Anthropic's most capable model is losing ground to cheaper rivals even as its investors model a $2 trillion October IPO, Goldman Sachs argues open-source AI could be Big Tech's unexpected winner after AT&T cut coding costs 56%, and a wave of GitHub projects from ByteDance's OpenViking to Basecamp's omarchy are reshaping how agents, languages and developer machines are built. Here is the daily tech digest for August 24, 2026.
AI and ML News
Anthropic's flagship model is losing the room to cheaper rivals
Financial Times reports that Anthropic's most capable model is struggling to win users as cheaper alternatives pull ahead. The friction is price: Anthropic's Opus 5 lists at $5 input and $25 per million output tokens, and Sonnet 5 at $2/$10, while OpenAI's GPT-5.6 Luna is free to consumers and open-weight models from DeepSeek and Zhipu undercut on every routine task. On Hacker News the story hit the front page with 253 points and 220 comments, a signal that developers are openly questioning whether premium frontier pricing still pays off for everyday work.
Anthropic's investors are modeling a $2 trillion IPO, bigger than SpaceX
Investors are modeling an October 2026 IPO for Anthropic at a valuation of $2 trillion or more, a figure that would surpass SpaceX's $1.77 trillion listing in June as the largest public debut in history, according to the Financial Times. Backers project annualized revenue of $100 billion to $120 billion by year-end, up from a fraction of that at the start of 2026, and have poured nearly $100 billion into the five-year-old Claude maker during 2026 alone. At least one investor sketched a $3 trillion scenario on 30-times-revenue multiples given roughly 800% annual growth.
Goldman Sachs: open-source AI, not Big Tech, could be the unexpected winner
Goldman Sachs equity research chief Jim Covello is flipping the open-source-AI panic on its head: cheaper models, he argues, make enterprise AI cheap enough to drive far more usage, helping hyperscalers Amazon, Microsoft and Google fill the capacity they are spending billions to build. The thesis has a hard number behind it. AT&T cut the cost of coding and other advanced AI tasks by 56% with only a 2% quality decline by routing routine queries to open-weight models including Meta's Llama, Google's Gemma and Nvidia's Nemotron, and now aims to lift open-model share from 40% to between 60% and 70%.
Researchers decode silent reading straight from the brain
A 45-page preprint on arXiv (cs.LG / q-bio.NC, August 2026) reports decoding silent reading directly from non-invasive EEG, reconstructing the words a person is reading in their head from commodity scalp electrodes. The work sits at the intersection of machine learning and computational neuroscience and suggests that cheap, safe brain-computer interfaces are inching toward reading internal language, not just motor intent, a leap with obvious stakes for accessibility and privacy.
Twin1 AI raises $20M to build an AI digital twin of every worker
Twin1 AI emerged from stealth on August 20 with a $20 million seed round co-led by Bessemer Venture Partners, Tribeca Venture Partners and Aramco Ventures, betting that the next enterprise AI interface is a digital copy of you. The San Mateo and London startup builds AI twins from a worker's authorized email, meetings, documents, Slack and Microsoft Teams, and says customers already include law firms Linklaters and Orrick; its founders previously built AI at Eigen that processed $100 trillion in contracts.
Big GitHub Releases
harry0703/MoneyPrinterTurbo - AI short-video factory hits 115K stars
MoneyPrinterTurbo is having a moment: the Python project that generates HD short videos from a single topic or keyword using an automated AI workflow crossed 115,000 stars and pulled in 2,221 stars in a single day, more than 10,900 over the week. It stitches script generation, voiceover, asset gathering and editing into one pipeline, the kind of end-to-end content automation that used to require a small studio.
google-gemini/gemini-cli - Gemini in your terminal crosses 106K stars
Google's gemini-cli, the open-source agent that drops Gemini straight into your terminal, crossed 106,000 stars and was pushed to GitHub today. It lets developers query, refactor and script against a live model from the command line, and its momentum reflects how quickly coding agents are moving from IDE plugins to the shell itself.
langflow-ai/langflow - visual agent builder at 153K stars
Langflow, the visual canvas for building and deploying AI-powered agents and workflows, sits at 153,000 stars and was pushed to GitHub today. Drag-and-drop components replace boilerplate, and the project has become a default on-ramp for teams shipping RAG and agent pipelines without writing orchestration code by hand.
OpenHands/OpenHands - autonomous coding agent at 84K stars
OpenHands, the autonomous AI-Driven Development agent, reached 84,000 stars and was pushed today. It takes natural-language tasks and executes them across a real code workspace, editing files, running commands and opening pull requests, a direct glimpse at where agentic software engineering is heading.
Under-the-Radar Finds
volcengine/OpenViking - self-evolving context database for agents (32.3K)
Volcengine, the cloud arm of ByteDance, open-sourced OpenViking and it is now at 32,000 stars with 804 gained today. It is a self-evolving context database for AI agents that unifies agent memory, knowledge RAG and skills into one store that rewrites itself as it learns, tagged agent-memory, agentic-rag and self-evolving. For teams wiring memory into production agents, it is one of the most complete primitives to appear this year.
modular/modular - the MAX engine and the Mojo language (28.9K)
Modular's platform, which bundles the MAX engine and the Mojo language, holds 28,900 stars and added 2,017 this week. Mojo is the AI-first programming language from Chris Lattner, the creator of Swift and LLVM, built to deliver C++-level performance with Python ergonomics for tensor and model code, and it is the most serious attempt yet to give AI workloads a native systems language.
basecamp/omarchy - DHH's opinionated Linux (28.7K)
Basecamp's omarchy is an opinionated, beautiful Linux built for developers, and it rocketed to 28,700 stars with 3,151 added in a week. From the camp of David Heinemeier Hansson, it is a statement that the Rails crowd thinks it can ship a better default desktop than the mainstream distributions, with curated defaults instead of endless configuration.
AprilNEA/OpenLogi - privacy-first Rust peripheral tool (14.5K)
OpenLogi is a native, local-first replacement for Logitech Options+ written in Rust, at 14,400 stars and a remarkable 4,993 gained this week. It remaps mouse buttons, DPI and SmartShift over HID++ with no account and no telemetry, a privacy-first utility that is clearly striking a nerve with users tired of cloud-tethered peripheral software.
marceloprates/prettymaps - turn OpenStreetMap data into art (13.9K)
prettymaps turns OpenStreetMap data into publication-ready artistic maps using osmnx, matplotlib and shapely, and sits at 13,900 stars. It is a small, delightful toolkit that reframes raw geodata as design, popular with researchers and cartographers who want a map that looks intentional rather than exported.
Why it matters. The throughline of today is leverage and cost. Anthropic is discovering that frontier capability alone no longer justifies frontier pricing when open-weight models close the gap on routine work, and Goldman's Covello is arguing the real winner is the platform that absorbs all that cheap inference. Meanwhile the GitHub trenches show where the next layer of leverage is being built: unified agent memory (OpenViking), a native AI systems language (Mojo), and developer machines that ship opinionated by default (omarchy). The companies that win the next phase will be the ones that make AI dramatically cheaper to run and dramatically easier to wire into real work, not the ones that simply charge the most for the smartest model.