Another Weekly AI Newsletter: Issue 78
OpenAI previews GPT-5.6 Sol as Washington asks to slow roll. AI patches the bugs it finds. OpenAI releases Jalapeño. A24 Films x Deepmind. xAI ships /goals. Claude Tag makes a splash in the enterprise
Washington’s grip extended from Anthropic to OpenAI.
OpenAI previewed GPT-5.6 Sol, its next-generation model. The preview introduced Sol alongside Terra and Luna variants, with METR publishing a predeployment evaluation of Sol’s capabilities.
The White House asked OpenAI to slow-roll the release. OpenAI limited the rollout after the government request over safety concerns, but said publicly that such restrictions should not become the norm.
TechCrunch’s read: it’s no longer Anthropic vs. OpenAI. The framing shifted to labs vs. a government that now weighs in on when frontier models ship.
The thread: a week after the Anthropic ban, the same dynamic reached OpenAI in a softer form, a request to wait rather than an order to stop. The throughline is that a frontier release is now a government conversation, which is exactly what every lab and enterprise spent last week preparing for.
AI crossed from finding software bugs to fixing them.
OpenAI’s full GPT-5.5-Cyber set a new state of the art on the CyberGym benchmark. It topped Anthropic’s Mythos 5 as part of Daybreak, a push to automate the patch rather than just the discovery.
Codex Security has scanned more than 30 million commits across 30,000 codebases. Sam Altman framed it as solving security problems instead of only finding them, with over 500,000 fixes auto-confirmed.
Patch the Planet put OpenAI’s models and Trail of Bits on 30-plus open-source projects. The initiative hardened cURL, Go, and Python and merged dozens of patches.
The thread: for a decade the hard part of security was finding the hole, and the bet this week was that the model can write the fix too. The same capability cuts both ways, which is why the Five Eyes alliance spent the week warning that frontier models could supercharge offensive hacking.
The compute land grab went into overdrive.
OpenAI unveiled its first custom chip, built with Broadcom. The in-house silicon is OpenAI’s move to own its inference stack, alongside an internal accelerator it calls Jalapeno.
SpaceX will rent open-source lab Reflection $6.3B of AI compute. Reflection pays $150M a month through 2029 for Nvidia GB300 chips at the Colossus 2 data center near Memphis.
Amazon committed another $13B to AI infrastructure in India. The investment deepens a week of India bets and adds to the global data-center build-out.
Cursor is training its own model, with SpaceX, and Groq confirmed a $650M raise and neocloud pivot, while Cursor’s keynote headlined the in-house model.
SK Hynix overtook Samsung as South Korea’s most valuable company. High-bandwidth memory demand rewired the chip hierarchy, and SpaceX even named its planned AI satellite network Starmind.
The thread: the week’s deals all point one way. When access to a model, or the chips under it, can shift on a policy decision, owning the compute, the silicon, and the weights is the hedge everyone is suddenly paying for.
Agents got real permissions.
xAI shipped /goal, multi-agent coding in Grok Build. The command hands Grok an objective and coordinates planners, implementors, and reviewers to reach it.
Google’s Interactions API hit general availability, a unified endpoint for Gemini agents with server-side state, while DeepMind also shipped computer use in Gemini 3.5 Flash.
AWS shipped Bedrock AgentCore Payments, so agents can pay per call mid-task, and Notion launched External Agents with Claude and Cursor.
Self-Harness lets agents rewrite their own rules. The framework reports gains up to 60%, while Stanford’s DeLM cut multi-agent costs 50% with no central orchestrator.
The warnings arrived with the tools. CrewAI told teams to stop handing agents raw database credentials, and Simon Willison reframed prompt injection as “role confusion.”
The thread: the tooling to give agents money, goals, and self-modification landed the same week as the arguments about what happens once they have it. TechCrunch called the result a world getting “loopy,” with swarms of agents running in the background instead of waiting for a prompt.
The AI video race and the culture fight both escalated.
Alibaba’s HappyHorse passed Sora to No. 2 in global video rankings. The 1.1 model is API-first with reference-to-video character consistency.
ByteDance answered with Seedance 2.5, the Doubao team’s model doing native 4K and roughly 30-second clips with up to 50 reference inputs.
Google DeepMind bet $75M on Hollywood, a research partnership with A24 pitched as working with artists rather than around them, while Adobe acquired Topaz Labs and shipped agentic workflows across Creative Cloud.
SZA called musicians who support AI “disgusting.” Her comments followed claims that 200-plus of her songs were used as training data, sharpening the artists-versus-AI backlash.
The thread: the labs courting Hollywood and the artists rejecting AI are reacting to the same fact, that generative video crossed from demo to production, and tools like Runway’s new Agent 2.0 are pushing it into ad and campaign workflows.
⭐ Featured: Anthropic put Claude inside Slack as a teammate.
Claude Tag is Anthropic’s attempt to make Claude an always-on coworker instead of a chat window you visit. In Slack, Claude joins a workspace as a member: you grant it access to chosen channels, tools, data, and even codebases, then tag @Claude to delegate a task the way you would a colleague, while it keeps persistent context on how your company works. It is in beta for Claude Enterprise and Team customers.
What makes it notable is the shift in interaction model. Andrej Karpathy called it a new paradigm, Claude operating inline with the rest of org-wide activity rather than in a separate app, and argued the hard part is the under-the-hood plumbing across tools, integrations, compute, and memory. TechCrunch’s read is sharper: Claude Tag is learning your company one Slack message at a time. The upside is an assistant that actually knows your context. The open question is the one every broadly-scoped agent raised this week, which is how much access you are comfortable giving something that quietly reads everything.
🎙️ Worth a Listen
Google DeepMind’s podcast on agentic economies. Host Hannah Fry digs into what happens when millions of AI agents start negotiating, transacting, and delegating to one another, and how to diversify their decision-making to avoid “AI groupthink.” A useful frame for the week agents got payments and goals.
Quick Hits
Krea released open weights for Krea 2 | Krea — an undistilled base plus a fast distilled version, built to fine-tune.
Mistral shipped OCR 4 | Mistral — bounding boxes, block classification, and confidence scores across 170 languages.
Z.ai’s GLM-5.2 beat GPT-5.5 on coding benchmarks and landed in Cursor | VentureBeat — open weights at roughly a sixth of the cost.
NVIDIA’s Nemotron 3 Ultra ranked among the top open models | NVIDIA — the open-weights field kept filling in.
Amazon is testing a Hindi Alexa+ in India | TechCrunch — localization aimed at a massive market.
Meta named Cred’s Kunal Shah WhatsApp chief and invested $900M | TechCrunch — a sharp India tilt.
General Intuition raised $2.3B to train agents on video games | TechCrunch — games as a world model for real-world agents.
Agility Robotics is going public via a $2.5B SPAC | TechCrunch — humanoids hit the public markets.
Databricks’ former AI chief says he can cut AI’s power bill 1,000x | TechCrunch — a swing at inference economics.
Anthropic’s Claude is winning over paid consumers | TechCrunch — eating into a market ChatGPT owned.
New data suggests engineering jobs are AI’s most resilient | TechCrunch — against the “AI kills coding” narrative.
Oracle cut 21,000 jobs and cited AI | TechCrunch — now on the running list of 2026 AI-attributed layoffs.
An ECB study found AI’s wage and job impact muted so far | U.S. News — a data point against the loudest displacement claims.
Satya Nadella said AI monopolies are a problem | TechTimes — concentrated AI power is unstable, he argued.
Google redesigned the search box for the first time in 25 years | VentureBeat — AI Overviews and AI Mode merged into one box.