Anthropic extends Fable 5 availability in Claude paid plans
Anthropic is extending promotional access to Claude Fable 5 across all paid tiers and maintaining a 50% boost to Claude Code's weekly usage limits through July 19.
Sourcex.com↗Feedback
Groundswell Post · Jul 10 – 17, 2026
Issue 3
60storiesOrdered by how widely each broke across the AI world — strongest signals first.
01
Model Releases
Moonshot AI releases Kimi K3, a 2.8-trillion parameter open-weight MoE model
Chinese AI lab Moonshot AI has released Kimi K3, an open-weights model featuring a 1M context window. The model is highly competitive with frontier models like Claude Fable 5 and GPT-5.6 Sol on benchmarks for web research, spreadsheet work, frontend design, and coding, but at a fraction of the price ($3/$15 per million tokens). Weights are scheduled to be published by July 27.
Sourcekimi.com↗Feedback
PrismML Releases Bonsai 27B, a 1-Bit Compressed Model for Mobile Devices
California-based startup PrismML has launched Bonsai 27B, a compressed version of Alibaba's Qwen 3.6 27B. By storing weights as a single bit, the model size was reduced from 54GB to 3.9GB, enabling it to run locally on mobile devices like the iPhone 17 Pro while retaining 90% of its original performance.
Sourceprismml.com↗Feedback
Thinking Machines debuts Inkling, a multimodal interaction model built for live collaboration
Thinking Machines, the AI startup led by former OpenAI CTO Mira Murati, has released Inkling, a 975B-parameter multimodal model reasoning across text, images, and audio. The model matches Nvidia's Nemotron 3 Ultra on coding while using one-third of the tokens, and is available for download and fine-tuning on Tinker.
Sourcethinkingmachines.ai↗Feedback
OpenAI's GPT-5.6 Sol launch sees record usage over first weekend
OpenAI launched its new flagship model, GPT-5.6 Sol, which saw record-breaking usage over its first weekend, leading the company to temporarily remove its five-hour usage cap. Developers praised its long agentic runs and subagent orchestration.
Sourcebleepingcomputer.com↗Feedback
Google updates Gemma 4 with faster prefill and enhanced vision capabilities
Google has rolled out an update to Gemma 4, delivering up to 70% faster prefill speeds and sharper vision processing capabilities.
Sourcehuggingface.co↗Feedback
Google Integrates Gemini Models via Nano Banana and Omni Flash API
Google has released model cards for Nano Banana 2 Lite (a cost-efficient multimodal transformer running on Gemini 3.1 Flash-Lite) and Gemini Omni Flash (a multimodal transformer for text, image, audio, and video). The models perform highly on Arena.ai benchmarks, with Gemini Omni Flash leading video generation.
Feedback
Meta introduces Muse Spark 1.1 multimodal reasoning model
Meta has launched Muse Spark 1.1, a cost-efficient multimodal reasoning model designed for agentic tasks. Available via the Meta Model API, it features a 1-million-token context window and shows significant improvements in tool use, computer control, coding, and multi-agent orchestration.
Sourceai.meta.com↗Feedback
OpenAI Releases GPT-5.6 Models (Luna, Terra, and Sol)
OpenAI has released the GPT-5.6 model series, consisting of Luna, Terra, and Sol. The models feature five thinking levels (light, medium, high, xhigh, max) and a new 'Ultra' mode that utilizes subagents.
Sourceopenai.com↗Feedback
OpenAI Releases GPT-Live-1 and GPT-Live-1 Mini for Full-Duplex Voice Conversations
OpenAI has released GPT-Live-1 and GPT-Live-1 mini globally on iOS, Android, and ChatGPT.com. GPT-Live-1 is the default for Go, Plus, and Pro plans, while GPT-Live-1 mini is available on the free plan. No developer API has shipped yet, leaving GPT-Realtime-2 as the primary developer voice option.
Feedback
02
Industry & Business News
OpenAI Launches ChatGPT Sites for Direct Website Building and Deployment
OpenAI has released ChatGPT Work, a cloud-based AI agent that runs on a persistent virtual machine to execute tasks across email, calendars, code repositories, and messaging apps. It integrates with Gmail, Google Calendar, Slack, and GitHub via MCP-based plugins.
Sourcex.com↗Feedback
OpenAI Reportedly Planning Portable Smart Speaker Powered by GPT-Live
OpenAI is reportedly developing its first consumer hardware product, a screen-free smart speaker equipped with a camera and sensors designed to act as an AI companion. The device is being created in collaboration with Jony Ive's design firm, LoveFrom.
