{"id": "claude-opus-5", "added": "2026-08-11", "name": "Claude Opus 5", "org": "Anthropic", "what": "Frontier model at unchanged pricing while more than doubling its predecessor's benchmark performance.", "category": "model", "tags": ["frontier", "commercial"], "url": "https://www.anthropic.com/news/claude-opus-5", "url_status": "verified", "note": "Launched July 24th at $5/$25 per million tokens. Capability moved and the number did not, which puts pressure on everyone selling a mid-tier model.", "sources": ["https://www.anthropic.com/news/claude-opus-5", "https://www.axios.com/2026/07/24/anthropic-releases-new-model-opus-5"], "scale": "product", "reason": "consequential"}
{"id": "claude-sonnet-5", "added": "2026-08-11", "name": "Claude Sonnet 5", "org": "Anthropic", "what": "The volume tier of the Claude line, with a price step-up scheduled for September 1st.", "category": "model", "tags": ["frontier", "commercial"], "url": "https://www.anthropic.com/news/claude-sonnet-5", "url_status": "verified", "note": "Worth watching for whether the September price rise causes visible usage decline. That is the cleanest read available on how elastic demand for mid-tier inference actually is.", "sources": ["https://www.anthropic.com/news/claude-sonnet-5"], "scale": "product", "reason": "consequential"}
{"id": "claude-code-auto-mode", "added": "2026-08-11", "name": "Claude Code auto mode", "org": "Anthropic", "what": "Coding agent that stops asking permission for each action, with a classifier blocking the irreversible ones.", "category": "devtool", "tags": ["agents", "safety"], "url": "https://www.anthropic.com/engineering/claude-code-auto-mode", "url_status": "verified", "note": "Default from August 14th on paid plans. The engineering post reports the classifier caught 89% of planted attacks against 13.6% for human reviewers, and separately missed 17% of real cases where the model exceeded its authorization. The second number is the one about production.", "sources": ["https://www.anthropic.com/engineering/claude-code-auto-mode", "https://techcrunch.com/2026/08/09/anthropic-is-turning-claude-codes-auto-mode-on-by-default/"], "scale": "product", "reason": "consequential"}
{"id": "muse-glimmer", "added": "2026-08-11", "name": "Muse Glimmer", "org": "Meta", "what": "30B agentic model under Apache 2.0, multimodal, small enough for a consumer GPU.", "category": "model", "tags": ["open-weights", "agents"], "url": "https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model", "url_status": "verified", "note": "Distilled from Muse Spark, weights on Hugging Face, with Meta saying an open-weight version of the flagship will follow. A reversal after a year of Meta Superintelligence Labs drifting closed.", "sources": ["https://research.meta.ai/blog/introducing-muse-glimmer-open-agentic-model", "https://www.cnbc.com/2026/08/10/meta-muse-glimmer-open-weight-ai.html"], "scale": "platform", "reason": "consequential"}
{"id": "genesis-open-models", "added": "2026-08-11", "name": "Genesis Open Models Initiative", "org": "US Department of Energy", "what": "Government-run library of open-weight scientific models, starting with Genesis-Science-1.", "category": "model", "tags": ["open-weights", "science", "policy"], "url": "https://www.energy.gov/undersecretaryforscience/articles/us-department-energy-launches-genesis-open-models-initiative", "url_status": "verified", "note": "Built with Arcee AI. The US government is now an open-weight model producer. That is a structural answer to China's majority token share, and it is not a research project.", "sources": ["https://www.energy.gov/undersecretaryforscience/articles/us-department-energy-launches-genesis-open-models-initiative"], "scale": "platform", "reason": "new"}
{"id": "openai-astra", "added": "2026-08-11", "name": "Astra", "org": "OpenAI", "what": "Unreleased model family that resolved ten decade-old mathematics problems with machine-checkable proofs.", "category": "model", "tags": ["frontier", "science"], "url": "https://openai.com/index/ten-advances-in-mathematics/", "url_status": "verified", "note": "Nobody outside OpenAI can run it, so it is here to know about and not to use. Compute cost roughly $2,000 per result, with Lean 4 certificates published so correctness does not rest on trust.", "sources": ["https://openai.com/index/ten-advances-in-mathematics/", "https://openai.com/index/model-disproves-discrete-geometry-conjecture/"], "scale": "platform", "reason": "consequential"}
