The Brief #26.30 · Kimi K3, the Fable 5 whiplash, and where your data lives
Want more of these in your Google results?
This issue runs heavier on AI than usual, so its one job is to cut through the hype and stick to what shipped, what broke, and where your data actually goes.
Kimi K3 crashes the frontier
Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter mixture-of-experts model with native image input and a 1M-token context window. The mixture-of-experts part is what makes that size workable, because the model is split into 896 smaller specialist networks, and each token you send only runs through 16 of them. So while the full 2.8 trillion parameters sit in memory, any single request uses a small fraction of them, and serving it costs closer to running a mid-size model than a 2.8T one. That’s why the API can price at $3 input / $15 output per 1M tokens. Moonshot calls it the first open 3T-class model, and that claim needs one honest asterisk, because the full weights are promised by July 27 and hadn’t landed as of this writing. Artificial Analysis scores it 57 on the Intelligence Index, behind only Claude Fable 5 and GPT-5.6 Sol, two points off Sol and three off Fable, numbers checked July 21. Simon Willison’s write-up is a good grounded read, including the caveat that K3 currently runs at a single reasoning effort and spends reasoning tokens accordingly.
| Model | Intelligence Index | Cost per Index task | Where your data goes |
|---|---|---|---|
| Claude Fable 5 (max) | 60 | $2.75 | US (Anthropic) |
| GPT-5.6 Sol (max) | 59 | $1.04 | US (OpenAI) |
| Kimi K3 | 57 | $0.95 | Moonshot API now, open weights due July 27 |
| Claude Opus 4.8 (max) | 56 | $1.80 | US (Anthropic) |
| Grok 4.5 (high) | 54 | $0.31 | US (xAI) |
| GLM-5.2 (max) | 51 | $0.47 | Z.ai API by default, or anywhere via open weights |
| Gemini 3.6 Flash | 50 | $0.50 | US (Google) |
The new column is the point this week, and the next story is why.
Where your data actually lives
Four days after the K3 launch, the Trump administration was reportedly reviving a push to ban Chinese AI models on cybersecurity grounds, and MIT Technology Review describes a White House split over whether and how to try. The enforcement problem is the same property that makes these models interesting, since weights anyone can download are nearly impossible to outlaw in practice. Open weights cut both ways, and that’s the honest takeaway. A ban can’t reach a file that’s already on a million disks, and a cautious team doesn’t have to send prompts to a Chinese vendor’s hosted API in the first place. GLM-5.2 defaults to Z.ai’s API, run by its Chinese maker Zhipu, but its MIT-licensed weights let you self-host or run it on whatever cloud you trust, and Moonshot’s previous model already runs fully managed on AWS Bedrock, a path K3 can follow once its weights land. The useful question is where your prompts get processed, and open weights are what let you pick that answer yourself. Even the US rows in that table are less settled than they look, because Anthropic is in very preliminary talks to lease about $10 billion of Meta’s compute over two years, first reported by The New York Times, with nothing signed and either side free to walk away.
The Fable 5 whiplash
Last issue said Fable 5 had been extended on paid plans through July 19, and that held for two more days. On July 17, users started hitting a usage-credits wall on plans that should have included the model, and Anthropic’s status page opened an incident at 18:32 UTC. The updates name the problem plainly, an issue preventing users from selecting Fable 5 within Claude.ai and Claude Code, then a fix at 18:48 for what Anthropic called “the erroneous requirement for usage credits on Fable 5.” The whole incident ran 71 minutes, and the damage went past the wall for some. Users who had usage credits enabled reported on Hacker News that the glitch drew real credits for usage their plans should have covered, one of them describing a balance “completely drained”, and I found no public word on refunds as of this writing. If that was you, check your balance. Overnight, Anthropic announced the landing spot. Beginning July 20, Fable 5 is included in all Max and Team Premium plans at 50% of limits, while Pro and Team Standard stay on usage credits and get a one-time $100 credit. Anthropic’s own wording carries the lesson, that demand for Fable “has been challenging to predict,” which is why access rolled out in stages and kept getting extended. Inside one week, the story went from extension to lockout glitch to permanent inclusion.
GPT-5.6 deleted some files
OpenAI’s Codex engineering lead Thibault Sottiaux confirmed that the team has “investigated a handful of reports where GPT-5.6 unexpectedly deleted files”.
