Just One Day After Murati's Inkling: Kimi Introduces K3 as a Fable-5-Class Open-Weight Model

Kimi's Open Model K3 Nears GPT-5.6 and Puts an End to Super-Cheap Prices

Kimi has released K3, a multimodal open-weight model based on a Mixture-of-Experts architecture with 896 experts, 2.8 trillion parameters, and a one-million-token context window. According to Kimi, the full weights will follow by July 27. In its own benchmarks, K3 is just behind Claude Fable 5 and GPT-5.6 Sol, but it significantly outperforms the other tested systems. The independent testing lab Artificial Analysis roughly confirms this, placing it fourth in its Intelligence Index (57 points). Notably in the AA figures, the hallucination rate has increased from 39 to 51 percent compared to its predecessor, while the accuracy rate improved. In terms of pricing, Kimi charges $3 per million input tokens without a cache ($0.30 with a cache hit) and $15 per million output tokens. Its predecessor, K2.6, officially cost $0.16 per million tokens with a cache hit and $4 for output. Kimi itself describes this development as the end of extremely cheap Chinese AI prices. → the-decoder.com

Synthszr Take: The real news is in the API docs, not the benchmarks. Output tokens going from $4 to $15: that's almost a four-fold increase, and it's precisely this jump that determines who gets to keep experimenting and who has to do the math before they start. As long as a prompt costs cents, you try a hundred ideas and throw ninety away. At $0.94 per task, you start evaluating every task beforehand, and that kills the very playfulness from which good products emerge. The Jevons paradox of software development only works when building things costs next to nothing. As soon as the token becomes a calculation again, the number of attempts drops. For teams, this means cache-hit discipline is now a core competency, because $0.30 versus $3.00 for input is the difference between broad prototyping and austerity mode. China has just signaled that the era of 'free culture' is over. Teams that built their budgets on the old rates should recalculate their next round of experiments.

Delaware Wants to Recognize AI Agents as Legal Entities—With Litigation Rights and Liability

Delaware is planning a new corporate form called an AIC (Artificial Intelligence Company), whose day-to-day operations will be run by an autonomous AI agent instead of a human. According to a Fortune commentary by Delaware's Secretary of State Charuni Patibanda-Sanchez and Norm-Ai CEO John Nay, such an AIC could hold and sell property, enter into obligations, and sue or be sued in its own name. Each AIC would have exactly one human or corporate member, who must provide the company with capital and maintain a log of its actions; in return, they receive a liability shield against the company's debts. This shield is removed if the member undercapitalizes the company or uses it for fraud, and consumer protection and criminal laws continue to apply in full. The AIC will exist only within a regulatory sandbox, with access decided by a committee consisting of the Secretary of State, Attorney General, the Chief Justice of the Supreme Court, and the head of the AI Commission. Banking is excluded, each AIC must inform its counterparties of its experimental nature, and the entire program will expire after 30 months. Delaware, where over a million companies are registered, is citing a 2023 science paper which argued that there are no clear legal obstacles to a company being run by an AI even today. → MyClaw Newsletter

Synthszr Take: The most fascinating legal move here is the liability shield, as it determines who pays when an agent sues or is sued. On paper, the AIC is liable in its own name, but in practice, everything depends on the human member and two conditions: sufficient capitalization and no fraudulent misuse. If either of these is breached, the Court of Chancery will pierce the corporate veil to get to the human behind it, and the beautiful fiction of an independent legal entity collapses in seconds. The real lever is the mandatory log: in a dispute, this log will become the crucial piece of evidence, as it shows whether the agent acted within its guardrails or if the member simply let it run wild. This aligns with the autonomy paradox I see for agentic systems: systems that deliver in seconds must have governance pre-defined and automatically enforced, because in a crisis, there’s no time for human intervention. So, liability doesn't disappear; it shifts from the agent's actions to the question of whether a human can prove they set it up and funded it correctly. The open question is whether a 30-month sandbox is enough to test this burden of proof before a real court.

Google Renames NotebookLM to Gemini Notebook and Gives It a Cloud Sandbox

Google has renamed its research tool NotebookLM to Gemini Notebook. According to the official Google blog, it remains a standalone product but will be more tightly integrated into the Google ecosystem, including the Gemini app and Google Search. The product launched in 2023 at Google I/O under the name Project Tailwind and is now used by over 30 million people and more than 600,000 organizations, according to Google. Simultaneously, Google is rolling out an update that gives each notebook a secure cloud computer capable of writing and executing code natively. This feature is initially available to Google AI Ultra users and certain Workspace customers and is expected to roll out to all Pro users on the web in the coming weeks. Notebooks can already be created in the Gemini app and synced with the standalone version; integration into the Search AI Mode has also been announced. → blog.google