{"id": "kimi-k3", "added": "2026-08-11", "name": "Kimi K3", "org": "Moonshot AI", "what": "2.8 trillion parameters, the largest open-weight model released to date.", "category": "model", "tags": ["open-weights", "china"], "url": "", "url_status": "unconfirmed", "note": "Part of the batch that took Chinese open weights from negligible to a majority share of all tokens processed. Western coverage of these releases is thin, which says more about the coverage than about the models.", "sources": ["https://www.datagravity.dev/p/chinas-open-weight-takeover"], "scale": "platform", "reason": "consequential"}
{"id": "deepseek-v4-flash", "added": "2026-08-11", "name": "DeepSeek V4 Flash", "org": "DeepSeek", "what": "Exited preview at $0.14 per million input tokens, $0.28 output.", "category": "model", "tags": ["open-weights", "china", "pricing"], "url": "", "url_status": "unconfirmed", "note": "The price floor for the industry. Anyone whose business rests on charging for a mid-tier model is competing with this number.", "sources": ["https://www.datagravity.dev/p/chinas-open-weight-takeover"], "scale": "product", "reason": "consequential"}
{"id": "qwen-3-8-max", "added": "2026-08-11", "name": "Qwen3.8-Max", "org": "Alibaba", "what": "2.4 trillion parameter open-weight model.", "category": "model", "tags": ["open-weights", "china"], "url": "", "url_status": "unconfirmed", "note": "Released alongside Kimi K3 and DeepSeek V4 in the same window. Three labs, three of the largest open-weight models in existence, one month.", "sources": ["https://www.datagravity.dev/p/chinas-open-weight-takeover"], "scale": "platform", "reason": "consequential"}
{"id": "perplexity-comet", "added": "2026-08-11", "name": "Comet", "org": "Perplexity", "what": "Shopping agent that browses and buys on a user's behalf.", "category": "agent", "tags": ["agents", "commerce", "legal"], "url": "", "url_status": "unconfirmed", "note": "The subject of the first federal appellate ruling on agent access. The Ninth Circuit held that when a user directs an agent, it is the user who accesses the site under the CFAA. Agentic commerce had been waiting on that.", "sources": ["https://www.courthousenews.com/ninth-circuit-lifts-block-on-ai-powered-shopping-assistant/", "https://www.cooley.com/news/insight/2026/2026-08-06-ninth-circuit-rules-on-ai-agent-access-to-third-party-websites-under-cfaa"], "scale": "product", "reason": "new"}
{"id": "awesome-ai-agent-attacks", "added": "2026-08-11", "name": "awesome-ai-agent-attacks", "org": "Community", "what": "Open timeline of real, documented AI agent security incidents.", "category": "repo", "tags": ["open-source", "security", "agents"], "url": "https://github.com/webpro255/awesome-ai-agent-attacks", "url_status": "verified", "note": "Useful because it collects incidents that actually happened, not proof-of-concept attacks. If you are deploying agents, read this before writing the threat model.", "sources": ["https://github.com/webpro255/awesome-ai-agent-attacks"], "scale": "project", "reason": "new"}
{"id": "lean-4", "added": "2026-08-11", "name": "Lean 4", "org": "Lean FRO", "what": "Proof assistant that machine-checks mathematics, now load-bearing for AI-generated results.", "category": "devtool", "tags": ["open-source", "verification", "science"], "url": "", "url_status": "unconfirmed", "note": "The reason the Astra results are credible at all. Verification moved from an academic niche to the bottleneck technology for AI-generated knowledge, and every domain that wants machine-checked output now needs its own version.", "sources": ["https://openai.com/index/ten-advances-in-mathematics/"], "scale": "tool", "reason": "new"}
{"id": "anthropic-claude-timeline", "added": "2026-08-11", "name": "anthropic-claude-timeline", "org": "jqueryscript", "what": "Community-maintained timeline of every Claude model release and change.", "category": "repo", "scale": "project", "tags": ["open-source", "reference"], "url": "https://github.com/jqueryscript/anthropic-claude-timeline", "url_status": "verified", "note": "The kind of small repo that is more useful than the vendor's own changelog, because it is dated, complete and diffable. Someone maintains this by hand and it stays current.", "sources": ["https://github.com/jqueryscript/anthropic-claude-timeline"], "reason": "new"}