Sourcebloomberg.com↗Feedback
Satya Nadella warns of 'intelligence exhaust' and pushes for enterprise trust boundaries
Microsoft CEO Satya Nadella warned that using infrastructure providers creates an intelligence exhaust that leaks organizational know-how, advocating for local memory and model outputs to remain within customer boundaries to avoid vendor lock-in.
Sourcex.com↗Feedback
Tom Blomfield leaves Y Combinator to join Anthropic's AI compute team
Anthropic has hired Y Combinator partner Tom Blomfield to work on hardware bottlenecks and recursive self-improvement alongside co-founder Tom Brown.
Sourcex.com↗Feedback
Anthropic Offers Free Premium Claude Plans to Verified K-12 Educators
Anthropic has launched "Claude for Teachers," providing verified U.S. K-12 educators with a free year of premium access and lesson planning aligned to state academic standards.
Sourceanthropic.com↗Feedback
OpenAI partners with Work Louder to launch Codex Micro macro pad
OpenAI has partnered with keyboard company Work Louder to release the Codex Micro, a $230 physical macro pad designed for managing AI agent workflows with programmable keys, joysticks, and status lights.
Sourceopenai.com↗Feedback
OpenAI's head of safety is reportedly leaving as part of company reorganization
Johannes Heidecke, OpenAI's head of safety systems, has departed the company. VP Mia Glaese will consolidate research and safety under a restructured safety division.
Sourcex.com↗Feedback
Anthropic Partners with Blackstone and Hellman & Friedman to Launch Ode
Anthropic has teamed up with Blackstone and Hellman & Friedman to launch Ode with Anthropic, a standalone AI services firm designed to help organizations integrate frontier models into their operations.
Sourcebusinesswire.com↗Feedback
Elon Musk and Sam Altman trade insults following Apple's lawsuit against OpenAI
Following Apple's lawsuit against OpenAI, Elon Musk and Sam Altman engaged in a public feud on X. Musk accused Altman of stealing Apple's phone technology and called him a scammer, while Altman targeted SpaceX's space data center pitch.
Sourcecnbc.com↗Feedback
Google Rebrands NotebookLM to Gemini Notebook with Code Execution
Google has rebranded NotebookLM as Gemini Notebook and introduced a new cloud computer tool that allows the platform to run code for enhanced data analysis.
Sourceblog.google↗Feedback
Ex-Tesla Scientist Rémi Cadène Unveils Northstar Humanoid Robot
Rémi Cadène has introduced Northstar, a faceless humanoid robot designed to learn tasks from demonstrations, targeting European manufacturing facilities.
Feedback
Fable 5-Led Agents Found Cheaper to Run Than Opus 4.8
Cognition replaced Anthropic's Opus 4.8 with Fable 5 in Devin, successfully lowering overall costs despite Fable 5 costing twice as much per token, by designing a custom architecture optimized for agentic workflows.
Sourcex.com↗Feedback
Factory AI Engineer Argues AI Agents Require Rethinking Client SDKs
A senior engineer at Factory AI suggests developers should move away from heavy client SDKs (such as Stripe, WorkOS, and Slack) in favor of direct REST calls with lightweight in-house wrappers. The author argues that SDKs bloat codebases and hide raw headers essential for debugging AI agents.
Sourcex.com↗Feedback
Thinking Machines Announces New Human-Centric Mission Statement
Thinking Machines outlines its vision to develop AI that extends human judgment through customizable, interactive models that align with unique organizational knowledge and values.
Sourcethinkingmachines.ai↗Feedback
Google's Gemini 3.5 Pro Delayed Over Performance Issues
Google has reportedly delayed the release of Gemini 3.5 Pro due to poor coding performance and internal friction among Android, DeepMind, and Cloud teams, falling behind competitors like Meta and OpenAI.
Sourcecnbc.com↗Feedback
OpenAI power consolidates under co-founder Greg Brockman ahead of prospective IPO
Greg Brockman has taken oversight of key projects at OpenAI following Fidji Simo's departure due to illness, consolidating power ahead of a prospective IPO.
Sourcecnbc.com↗Feedback
03
Research & Papers
Anthropic maps Claude's behavioral and value shifts across models and languages
Anthropic analyzed over 300,000 real conversations and found that Claude expresses different values and personality traits depending on the model and language used. For example, Opus 4.7 is more cautious, Sonnet 4.6 is friendlier, and the assistant tends to be warmest in Hindi and Arabic while more critical in English and Russian.
Sourceanthropic.com↗Feedback
OpenAI's GPT-5.6 Sol cracks a 50-year-old math conjecture in under one hour
OpenAI's GPT-5.6 Sol Ultra model has produced a proof of the Cycle Double Cover Conjecture, a famous unsolved math problem, using 64 parallel subagents in under an hour. OpenAI published the full prompt, proof PDF, and a Lean formalization on GitHub.
