Models, tools, agents, and moves across the AI world — always with the source in plain sight, and how it connects to what we build at NeuralOS.
Anthropic pulled Claude Cowork off the desktop and onto the web and mobile: sessions live in the cloud, keep running multi-step tasks even if you power down every device you own, and data from 1.2 million sessions reveals something uncomfortable — only 8.7% of usage is coding. The rest is the office.
Anthropic redefined what it means to "publish a page": a Claude Code Artifact is no longer static HTML, but an app that pulls live data and runs actions through the viewer's own MCP connectors. That same week brought public links, editor roles, screen reader mode, and /fork.
Fireworks AI closed a $1.505 billion Series D at a $17.5 billion valuation, the largest round of the week according to Crunchbase. Their business isn't training the next giant model — it's helping companies turn generic models into specialized intelligence using their own data. The signal is sharp: capital has moved from the race for the base model to the specialization layer.
Moonshot AI unveiled Kimi K3, a 2.7-trillion-parameter reasoning model that Fortune calls the largest open-weight model available today. With a coding obsession and in-house benchmarks that brush up against the best of the West, China's open frontier says it's going toe-to-toe with Anthropic.
Thinking Machines Lab released Inkling, its first model: a multimodal Mixture-of-Experts with 975B total parameters that activates just 41B per token, with 1M context and trained on 45 trillion tokens of text, image, audio, and video. Open-weights on Hugging Face, plus an adjustable 'thinking effort' dial that decides, per request, how much the model reasons and how much it costs.
Lovable launched Agent Integrations: a published app stops being something only a human opens in a browser and becomes a tool that ChatGPT or Claude can invoke over MCP. You turn it on from a menu and the assistant connects to it with a link. Who consumes what you build just changed.
Twenty-six workers accuse Meta of letting its internal AI systems —activity monitoring, token-usage dashboards, and algorithmic performance rankings— shape who got laid off, penalizing those on medical or parental leave. It's being called the first major case of workplace algorithmic discrimination against a tech company. Meta denies it: "people decided, not the AI."
The Future of Life Institute graded nine AI giants with a seven-expert panel, 37 indicators, and a university grading scale. Nobody earned an A or a B. The top of the class, Anthropic, barely passed with a C+. Three labs failed outright with an F. It is the only independent AI safety ranking out there, and its headline is a warning to the whole industry.
Lovable launched "App user connectors": every person who uses your published app connects their own Google, Slack, Salesforce, or HubSpot account, and the app acts on their behalf, with their permissions and only their data. No more shared account for everyone. It's the multi-tenant pattern that separates a toy from a real product.
While everyone was watching the 2027 extension, a legal analysis from Latham & Watkins confirms what actually takes effect in weeks: mandatory transparency, labeling of AI-generated content, and enforcement powers over GPAI providers. Fines of up to €35M or 7% of worldwide revenue.
Cursor 3.11 introduces "side chats": parallel, persistent agent conversations that open with /side, /btw or a + button, run alongside the main thread without interrupting it, and can be re-mentioned with @ to bring their findings back to the conversation that matters. It also adds agent transcript search from the command palette and hooks for cloud agents. The IDE stops being a chat and turns into a control panel for several live agents.
Meta Superintelligence Labs unveiled Muse Spark 1.1, a reasoning model built for agentic tasks — it uses subagents, computer use, and a 1M-token window it actively manages — available through the new Meta Model API in public preview. The open-weights company is starting to offer a hosted service, and it's doing so by speaking OpenAI's format.
SpaceXAI announced Grok 4.5 on July 8 and opened it to the public the next day. Elon Musk described it as "an Opus-class model, but faster, more token-efficient, and lower cost," priced at $2 per million input tokens and $6 output. AI's frontier just got a lower price tag, and that reshuffles the math for everyone building on APIs.
Google added four capabilities to the Gemini API's Managed Agents: background execution, connection to remote MCP servers, function calling with a transition to requires_action, and credential refresh without losing state. The least flashy piece — renewing the token without restarting the environment — is what separates an agent that dies every hour from one that actually gets work done.
Microsoft announced general availability of the Sales Agent and Service Agent across Microsoft 365 Copilot, Dynamics 365, Outlook and Teams. A purpose-built MCP server, with more than 90 tools, is what drives the Service Agent. The signal is clear: MCP has stopped being an experiment and is now production infrastructure.