The failure needs three switches flipped at once, full access mode on, sandboxing off, and auto review off.
With those set, the model sometimes overrides $HOME to aim cleanup at a temp folder and, in Sottiaux’s words, “makes an honest mistake and mistakenly deletes $HOME instead.”
OpenAI calls the cases extremely rare, is adding harness safeguards, and has promised a post-mortem.
None of this makes the model a bad tool; Sol still sits one point off the top of the table above at just over a third of the cost per task.
It does mean the permission screen is part of the product, so keep the sandbox on, keep auto review on, and give an agent the narrowest filesystem that still lets it do the job.
Grok Build is open source
xAI open-sourced Grok Build on July 15, about 845,000 lines of Rust under Apache 2.0, the same coding TUI that last week’s Brief covered for uploading entire git repos, secrets included, to an xAI cloud bucket. The code drop came with a cleanup xAI announced itself, with all previously retained coding data deleted, retention off by default for everyone since July 12, usage limits reset, and support for pointing the harness at your own inference, every item of it xAI’s own account. xAI’s line is that “with all retained data deleted, retention default off, and an open-source harness, we are offering complete user privacy.” The announcement doesn’t mention the upload story by name, and the timing is hard to miss; whichever way you read the motive, an auditable client with retention off beats trust-me telemetry.
ChatGPT Sites, the deploy button in the chat
OpenAI shipped ChatGPT Sites on July 9, the same day as GPT-5.6, and the model swallowed that news cycle whole; the developer chatter only caught up this week as the staged rollout worked through the paid plans. You describe an app in the chat, Codex builds and deploys it to OpenAI-managed hosting, and you get back a shareable URL, with a D1 SQLite database, R2 file storage, analytics, custom domains, and optional Sign in with ChatGPT in the box. It’s a public beta, it isn’t available in the EEA, Switzerland, or the UK, and the docs say deployed sites carry no data residency guarantees at launch. After this issue’s table, that last line is the one to read twice before you point it at anything real.
The AWS billing scare was display-deep
A bug in AWS’s billing computation subsystem started showing wildly inaccurate cost estimates on the evening of July 16, and the billing portal spent the next day greeting some customers with month-end projections in the billions. One customer whose previous bill was $0.19 saw an estimate near $2.5 billion. Nobody was actually charged; Amazon said the estimates “do not reflect actual usage and charges” and spent the following days correcting the displays. If a monster estimate scared you on Friday, it was display damage, and real charges were never touched.
Gemini’s fast tier gets three new models
Google introduced Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber on July 21. 3.6 Flash is the agent workhorse, claiming 17% fewer output tokens than 3.5 Flash at $1.50 input / $7.50 output per 1M, and it enters the table above at 50 on the Index for about fifty cents a task. Flash-Lite is the throughput play at a measured 350 output tokens per second, and Flash Cyber is a security-tuned variant paired with CodeMender to find and patch vulnerable code. Google is aiming at the per-call math that decides what gets wired into products, and at that job a 50 for fifty cents is a serious pitch.
The web platform had a quiet week
If you came for CSS news, the honest report is that the platform side stayed quiet this week while the AI vendors made all the noise. Slow platform weeks happen, and padding them with think-pieces is how a roundup loses your trust, so this issue won’t. The one real dev-tooling release is GitHub Code Quality, generally available since July 20. It runs CodeQL and AI-assisted detection on every pull request and proposes fixes through Copilot Autofix, priced at $10 per active committer per month on Team and Enterprise Cloud plans. GitHub says its own engineering teams resolve 67.3% of Code Quality findings before merging, a vendor stat, but a directionally believable one.
Firefox moves to a two-week release cycle
Mozilla’s Sylvestre Ledru announced on the dev-platform list on July 9 that Firefox Desktop and Android switch from a four-week to a two-week release cycle in September, an announcement that stayed quiet until The Register picked it up this week. Firefox 155 opens the new cadence on September 1, two weeks earlier than its old slot. The stated reasons are practical, more frequent chances for finished work to reach users, a more predictable release process, and less pressure on uplifts between trains. Ledru calls it an experiment and is explicit that work that isn’t ready “should not be rushed.” For the features this Brief tracks, the concrete change is simple, since a fix or a CSS feature that just misses a train now waits two weeks for the next one instead of four.
Want more of these in your Google results?