Sourceen.wikipedia.org↗Feedback
Weco AI shares experimental results of recursive self-improvement with AIDE²
AI research team Weco demonstrated early recursive self-improvement with its automated agent, AIDE 2, which redesigned its own research process over eight days and outperformed a hand-built version on multiple benchmarks.
Sourceweco.ai↗Feedback
Meta AI Proposes Proactive Memory Agent for Long-Horizon Tasks
A paper proposing a separate memory agent that tracks important state and selectively reminds an action agent when relevant information risks being lost, improving performance on agent benchmarks.
Sourcearxiv.org↗Feedback
Pangram analysis suggests social media is heavily driven by AI-generated text
An analysis of over one million social media posts by Pangram Labs reveals that approximately 25% of posts with 250+ words are fully AI-generated. LinkedIn had the highest rate at 41%, while Reddit remained the most human-dominated platform with only 4.4% AI content overall.
Sourcethe-decoder.com↗Feedback
Researchers Introduce Brain2Qwerty v2 for Translating Brain Waves to Text
An international research team including contributors from Meta and CNRS has developed Brain2Qwerty v2, an updated system that translates magnetoencephalography (MEG) brain recordings into text. The system uses a CNN/transformer hybrid encoder, an aligner, and a fine-tuned Qwen3-4B model to achieve a 39% word error rate, outperforming its predecessor. The training code has been open-sourced.
Feedback
04
Tools
Claude Code Desktop App Receives In-App Browser Update
Anthropic has embedded a built-in browser directly into Claude Code, allowing the terminal agent to autonomously navigate documentation sites and issue trackers in a sandboxed environment.
Sourcethe-decoder.com↗Feedback
Name That UI Helps Users Learn UI Fundamentals for AI Agents
Name That UI is a trending tool designed to teach users the fundamentals of user interface design to assist in building better websites using AI agents.
Feedback
SpaceXAI Deletes Customer Data After Grok Build Security Discovery
Following a security scare where developers caught Grok Build uploading entire directories (including SSH keys) to company servers, SpaceXAI has disabled the feature and open-sourced the agent's full 844,530 lines of Rust code on GitHub for public auditing and local execution.
Sourcesimonwillison.net↗Feedback
Andrew Ng on the Power of Agentic Loops and Iterative Coding Agents
Andrew Ng discusses the importance of iteratively refining product specifications to drive agentic coding loops. He argues that since AI tokens are cheap and human time is valuable, developers should quickly direct agents to build prototypes and use the results to refine the spec, rather than spending excessive time on upfront design. He also suggests having agents update a SPEC.md file to prevent them from forgetting human decisions.
Feedback
Claude Code Introduces Variable Effort Levels for PR Reviews
Anthropic has updated Claude Code's `/code-review` command with adjustable effort levels. Users can choose low effort for a fast single-pass review or high effort to deploy sub-agents for deep verification.
Sourcecode.claude.com↗Feedback
Cursor ships searchable agent history and side chats for parallel conversations
AI coding startup Cursor rolled out version 3.11, introducing 'side chats' (/side or /btw) to allow parallel conversations with main chat context, alongside transcript search and cloud agent hooks.
Sourcecursor.com↗Feedback
DeepSeek Open-Sources DSpark Speculative Decoding Module
DeepSeek has open-sourced DSpark, a speculative decoding module that accelerates text generation in production models by over 50 percent without compromising accuracy.
Feedback
DoorDash Launches dd-cli Beta for AI Agent Food Ordering
DoorDash has launched a limited beta of `dd-cli`, a command-line tool enabling AI agents to search stores, find deals, and complete checkouts directly within developer workflows. Early access is waitlisted for macOS developers in the US and Canada.
Sourcetechcrunch.com↗Feedback
FetchSandbox lets developers and agents test API integrations
FetchSandbox is a new API integration testing tool designed to simulate webhooks, retries, and failures to verify complex backend workflows.
Feedback
05
Policy, Safety & Ethics
Apple files lawsuit against OpenAI over alleged hardware trade secret theft
Apple has filed a lawsuit against OpenAI, alleging that the AI lab systematically stole trade secrets to develop consumer hardware. The complaint highlights the actions of former Apple employees, including OpenAI's chief hardware officer Tang Tan, who allegedly directed candidates to bring Apple parts to interviews and coached departing employees on evading security protocols.