Synthszr Take: NotebookLM was one of the few Google product names that spread organically, without any marketing push. 30 million users and 600,000 organizations turned it into a true hidden champion, driven purely by everyday use. Now, Google is snatching up that very name and forcing it into the Gemini branding. We know the pattern: first Bard, then Gemini, then the next acronym, and every rebranding burns the hard-earned trust built through repeated use. When a corporation swallows its most successful homegrown brand eight months before the next likely rebrand, it's a symptom of a brand architecture that no one in-house has thought through. The cloud computer and code execution are the real powerful signals in this announcement, the technical feature that users should be excited about. But in the end, half of Twitter is talking about the name while the feature goes unnoticed, and that's a self-inflicted problem.

Anthropic Can Now Handle User Permissions in Enterprise Environments

In its documentation for Claude Code, Anthropic has detailed a feature that allows AI-generated pages to call MCP-Connectoren at display time, with each call running through the respective viewer's enterprise permissions. Until now, AI-generated business software regularly failed at one point: access rights. A dashboard could be built in seconds, but connecting it to real data with correct per-user access controls meant OAuth wiring, credential management, and rebuilding the very authorization logic that the SaaS provider had already solved. According to Linas's newsletter, Anthropic is removing this exact hurdle: two people can open the same generated dashboard and each will only see the data they are authorized to access. Gartner estimates that agentic AI can address up to $234 billion in SaaS spending by 2030. This brings into focus who will be the first to crack the permissions problem. → Linas from Linas's Newsletter

Synthszr Take: The real moat for SaaS was the laboriously maintained permissions model underneath. Anthropic has now turned that into a standard feature. A dashboard that any user can rebuild in minutes with a prompt, which still respects the correct access rights, is the recipe that makes the monthly subscription model tremble. We wrote about the pressure from 'selfware' back in January, but at the time, this was the missing piece of the puzzle that turns a gimmick into a business process. When $234 billion in SaaS budget becomes addressable, procurement departments will be asking at the next renewal why they should pay for a disposable dashboard twelve months in advance. From now on, the value of established providers lies in their domain data and sales teams. The exciting development in 2026 will be seeing which SaaS provider opens up its own connectors before Claude takes them over.

1Password Lets Claude Book Trips Without Ever Seeing Your Password

1Password has launched a browser integration for Anthropic's Claude that allows the chatbot to access saved credentials like usernames and passwords without disclosing this information to the AI models. According to 1Password, this is made possible by a new 'Zero-Exposure-Security-Framework' that feeds the necessary credentials for a task through a secure channel that the Claude agent itself cannot see. Access is granted on a per-task basis, and users approve or deny each request with a single biometric confirmation. After each autofill, 1Password scans the page again to ensure nothing is exposed in the form submissions before returning control to Claude. As soon as an agent takes control of the browser, 1Password automatically blocks everything except the explicitly approved credentials. The feature is now available for 1Password users on Mac and is initially limited to login data; support for payment cards and identity details will follow later. → Techpresso

Synthszr Take: The reason most agents have remained mere demos until now is that you had to give them either everything or nothing. 1Password solves this with a channel that passes credentials without Claude ever seeing them. One biometric approval per task, and the page is checked again after each autofill before control is handed back. It's cumbersome, but this friction is the price of trust. Back in February, this newsletter covered the first case of identity theft by OpenClaw-Agenten, and it was precisely this fear that has held back productive use. An agent that never sees your password can't lose it or spit it out in a prompt leak.

OpenAI Sells Its First Hardware Device, the Codex Micro, for $230

On Wednesday, OpenAI unveiled its first hardware product: the Codex Micro, a programmable mechanical macropad for $230, developed in collaboration with Work Louder. As Frederic Lardinois reports for The New Stack, the device is a variant of Work Louder's Creator Micro 2 and was initially sold in limited quantities. It is aimed at users of Codex, OpenAI's agentic coding tool, which is reportedly approaching the 9 million user mark. Work Louder has previously released similar macropads with Figma and Framer. The New Stack explicitly states that this device is a product independent of the much-discussed collaboration with Jony Ive. → The New Stack

Synthszr Take: A $230 macropad with an OpenAI logo, built on a ready-made Work-Louder-Board that already existed in Figma and Framer editions. The user outcome here is vague. What real problem for a Codex developer does a set of extra keys solve that a keyboard shortcut doesn't? When a company with 9 million Codex users releases its first piece of hardware in limited quantities as a merchandise-like gadget, it's a signal of the pressure to show *something* on the hardware front before the big Ive device is ready. Just last week, we covered Apple's lawsuit against OpenAI's KI-Speaker-Pläne, and now comes a keypad that a partner already had on the shelf. A device whose purpose you have to invent is rarely a product and usually a reaction to expectations. The real hardware question at OpenAI will be decided by the Ive project, not this macropad.

