Anthropic launched Claude Sonnet 5, a cheaper Sonnet model meant for agent-style work: planning, using browsers and terminals, checking its own output, and finishing multi-step coding tasks with less hand-holding.

the notable bit is positioning: Anthropic says it gets close to Opus 4.8 on agentic tasks, is now the default for Free and Pro, and is on the API at intro pricing of $2 / $10 per million input/output tokens through Aug 31.

🔗 https://www.anthropic.com/news/claude-sonnet-5

Anthropic says the Commerce Department has lifted the export controls that forced it to pull Claude Fable 5 and Mythos 5 earlier this month. access is supposed to start coming back from July 1 U.S. time, after a strange 18-day pause where a model launch became a live national-security and compliance fight.

the bigger story is that frontier model access is now getting negotiated in public with Washington, not just shipped through product pages. for builders, the question is less "is the model good?" and more "who is actually allowed to use it, and under what rules?"

🔗 https://x.com/AnthropicAI/status/2072106151890809341

Etched is out of stealth with its first inference racks in customer validation: a successful A0 tapeout, $800m raised, and $1b+ in customer contracts. it says the first racks ship this summer.

the interesting part is that this is a full systems bet, not just a chip launch. Etched is pitching rack-scale inference clusters built around low-voltage compute and shared cluster memory to make frontier model inference faster, cheaper, and less power-hungry than today's GPU setups.

🔗 https://x.com/Etched/status/2071972062202343590

oasis launched OASIS 1, a smart ring for private dictation and touch control. the pitch is less fitness tracker and more input device: whisper to write, then tap or swipe on the ring to edit, scroll, pause music, or move around apps.

the useful angle is how narrow it starts. OASIS is selling the awkward middle ground where voice is faster than typing but speaking out loud feels too public. preorders are listed at $289, with shipping around Christmas 2026.

🔗 https://x.com/oasisdevices/status/2072033581241683988

very very sorry for the lack of news!!!!!! i’ve been spending too much time in person in San Francisco, and honestly it’s been jsfhifasjkndhgfq!!!!!!!!

there’s something about being here that’s hard to get online. even at afters, every single person you meet is building something, poking at some weird idea, or asking better questions than the internet usually gives you. that kind of curiosity is contagious. lime rides all around the city, the weather helps too.

17 hours flight to go!!!!!!!!!!!!! in the meantime, come check out this event where we'll try and talk about some of the cool shit that transpired!

speak more soon!
🔗 https://luma.com/olsz24cx

dayflow launched an open source automatic work journal for Mac. it turns screen activity into a readable timeline of what you actually worked on, so you can reconstruct the day without timers, tags, or pretending your calendar was the truth.

the useful bit is the privacy model: recordings and the database stay on your Mac, and you can run analysis with local models or bring Gemini, ChatGPT, or Claude if you want better summaries. it’s MIT licensed, macOS 14+, and the repo is at ~6.5k stars.
🔗 https://x.com/jerryliu/status/2073116662602342734

humanoids are usually framed as an AI problem, but Humanity's Last Machine is a good reminder that the body is the hard part too. it maps the actual hardware stack: skeleton, motors, reducers, screws, bearings, actuators, batteries, compute, sensors, tactile skin, and hands.

the most useful takeaway: costs probably compress more like cars than iPhones. actuators and hands are still the blocker, and the US-China split is less "who has better demos" than "who can iterate hardware, suppliers, and demand loops faster."

🔗 https://www.humanityslastmachine.com/

Anthropic published a short history of Claude Code, and the interesting part is how long the core bet has been around: coding was the path to agents that can actually do work, not just answer questions.

the story traces it from an early VS Code assistant and internal clide / Claude CLI experiments to the february 2025 Claude Code research preview. the takeaway is that the product was less one big reveal than years of scaffolding, tool use, bash access, permissioning, and users teaching Anthropic what the workflow should become.

🔗 https://www.anthropic.com/features/making-of-claude-code

Anthropic published new interpretability research on what it calls Claude's "J-space": a small internal workspace where the model appears to hold concepts it can report, focus on, and use for multi-step reasoning, even when those thoughts never show up in the output.

important caveat: this is not proof that Claude has feelings or human-like consciousness. the practical part is a lens into some hidden internal state, which could help audit when a model notices tests, fabricates data, or pursues a goal it is not saying out loud.

🔗 https://www.anthropic.com/research/global-workspace

grok's voice stack is getting a bigger cast: SpaceXAI added 21 new multilingual flagship voices to the Grok API, alongside upgrades to the original five.

the useful part is that these are not just app voices. they're available through the realtime Voice Agent API, Text to Speech API, and the new Voice Agent Builder, so developers can plug them into support agents, characters, ads, education, and other voice workflows.

🔗 https://x.ai/news/new-flagship-voices

Meta introduced Muse Image and previewed Muse Video, the first media generation models from Meta Superintelligence Labs. the interesting bit is that Muse Image is more agentic than a normal image model: it can use search and code, self-refine, edit in place, and combine multiple reference images in one prompt.

Muse Image is available in Meta AI/meta.ai, Instagram Stories in the US, and WhatsApp in limited countries. Muse Video is still a preview, but Meta says it brings native audio support and is coming to creators and Meta AI.

