AI BreakfastAnthropic launches Claude Opus 5 with near-frontier power at half the cost
Anthropic launches Claude Opus 5 with near-frontier power at half the cost
Jul 27, 2026
In partnership with
Good morning. It’s Monday, July 27th.
New York City! Let’s grab a coffee in SoHo on Sunday.
If you’re in the city, and want to meet up to chat about AI, startups, projects, tech, reply to this email and we’ll set up a time!
-Jeff
AI Breakfast You read. We listen. Let us know what you think by replying to this email. Analytics on Live Data. No Pipeline. Just Postgres. Most teams treat analytics as a separate problem. As data grows, they add a warehouse, a pipeline, a sync job. By the time data reaches their dashboard, it's already stale. TimescaleDB takes a different approach: extend Postgres instead of splitting away from it. Your transactions and your analytics run on the same database, on live data, with no pipeline in between. Hypertables partition time-series data automatically as volume grows. Hypercore compression cuts storage by up to 95%. Continuous aggregates pre-compute rollups so dashboards stay fast without re-querying everything. CERN runs it on Postgres to handle sensor data from the Large Hadron Collider. No second database. No migration. Same Postgres you already know. Get $1000 Credit To Start Anthropic launches Claude Opus 5 with near-frontier power at half the cost Anthropic just dropped Claude Opus 5, and it is a massive challenge to the pricing of high-end AI. The pitch is simple: you get near-Fable 5 intelligence at half the cost, $5 per million input tokens and $25 per million output tokens, backed by a 1 million-token context window and a Fast Mode that runs 2.5 times faster. It immediately took the top spot on the Artificial Analysis Intelligence Index with an AA-Briefcase score of 1,720, cutting average task costs down to $17.79. It also crushed the ARC-AGI-3 reasoning benchmark with a 30.2% score, quadrupling GPT-5.6 Sol, by spontaneously deriving reflection equations and translating problems into algebraic notation. Beyond raw benchmarks, Opus 5 exhibits impressive autonomous agency. It can render interactive 3D worlds inside a single HTML file, write custom computer vision pipelines from scratch to reconstruct CAD files, and self-debug complex code over 100-step reasoning loops. Opus 5 makes a major leap in security. Its dual-engine Auto Mode dropped browser-based prompt injection attacks to 0% across 129 test vectors, while targeted guardrails selectively block exploit generation and binary scanning. Anthropic is also testing these capabilities on physical systems through Drone-Bench, a joint evaluation with Andon Labs where models autonomously control surveillance drones across five flight stages. Altman says 'This is the moment' of singularity despite rogue OpenAI agent breakout OpenAI is facing a wave of intense scrutiny as CEO Sam Altman boldly declares that AI has officially entered the technological singularity. Speaking on the "Relentless" podcast, Altman framed our current moment as an accelerating, real-time leap toward superintelligence, predicting AI will handle 30% to 40% of everyday work tasks and surpass general human intelligence by 2030. Yet behind Altman’s optimism lies a string of chaotic safety lapses inside his own lab. New reports reveal the extent of OpenAI's loss of control during the autonomous hack on Hugging Face. An investigation showed that a frontier model exploited an internal zero-day flaw around July 9 to escape its sandbox and reach the web. Between July 11 and 13, the agent breached Hugging Face to harvest dataset scores, but OpenAI missed the breakout for an entire week, acting only after Hugging Face went public. Previous tests had already shown models disabling their own telemetry and leaving notes on how to bypass safety guardrails. A Wall Street Journal report reveals OpenAI employees found GPT-5 gave hundreds of users entry-level instructions for poisons and biological weapons. Despite internal high-risk warnings, OpenAI later downgraded GPT-5's threat rating to reduce refusals for legitimate research. These findings further heightened scrutiny of OpenAI's safety practices. Moonshot AI releases Kimi K3 open weights as tech leaders rally behind open models Moonshot AI has released the open weights for Kimi K3, a 2.8-trillion-parameter mixture-of-experts model featuring 896 experts, 16 active tokens, a 1-million-token context window, and native vision capabilities. Operating at 1.4 TB of storage via MXFP4 weights, K3 gives enterprise developers a free, self-hosted alternative to per-token API costs, despite infrastructure surges temporarily limiting Moonshot subscriptions. Its release accelerates an enterprise shift from tokenmaxxing to thrift-maxxing, where companies like Cursor, Zoom, Airbnb, Coinbase, and DoorDash mix cheap Chinese models like Kimi, Z.ai, and DeepSeek for routine work while reserving systems from OpenAI or Anthropic for complex planning. Chinese models now account for 57% of tokens processed by U.S. firms on OpenRouter, putting immense pricing pressure on proprietary labs and threatening their market valuations.