The Challenger, Gray & Christmas report puts it in black and white: so far in 2026, AI was the reason cited for 101,743 job cuts, close to 23% of the year's total, and it has spent four consecutive months as the number-one reason. It already doubles the 54,836 from all of 2025, and the year isn't even over. It's the flip side of the 62% pay premium: the market isn't destroying or creating blindly — it's choosing.
The first half of 2026 was the biggest in history for investment: $510B, more than all of 2025 put together. What almost nobody says out loud: more than 70% of that money went to AI, and just two companies took 43% of the total. AI capital doesn't spread out — it concentrates.
GitHub shipped a public preview of Copilot agent session streaming for Enterprise Cloud: prompts, responses, and tool calls flowing into your SIEM or Microsoft Purview, plus a REST API that returns the last 48 hours on demand. Observability for the coding agent stopped being a nice-to-have and became infrastructure.
Ever launched the tool you'd been waiting weeks for, only to have it vanish days later? That's what happened to millions with Claude Fable 5: it shipped, the US restricted it over security, and it came back with fresh armor. The most talked-about roller coaster in AI this year.
DeepSeek V4 introduces time-of-day pricing: during two daily windows, tokens —input and output— cost double, and drop off-peak. It is the first time a major lab has charged 'by the clock,' Uber-style surge pricing. Inference is starting to behave like electricity.
Together AI closed an $800 million Series C that more than doubled its valuation to $8.3 billion, with Aramco Ventures leading and NVIDIA, Vista Equity, General Catalyst and Pegatron joining. Its annual bookings already crossed $1.15 billion, and industry usage of open models tripled in a year. The signal to the market is crystal clear: running open models is no longer an experiment, it's a business decision.
Anthropic launched Claude Sonnet 5 on June 30 as the default model for Free and Pro plans. It plans, drives a browser and terminal, runs on its own — and costs a fraction of Opus 4.8. If you build with AI, your token budget just caught a break.
X launched a hosted MCP server that lets AI assistants connect directly to its API with your account's permissions. It's read-only (no posting on your behalf): they can search, read timelines, and analyze trends. A sign that every social platform will have to do the same.
Not even the Mountain View giant can keep up. Google rationed Meta's access to Gemini because it didn't have enough compute, and Meta had to tighten its belt and lean on its own model. The GPU shortage is hitting even the biggest players.
On June 26, OpenAI showed off GPT-5.6 in three flavors — Sol the powerful, Terra the balanced, Luna the cheap — with a new "ultra" mode that splits the work across subagents. The odd part: it launched as a closed preview for around 20 partners, at the request of the U.S. government.
The AI code editor you might have open right now just sold to Elon Musk's rocket company for $60 billion, all in stock. It's the largest acquisition of a venture-backed startup in history. And it started with the same problem that holds you back: not enough compute.
PwC analyzed over a billion job postings, and the verdict pops the panic bubble: roles with AI skills are growing eight times faster and pay 62% more. The market is splitting in two, and it's worth knowing which side you're on.
Europe hit the brakes on its AI Act: the strictest rules for high-risk AI have been pushed back to December 2027. If the thought of complying with everything right now was stressing you out, breathe easy — you just got 16 more months of breathing room.
At WWDC (June 8), Apple announced Xcode 27 with "the best of Anthropic, Google and OpenAI's models and agents" built straight into the editor. It also opens an Agent Client Protocol and MCP support to plug in any agent. The message: building with AI is no longer a single-vendor game.
The app that turns your ideas into software without writing code is in talks to nearly double its valuation in six months. 8 million users and $400 million in revenue: vibe coding has become very serious business.
Anthropic has made Dynamic Workflows in Claude Code generally available: Claude plans, splits the task, and spins up anywhere from dozens to hundreds of subagents in parallel, with agents that challenge each other until they converge. The killer example: porting roughly 750,000 lines of the Bun runtime in 11 days.
The Model Context Protocol shipped the release candidate of its next version and it's changing the rules: a session-free core, extensions versioned separately, apps with UI in an iframe, and auth aligned with OAuth. Translation: wiring tools to agents is about to get a whole lot easier to scale.
Join 4,200+ builders. No credit card. Build your first app with AI in minutes.