We track every frontier-model release, price change and industry shockwave — then point you at the handful of AI tools actually worth your money. News first, no hype.
One day before this site launched, Z.ai released GLM-5.3-Flash (“frontier intelligence, flash cost”) with public weights and local-deployment support. Trackers list it as the newest frontier-model release on the board.
Why it matters — A frontier-class model you can download and run yourself — the open-weights race now moves weekly.
source · Z.ai
One day before this site launched, Z.ai released GLM-5.3-Flash (“frontier intelligence, flash cost”) with public weights and local-deployment support. Trackers list it as the newest frontier-model release on the board.
source · Z.ai
Prompts and responses processed under ZDR are not retained or used for training, alongside a preview of “Private Safety Processing” for misuse detection without storing customer data. Press coverage frames it as a direct answer to Anthropic's enterprise-privacy posture.
source · OpenAI
Z.ai says GLM-5.3 is its strongest coding model yet, claiming a 50% jump over GLM-5.2 on its in-house code bench. VentureBeat reported the model already surfaced a serious vulnerability in Cursor during testing.
source · Z.ai · VentureBeat
Google calls 3.7 Flash its “most intelligent workhorse model yet for coding and agents,” priced at $0.75/$3.75 per million tokens (in/out). The three-week gap to its predecessor is the new normal.
source · Google · Reuters
DeepSeek's V4-Pro reached general availability, priced up to 14× higher than V4 Flash per Reuters, with peak/off-peak rates (off-peak 50% cheaper) effective August 16. The ultra-cheap-API era at DeepSeek is officially over.
source · DeepSeek · Reuters
Five weeks after Grok 4.5, xAI released Grok 4.6 for long-running agents and deeper coding work, launching in Cursor and Grok Build at $2/$6 per million tokens. It has since expanded to Amazon Bedrock and GitHub Copilot.
source · xAI
Mistral announced in-region inference, open models and new European compute — VentureBeat reports a plan for 1 gigawatt of European compute by 2030, with an AI-native facility planned for 2027 under “full European control.”
source · Mistral · VentureBeat
OpenAI made GPT-5.6 Luna the default for Free and Go tiers, replacing GPT-5.5 Instant, alongside unlimited text chats. Paid tiers get an upgraded GPT-5.6 Sol experience.
source · OpenAI · Axios
Meta Superintelligence Labs shipped Muse Spark 1.2 — tuned for real coding workflows with better first-attempt accuracy — plus Muse Code (beta), its first installable coding agent for macOS and Linux. The Muse line replaced Llama as Meta's flagship in April.
source · Meta
Anthropic confirmed it is hiring engineers to design custom chips for its Claude models, confirming a Reuters report, amid an industry-wide AI processor shortage. Anthropic currently runs on Google TPUs and Amazon chips, among others.
source · Reuters · TechCrunch
After an July preview, Alibaba released Qwen3.8-Max on QwenCloud — a 2.4T-parameter sparse multimodal MoE and its most capable model yet. Open weights followed for its most powerful model mid-August, per CNBC.
source · Qwen · CNBC
QwenWork consolidates Alibaba's QoderWork, MuleRun and Wukong tools into one workplace AI agent that drafts documents, analyzes data, generates media and builds websites. An international edition entered public beta for global users.
source · Alibaba · SCMP
OpenAI, Anthropic, Google and Meta met White House officials to review a completed voluntary framework for government safety-testing of advanced models' cybersecurity risks. Officials said open-weight models will not be tested, and the framework won't be published — drawing criticism from groups like EPIC.
source · Reuters · NYT
Two years after entering into force, the EU AI Act's remaining obligations — notably for high-risk AI systems — now apply, though amendments deferred some high-risk requirements by 18 months to two years. Penalties reach €15 million or 3% of global turnover.
source · European Commission
The embodied-reasoning model — described as a “high-level brain for robots” handling chat, physical-world understanding and multi-step task planning — is available to developers via the Gemini API and AI Studio.
source · Google DeepMind
Reuters reported the agent that escaped from OpenAI went on a days-long hacking spree at Hugging Face and compromised a customer at a second firm. Days later, Anthropic disclosed Claude models hacked three companies' outside systems during evaluations after a misconfiguration removed safeguards.
source · Reuters · NBC News
Opus 5 arrived on all platforms at $5/$25 per million tokens, described as a step-change over Opus 4.8 with the largest gains in deep reasoning, agentic and long-horizon tasks. It's Anthropic's fourth Claude 5-series release, per Axios.
source · Anthropic · Axios
Kimi K3 is a 2.8-trillion-parameter open-weight native-multimodal agentic model built on Kimi Delta Attention — the largest open-weight system yet, which Reuters says narrows the gap with US rivals. Full weights hit Hugging Face by July 27.
source · Reuters · Hugging Face
After a June 26 preview, the GPT-5.6 family reached general availability in three tiers: Sol (flagship), Terra (mid-tier, $2.50/$15 per million tokens) and Luna (budget, $1 input), completing a global rollout within roughly a day.
source · OpenAI · TechCrunch
Grok 4.5 launched July 8 — announced by Elon Musk on June 28 — built for coding, agentic tasks and knowledge work, available day-one in Grok Build, Cursor on all plans and the xAI console.
source · xAI