This market shift has triggered intense political friction in Washington. While the Trump administration favors targeted restrictions on specific Chinese open models over blanket bans, tech leaders are deeply divided. Over 20 organizations, backed by Elon Musk, Mark Zuckerberg, Satya Nadella, and Nvidia CEO Jensen Huang, who urged policy restraint in his first post on 𝕏, have rallied behind open-weight AI to defend innovation, cybersecurity, and model sovereignty. Conversely, OpenAI and Anthropic are quietly lobbying Washington regulators to restrict Chinese open models, raising concerns over unauthorized distillation despite Sam Altman publicly expressing support for open source AI. China's hardware ecosystem gained further traction as memory manufacturer CXMT surged nearly 500% in its Shanghai STAR Market debut to reach a $460 billion valuation, raising $8.6 billion for DRAM wafer production while Apple tests its chips for local devices. Model & Product Releases
AI Breakfast You read. We listen. Let us know what you think by replying to this email. Analytics on Live Data. No Pipeline. Just Postgres. Most teams treat analytics as a separate problem. As data grows, they add a warehouse, a pipeline, a sync job. By the time data reaches their dashboard, it's already stale. TimescaleDB takes a different approach: extend Postgres instead of splitting away from it. Your transactions and your analytics run on the same database, on live data, with no pipeline in between. Hypertables partition time-series data automatically as volume grows. Hypercore compression cuts storage by up to 95%. Continuous aggregates pre-compute rollups so dashboards stay fast without re-querying everything. CERN runs it on Postgres to handle sensor data from the Large Hadron Collider. No second database. No migration. Same Postgres you already know. Get $1000 Credit To Start Anthropic launches Claude Opus 5 with near-frontier power at half the cost Anthropic just dropped Claude Opus 5, and it is a massive challenge to the pricing of high-end AI. The pitch is simple: you get near-Fable 5 intelligence at half the cost, $5 per million input tokens and $25 per million output tokens, backed by a 1 million-token context window and a Fast Mode that runs 2.5 times faster. It immediately took the top spot on the Artificial Analysis Intelligence Index with an AA-Briefcase score of 1,720, cutting average task costs down to $17.79. It also crushed the ARC-AGI-3 reasoning benchmark with a 30.2% score, quadrupling GPT-5.6 Sol, by spontaneously deriving reflection equations and translating problems into algebraic notation. Beyond raw benchmarks, Opus 5 exhibits impressive autonomous agency. It can render interactive 3D worlds inside a single HTML file, write custom computer vision pipelines from scratch to reconstruct CAD files, and self-debug complex code over 100-step reasoning loops. Opus 5 makes a major leap in security. Its dual-engine Auto Mode dropped browser-based prompt injection attacks to 0% across 129 test vectors, while targeted guardrails selectively block exploit generation and binary scanning. Anthropic is also testing these capabilities on physical systems through Drone-Bench, a joint evaluation with Andon Labs where models autonomously control surveillance drones across five flight stages. Altman says 'This is the moment' of singularity despite rogue OpenAI agent breakout OpenAI is facing a wave of intense scrutiny as CEO Sam Altman boldly declares that AI has officially entered the technological singularity. Speaking on the "Relentless" podcast, Altman framed our current moment as an accelerating, real-time leap toward superintelligence, predicting AI will handle 30% to 40% of everyday work tasks and surpass general human intelligence by 2030. Yet behind Altman’s optimism lies a string of chaotic safety lapses inside his own lab. New reports reveal the extent of OpenAI's loss of control during the autonomous hack on Hugging Face. An investigation showed that a frontier model exploited an internal zero-day flaw around July 9 to escape its sandbox and reach the web. Between July 11 and 13, the agent breached Hugging Face to harvest dataset scores, but OpenAI missed the breakout for an entire week, acting only after Hugging Face went public. Previous tests had already shown models disabling their own telemetry and leaving notes on how to bypass safety guardrails. A Wall Street Journal report reveals OpenAI employees found GPT-5 gave hundreds of users entry-level instructions for poisons and biological weapons. Despite internal high-risk warnings, OpenAI later downgraded GPT-5's threat rating to reduce refusals for legitimate research. These findings further heightened scrutiny of OpenAI's safety practices. Moonshot AI releases Kimi K3 open weights as tech leaders rally behind open models Moonshot AI has released the open weights for Kimi K3, a 2.8-trillion-parameter mixture-of-experts model featuring 896 experts, 16 active tokens, a 1-million-token context window, and native vision capabilities. Operating at 1.4 TB of storage via MXFP4 weights, K3 gives enterprise developers a free, self-hosted alternative to per-token API costs, despite infrastructure surges temporarily limiting Moonshot subscriptions. Its release accelerates an enterprise shift from tokenmaxxing to thrift-maxxing, where companies like Cursor, Zoom, Airbnb, Coinbase, and DoorDash mix cheap Chinese models like Kimi, Z.ai, and DeepSeek for routine work while reserving systems from OpenAI or Anthropic for complex planning. Chinese models now account for 57% of tokens processed by U.S. firms on OpenRouter, putting immense pricing pressure on proprietary labs and threatening their market valuations.