🔗 https://ai.meta.com/blog/introducing-muse-image-muse-video-msl/
🔗 https://x.com/mattdeitke/status/2074556783583191432

Brainbase launched what it calls an AI agent cloud: infrastructure for deploying lots of agents across different models and tool harnesses, with the boring-but-important pieces handled for you: sandboxing, routing, evals, monitoring, and scaling.

the bigger idea is agents that provision more like compute. instead of one giant model doing everything, teams can spin up smaller purpose-built agents for workflows like PR review, incident triage, or customer requests, then route between models based on cost and performance.

🔗 https://x.com/BrainbaseHQ/status/2074530735911047608

OpenAI says GPT-5.6 Sol, Terra and Luna will launch publicly on Thursday, July 9, with preview access expanding globally now.

the bigger signal is access: this was previously a limited preview, and now the full model family is moving toward public availability instead of staying behind case-by-case approvals.

ARE... YOU... READY???????????????????????????/

🔗 https://x.com/openai/status/2074704958419792299

Cognition launched SWE-1.7, its new coding model for Devin, across web, desktop, and CLI. the practical bit: Cognition says it runs at 1000 tokens/sec via Cerebras, is free for paid users for the next month, and lands close to frontier models on its own FrontierCode eval while costing much less per task.

the interesting part is the training story: SWE-1.7 starts from Kimi K2.7, then Cognition pushes it with RL for longer software tasks, including self-compaction so it can summarize its work and keep going. the caveat is the headline benchmarks are mostly Cognition-run, so treat the numbers as launch data, not neutral third-party validation.

🔗 https://x.com/cognition/status/2074882968770728416
🔗 https://cognition.com/blog/swe-1-7

SpaceXAI launched Grok 4.5, a coding-and-agent model trained alongside Cursor. the practical pitch is less about chatbot vibes and more about engineering workflows: big codebases, long-running tasks, multi-repo work, and tool-heavy agent loops.

the launch claims 80 tokens/sec, roughly 2x token efficiency versus comparable models, and pricing at $2/M input and $6/M output tokens. it’s available in Grok Build, Cursor, and the SpaceXAI console, with EU availability still pending.

🔗 https://x.ai/news/grok-4-5

OpenAI says GPT-Live is now fully rolled out for ChatGPT Go, Plus, and Pro users, with Free still rolling out. update the ChatGPT app on iOS or Android and the new Live voice mode should show up if your account has access.

the bigger change is that this is not just a nicer voice. GPT-Live is full-duplex, so it can listen and talk at the same time, handle pauses and interruptions better, and hand harder questions to GPT-5.5 in the background while keeping the conversation going.

🔗 https://openai.com/index/introducing-gpt-live/
🔗 https://x.com/OpenAI/status/2075019750569378007

Marble open sourced its primary school curriculum graph: 1,590 concepts and 3,221 prerequisite links across 8 subjects, mapped to US/UK standards and shipped as structured JSON on GitHub.

the useful part is that it is not just a list of topics. each concept includes age bands, mastery evidence, and hard/soft prerequisite links, so builders can compute learning paths for AI tutors, education agents, or kid-facing learning apps.

🔗 https://withmarble.com/curriculum
🔗 https://github.com/withmarbleapp/os-taxonomy

Vercel Labs has an open source Native SDK for building desktop apps without shipping a browser. the idea is a simpler native stack: .native markup for views, Zig for logic, hot reload, built-in components/themes, and a renderer that draws into real OS windows instead of using WebView.

the more interesting part is that it is built for agent-era software too: deterministic screenshots, record/replay, headless UI tests, and an automation server so agents can inspect and drive the app. still pre-1.0, with desktop as the mature surface and mobile experimental.

🔗 https://github.com/vercel-labs/native
🔗 https://x.com/ctatedev/status/2075016514860650685

launch videos are becoming a standard part of startup marketing, and andcut is basically a reference library for them.

it collects product, startup, and feature launch videos from X, YouTube, LinkedIn, etc., then makes them searchable by company, launch type, stage, platform, format, industry, and duration. small but useful if you’re trying to figure out how to frame your own launch without starting from a blank page.

🔗 https://www.andcut.video/
🔗 https://x.com/adilatwork/status/2074876337098289243

Meta is opening up Muse Spark 1.1, its new agentic/coding model, through the Meta Model API public preview. the interesting part is not another chat model, but that Meta is packaging long-context, tool-using, computer-use, coding, and multimodal work into one developer-facing API.

Meta says it can manage a 1m-token context, delegate to subagents, use desktops and browsers, and work with images, video, and docs. the eval claims put it near frontier models, but the part to watch is whether the API holds up for real agent builders outside Meta's own tests.

this is surely going to get washed away by 5.6Sol which drops later in about 45 minutes! (1am sgt)

🔗 https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/

official model card/deets are out!:

OpenAI launched GPT-5.6, a new model family with Sol as the flagship, Terra as the lower-cost everyday model, and Luna as the fastest budget option. the interesting part is not just benchmark lift, but the product shape: better coding and knowledge work, programmatic tool calling, and an ultra mode that coordinates multiple agents for harder tasks.

pricing starts at $5 / $30 per 1M tokens for Sol, $2.50 / $15 for Terra, and $1 / $6 for Luna. it is rolling out across ChatGPT, Codex, and the API, with OpenAI also saying GPT-5.6 gets stronger safeguards for cyber/bio work and more trusted-access controls for sensitive use cases.

🔗 https://openai.com/index/gpt-5-6/