{"id": "wan-animate-2", "added": "2026-08-11", "name": "Wan-Animate-2", "org": "Alibaba Tongyi Lab", "what": "Open-weight character animation model driven by raw video, with a Lite variant streaming 24 fps at 400x720.", "category": "model", "scale": "product", "tags": ["open-weights", "video", "china"], "url": "https://humanaigc.github.io/wan-animate-2/", "url_status": "verified", "note": "Apache 2.0, weights, inference code and paper on GitHub and Hugging Face since August 7th, with INT8/BF16 quantizations and ComfyUI nodes. Blind user studies put it at parity with ByteDance's Dreamina and Kling MotionControl, so the closed video-animation products just lost their moat to a free download.", "sources": ["https://humanaigc.github.io/wan-animate-2/", "https://pandaily.com/tongyi-wan-animate-2-character-animation-open-source-aug2026"], "reason": "new"}
{"id": "gpt-5-6-cyber", "added": "2026-08-11", "name": "GPT-5.6-Cyber", "org": "OpenAI", "what": "Cybersecurity model trained for vulnerability research and exploit validation, gated behind the Daybreak Red partner tier.", "category": "model", "scale": "product", "tags": ["security", "frontier", "restricted"], "url": "https://x.com/OpenAI/status/2086864365379010729", "url_status": "verified", "note": "First OpenAI model rated High cyber capability under its Preparedness Framework: 95.0% completion on advanced exploit-chain tasks against 57.3% for GPT-5.5-Cyber, and two chainable Chrome V8 discoveries. You cannot sign up for it. It ships to 16 vetted partners, and the access model is as notable as the model.", "sources": ["https://x.com/OpenAI/status/2086864365379010729", "https://techcrunch.com/2026/08/10/as-ai-led-attacks-multiply-openai-launches-a-new-cyber-model/"], "reason": "consequential"}
{"id": "bumblebee-scanner", "added": "2026-08-11", "name": "Bumblebee", "org": "Perplexity", "what": "Read-only supply-chain scanner that inventories packages, MCP configs, editor and browser extensions on developer machines.", "category": "security", "scale": "tool", "tags": ["security", "open-source", "mcp"], "url": "https://github.com/perplexityai/bumblebee", "url_status": "verified", "note": "Apache 2.0, Go with zero non-stdlib dependencies, and it never executes install scripts or package managers, so a scan cannot itself be an attack. Covers npm, PyPI, Go modules, RubyGems, Composer, MCP servers and extensions in one pass. The month tl;dv showed what unaudited tooling costs, an endpoint inventory tool this boring is the right kind of boring.", "sources": ["https://github.com/perplexityai/bumblebee", "https://www.marktechpost.com/2026/05/23/perplexity-open-sources-bumblebee-a-read-only-supply-chain-scanner-for-developer-endpoints/"], "reason": "new"}
{"id": "codebase-memory-mcp", "added": "2026-08-11", "name": "codebase-memory-mcp", "org": "DeusData", "what": "MCP server that indexes a codebase into a persistent knowledge graph so coding agents answer structural questions without re-reading files.", "category": "mcp", "scale": "tool", "tags": ["agents", "mcp", "open-source"], "url": "https://github.com/DeusData/codebase-memory-mcp", "url_status": "verified", "note": "Single Go binary, embedded SQLite, 158 languages via tree-sitter, 11 MCP tools. The authors' benchmark across 31 repositories claims 10x fewer tokens and 2.1x fewer tool calls than file-by-file exploration; their numbers, so discount accordingly. Attacks the same session-amnesia problem Dwarkesh's continual-learning essay names, from the cheap end.", "sources": ["https://github.com/DeusData/codebase-memory-mcp", "https://www.russ.cloud/2026/05/10/codebase-memory-mcp-giving-claude-code-and-codex-a-map/"], "reason": "new"}
{"id": "juggler-gui-agent", "added": "2026-08-11", "name": "Juggler", "org": "", "what": "Open-source GUI coding agent from the creator of the JUCE audio framework.", "category": "agent", "scale": "project", "tags": ["agents", "open-source", "coding"], "url": "https://news.ycombinator.com/item?id=48883305", "url_status": "verified", "note": "Surfaced via Show HN. A desktop GUI for running coding agents, written by someone who has maintained a widely used C++ framework for two decades. The agent-tooling wave has mostly lacked authors with that background. Filed here because it is hard to find, and with no claim that it wins its category.", "sources": ["https://news.ycombinator.com/item?id=48883305"], "reason": "new"}