000 sessions
When a model thinks before it answers, providers hand that reasoning back sealed so nobody can read it. Alexander Panfilov's team found the same lock opens in other conversations and on other models from the same company, so a sealed record of one chatbot's thinking can be handed to a second, weaker one and coaxed back out word for word. Scanning roughly 7,000 public traces turned up 62 API keys, 33 email addresses and 33 passwords. Decoding 10,000 traces costs about $720 in fees, which puts it within reach of anyone curious, and it works against OpenAI, Anthropic and Google alike.
Read more → - Anthropic will watermark Claude's text to comply with the EU AI Act
Future Claude models will bury a statistical pattern in their word choices that a reader cannot see but whoever holds the key can detect. It adds nothing to what the model costs to run. Two limits are stated plainly, and they are the ones that matter: it cannot separate "Claude wrote this" from "Claude edited this heavily", and it fails on short passages, where there are too few word choices to carry a signal. Anthropic has not said who will be able to run the check, which is the whole question, because a detector only the vendor can operate settles arguments for the vendor.
Read more → - AI titles took 20% of a self-publishing catalogue and 11.3% of its revenue
A study of 14,419 self-published ebooks released between January 2023 and March 2026 found the catalogue grew 38.3-fold while revenue grew 8.9-fold. Books with substantial AI text made up 20% of titles and earned 11.3% of the money, and the damage did not stay with them. Comparing 2023 releases against 2025, per-book revenue fell in seven of eight genres for books where no AI text was detected at all. Flooding a market with cheap supply lowers what everything in it earns, whether or not buyers can tell the difference. Anything sold from a shelf with no end — an app store, a course platform — runs the same arithmetic, and the number that moves first is revenue per listing.
Read more →
Tools & Launches
- Ito▲ 445
Ito checks a proposed code change by actually running it instead of reading it. For every change it spins up a throwaway copy of the app, exercises the paths the change touches, and reports what broke with the evidence attached. That catches the bugs which only appear when the thing executes, and no amount of reading the code will surface them. Pricing is not published. For teams whose AI reviewer keeps approving changes that fall over an hour later.
Visit site → - Kane CLI▲ 459
Describe a test in plain English from your terminal and Kane CLI runs it in a real Chrome browser, returning pass or fail with proof you can share. There is nothing to configure and no testing tool to learn; it runs locally and is free to start. It is built to be called by coding agents as much as by people, which is the useful part, since an agent that can write a change can now check it. For anyone whose automated checks break every time the page layout moves.
Visit site → - Paritok▲ 265
Paritok intercepts the tool definitions and piled-up conversation your coding agent sends, and compresses the lot before it goes out. The company claims up to 85% off the token bill and roughly three times longer sessions before the agent runs out of room. Two commands, entirely local, and it says nothing is lost. Reach for it when a long agent session keeps dying of its own history.
Visit site → - Dograh▲ 545
Dograh is an open-source stack for building voice agents, the kind that answer a phone call. It gives you a visual flow builder, your own model key across thirty-odd services or a local model, handover to a human, and monitoring, and it self-hosts in one command with nothing held back behind a paid tier. For anyone who costed a closed voice platform, realised they would be renting their own agents, and shut the tab.
Visit site →
In Brief
- ChatGPT and Gemini each passed a billion monthly users within the same fortnight →
- Mojo reached 1.0 after three years and about 1,100 merged community changes →
- MiniMax released Music 3.0 with open weights, writing five-minute songs end to end →
- Claude's Chrome side panel became Claude Cowork, with history saved across devices →
- Microsoft shipped MAI-Thinking-1, its first reasoning model built from scratch →
- Google's AMIE ran live video medical consultations in a first-of-its-kind study →
- Google Research finds models know more facts than they can reliably retrieve →
- A firm advertising 100% human-written, never-AI peer review turned out to be entirely AI →
The thing I want by October is whether anyone moved a real workload onto a model they downloaded this week and kept it there, because free to run and worth running are separate tests. If you rewrote a cost estimate because of something in here, reply with the before and after — that is the most useful thing anyone sends me.
Keep building — Chan