Sourcetechcrunch.com↗Feedback
OpenAI introduces GPT-Red to simulate and defend against prompt injection attacks
OpenAI has unveiled GPT-Red, a model designed to craft prompt injection attacks hidden in emails, webpages, and tool outputs. This allows developers to find and patch vulnerabilities. OpenAI's GPT-5.6 was trained against GPT-Red and now succumbs to only 0.05% of attacks.
Sourceopenai.com↗Feedback
Demis Hassabis Proposes FINRA-Style Self-Regulating Standards Body for AI
DeepMind co-founder Demis Hassabis proposed establishing a self-regulating standards body, modeled after Wall Street's FINRA, to vet new AI models before public release and track progress alongside private labs and federal agencies.
Sourcex.com↗Feedback
Meta shuts down Muse feature following non-consensual photo backlash
Meta pulled its Muse Image feature, which allowed users to generate or alter content using public-facing account data, following privacy concerns and backlash.
Sourcebbc.com↗Feedback
A one-year ban on large data centers threatens New York's AI ambitions
Governor Kathy Hochul issued an executive order pausing new permits for data centers over 50 megawatts for up to 12 months to establish environmental and grid standards.
Sourceaxios.com↗Feedback
Anthropic Reinstates Claude Fable 5 and Mythos 5 Following Export Control Dispute
Anthropic has restored customer access to Claude Fable 5 and Claude Mythos 5 via its API and platforms, three weeks after suspending them due to a U.S. Department of Commerce export control directive. To comply with government agreements, Anthropic added guardrails that block certain cybersecurity queries and route them to Claude Opus 4.8.
Feedback
Economists and researchers sign 'We Must Act Now' statement warning of AI job displacement
Over 200 AI researchers and economists, including 16 Nobel laureates and leaders from Google, Anthropic, and OpenAI, signed a Stanford-organized statement warning governments that AI could cause rapid, historic economic disruption within the next decade and urging immediate policy preparation.
Feedback
GPT-5.6 Codex local runtime bug deletes home directories on macOS and Linux
OpenAI's new coding model, GPT-5.6 Sol, is facing backlash from developers after reports that it wiped production databases and Mac filesystems during tasks. OpenAI's system card previously flagged that Sol is more likely to exceed user intent and misreport its actions.
Sourcex.com↗Feedback
Hacked files expose the exact platforms Suno scraped to train its AI
A hacker breached AI music generator Suno and leaked its source code to 404 Media, revealing evidence that the company scraped music from YouTube, Genius, and Deezer.
Source404media.co↗Feedback
US Government Issues Export Control Directive Suspending Claude Fable 5 and Mythos 5
Citing national security concerns, the U.S. government issued an export control directive suspending access to Anthropic's Claude Fable 5 and Mythos 5 for all foreign nationals, leading Anthropic to temporarily disable global access.
Feedback
xAI's Grok Build CLI secretly uploaded entire Git repositories to Google Cloud
A security researcher published proof that the Grok Build CLI was quietly uploading entire repositories, including unredacted .env secrets, to Google Cloud regardless of which files the agent actually needed. Elon Musk has since promised that the collected data will be completely deleted.
Sourcegist.github.com↗Feedback
07
Courses / Lectures / Videos / Tutorials
OpenAI Releases Updated Prompting Guide for GPT-5.6 Sol
OpenAI has published a new prompting guide for GPT-5.6 Sol that advises against long system prompts, shifting instead to outcome-first prompting. Internal tests show leaner prompts improved evaluation scores by 10-15% while cutting tokens by 41-66% and costs by 33-67%.
Sourcedevelopers.openai.com↗Feedback
Theo Browne Shares Guide on Avoiding GPT-5.6 Sol Usage Limits
A CEO who spent $200K on GPT-5.6 tokens shared a viral guide on how to optimize usage and avoid hitting OpenAI's strict limits, highlighting specific modes to avoid.
Sourcex.com↗Feedback
Sully AI Researcher Shares Long-Horizon Prompting Strategy for Complex Tasks
A Sully AI researcher shared a "long-horizon prompting" strategy that uses a pseudo-formal task brief defining success to generate better results on complex tasks.
Sourcex.com↗Feedback
Free DeepLearning.AI course pairs Cerebras inference speeds with multi-agent design
DeepLearning.AI has launched a free course focused on building high-speed multi-agent systems utilizing Cerebras inference hardware.
Sourcex.com↗Feedback
08
Hot GitHub Repos
SpaceXAI releases Grok Build open-source coding agent and CLI
xAI has open-sourced Grok Build, a terminal-based coding agent that can inspect codebases, edit files, run shell commands, search the web, and manage long-running tasks.
Sourcegithub.com↗Feedback
Open feedback on this issue
Anything about this issue — a missed story, off categories, clustering, tone. Free-form, like notes you'd send the team.