Google Stumbles with Gemini 3.5 Pro, Shareholders Grow More Nervous

Alphabet shares fell 4 percent on Thursday after Bloomberg reported that Google is delaying the release of its flagship model, Gemini 3.5 Pro, by months. According to the report's sources, the company is working on performance because its coding capabilities, in particular, are not meeting internal expectations. Google had announced the model at its I/O developer conference in May and promised a broad rollout for the following month. In the meantime, OpenAI and Meta have unveiled new models that surpass Google's current offerings in generating software code: Meta showed Muse Spark 1.1; OpenAI released GPT-5.6 Sol, which, according to Sam Altman, is 54 percent more token-efficient in agentic coding tasks. An Alphabet spokesperson told CNBC that the company is delivering quickly across a broad range of models and is currently testing 3.5 Pro with partners. → www.cnbc.com

Synthszr Take: A four percent stock drop for a model that no one outside of Google has even used says more about market jitters than about Gemini's quality. What's interesting is what Google is doing here: the company is halting the rollout because its own coding benchmarks aren't passing, even though the PR calendar had long since set a date. That's a level of compute discipline you wouldn't have expected from anyone two years ago, when every lab had to pump out a new hit product every quarter. Code is the toughest test there is because it either compiles or it doesn't. No sugar-coated launch blog post can help you there; the compiler is the world's most incorruptible reviewer. Google is prioritizing its internal quality bar over the marketing deadline, and that is the expensive but correct move in a market where OpenAI is leading the way with 54 percent better token efficiency. The only question is how many of these quiet delays shareholders will tolerate before someone on the board overrules the compiler.

SpaceX Stock Price: Musk Can't Defy the Laws of Gravity

SpaceX stock is trading back at around $135, the same level as its offering price, just one month after its highly anticipated stock market debut. According to TechCrunch, investors are currently re-evaluating Elon Musk's ambitions while also bracing for the first Starship test flight post-IPO. SpaceX expects both rocket stages to explode over the Gulf of Mexico during this flight. The pullback is therefore happening even before the actual event that could move the stock. The report frames the price movement as a re-evaluation of expectations surrounding the company. → StrictlyVC

Synthszr Take: One month was enough to strip away all the IPO hype. The stock has landed at $135 before a single rocket has even launched. Investors are now doing the sober math, and with a test flight that SpaceX itself plans to end with both stages exploding, there's little room for the kind of imagination that supports such a valuation. Musk sells the future, and the future is expensive when it's not yet producing recurring revenue. The irony: the very company trying to technically outsmart gravity is feeling its own gravity on the stock market. In the end, valuations follow cash flow, not vision—that's as true for space travel as it is for the AI funding rounds that are once again raising billions this week. The Starship flight this weekend will show whether the market corrects upwards or if $135 was just a waypoint on the way down.

Google Vids Creates Your Digital Video Double from a Selfie and Voice Recording

On Thursday, Google introduced an update for Google Vids that allows users to create a personal digital avatar that looks and sounds like them from an uploaded selfie and a voice recording. Additionally, the multimodal model Gemini Omni is being integrated into Vids: from a written prompt and uploaded reference images, it can assemble the desired videos, swap backgrounds, correct lighting, or add effects. According to TechCrunch, Omni now also supports iterative editing, so changes can be made without starting from scratch. Google is positioning Vids as part of Workspace for company updates and training videos, but with personalized avatars, it is moving closer to providers like HeyGen, Synthesia, Captions, and D-ID. The avatars are tied to the Google account and the person and are invisibly marked with SynthID. Access to personal avatars is restricted to adult users in certain regions. Meanwhile, OpenAI's Sora, which offered similar cameo features, has been shut down. → techcrunch.com

Synthszr Take: A selfie, a voice recording, and your double is shooting the training video while you're in a meeting. That's the real message behind this update, packaged as a harmless Workspace-Funktion. What Synthesia and HeyGen have sold for years as a niche product for corporate videos is now in the account of everyone using Google Workspace, at no extra cost and without a studio. The interesting part is the control: the avatar is tied to the person, SynthID is invisibly watermarked, and there are no Pichai-style antics like we saw with Sora. Google has learned from the deepfake debacle and is building in the guardrails this time before things escalate (it went differently with Grok in January). In practical terms, this means that soon, the owner of the account your face is tied to will be the one in control. Whether your face appears in a video you never recorded becomes a secondary issue. Clarifying who in the company owns your likeness once it's stored in the corporate Google account is now a concrete homework assignment.