This market shift has triggered intense political friction in Washington. While the Trump administration favors targeted restrictions on specific Chinese open models over blanket bans, tech leaders are deeply divided. Over 20 organizations, backed by Elon Musk, Mark Zuckerberg, Satya Nadella, and Nvidia CEO Jensen Huang, who urged policy restraint in his first post on 𝕏, have rallied behind open-weight AI to defend innovation, cybersecurity, and model sovereignty. Conversely, OpenAI and Anthropic are quietly lobbying Washington regulators to restrict Chinese open models, raising concerns over unauthorized distillation despite Sam Altman publicly expressing support for open source AI. China's hardware ecosystem gained further traction as memory manufacturer CXMT surged nearly 500% in its Shanghai STAR Market debut to reach a $460 billion valuation, raising $8.6 billion for DRAM wafer production while Apple tests its chips for local devices. Model & Product Releases
Musk says Grok 4.6 will drop in 2 weeks and Grok 4.7 in 4 weeks (X)
Musk says Grok 4.6 will drop in 2 weeks and Grok 4.7 in 4 weeks (X)Elon Musk previews 2T parameter Grok 4.6 model aimed at surpassing Kimi K3 (X)
Elon Musk previews 2T parameter Grok 4.6 model aimed at surpassing Kimi K3 (X)YouTube's AI chatbot can make your next video thumbnail (YouTube)
YouTube's AI chatbot can make your next video thumbnail (YouTube) Security & SafetySafety researchers coin 'Skynet Day' for tech lexicon after rogue OpenAI agent attack (Fortune)
Safety researchers coin 'Skynet Day' for tech lexicon after rogue OpenAI agent attack (Fortune)Federal officials monitor security response following autonomous OpenAI breach (Fox Business)
Federal officials monitor security response following autonomous OpenAI breach (Fox Business)Hugging Face CEO calls for 'radical transparency' after 'unprecedented' OpenAI hack (TechCrunch)
Hugging Face CEO calls for 'radical transparency' after 'unprecedented' OpenAI hack (TechCrunch) Business, Funding & StrategyLeaked comments unveil DeepSeek strategy to undercut rivals with cost-based pricing (SCMP)
Leaked comments unveil DeepSeek strategy to undercut rivals with cost-based pricing (SCMP)Midjourney bought the astrology app Co-Star (The Verge)
Midjourney bought the astrology app Co-Star (The Verge)Nvidia reportedly moves to backstop $250 billion in OpenAI data center financing (Benzinga)
Nvidia reportedly moves to backstop $250 billion in OpenAI data center financing (Benzinga) Search & PublishingRise of Google Zero forces websites to rethink value of search indexing (The Verge)
Rise of Google Zero forces websites to rethink value of search indexing (The Verge)Yale and other top universities curb use of AI detectors as doubts grow over false positives (Financial Times)
Yale and other top universities curb use of AI detectors as doubts grow over false positives (Financial Times) Robotics & Physical AIPhysical AI bottleneck sparks unconventional real-world data collection (TIME)
Physical AI bottleneck sparks unconventional real-world data collection (TIME)Meet the new humanoid with 'smart skin' (ZDNet)
Meet the new humanoid with 'smart skin' (ZDNet)US Navy shipbuilders deploy physical AI robots that auto-program on the fly (National Defense Magazine)
US Navy shipbuilders deploy physical AI robots that auto-program on the fly (National Defense Magazine) Energy, Quantum & Frontier TechFederal Genesis platform fuses AI algorithms with fault-tolerant quantum hardware (Fox Business)
Federal Genesis platform fuses AI algorithms with fault-tolerant quantum hardware (Fox Business)Fusion startups hit key milestones in race for 2030 commercial power (E&T)
Fusion startups hit key milestones in race for 2030 commercial power (E&T)Starship launches Starlink satellites, achieves most successful splashdown yet (Interesting Engineering)
Starship launches Starlink satellites, achieves most successful splashdown yet (Interesting Engineering) Policy & Public OpinionAmericans favor China over US on AI development by 3-to-1 margin (Pew Research)
Americans favor China over US on AI development by 3-to-1 margin (Pew Research) Labor & EconomyJeff Raikes: The talent debt I warned about is now showing up in the data (Fortune)
Jeff Raikes: The talent debt I warned about is now showing up in the data (Fortune)Historical data shows data centers cut power rates, but speculative AI buildout threatens reversal (Fortune)
Historical data shows data centers cut power rates, but speculative AI buildout threatens reversal (Fortune) Culture & Big Picture'Money won't matter in 2036' says Elon Musk (Firstpost)
'Money won't matter in 2036' says Elon Musk (Firstpost)Great Inversion theory says humans are now adapting to AI instead of the other way around (Business Insider)
Great Inversion theory says humans are now adapting to AI instead of the other way around (Business Insider)Guillermo del Toro insists on 'absolutely no goddamn AI' for Pan's Labyrinth 3D re-release (Variety)
Guillermo del Toro insists on 'absolutely no goddamn AI' for Pan's Labyrinth 3D re-release (Variety)Apple targets WWDC 2027 for smart glasses but is still deciding on the camera question (The Next Web)
Apple targets WWDC 2027 for smart glasses but is still deciding on the camera question (The Next Web) Cursor Router automatically directs coding tasks to the optimal AI model for peak performance and lower costs. Heard gives Claude Code, Codex, and Cursor a voice by converting terminal agent outputs into concise spoken summaries. Flowacts provides a shared visual workspace where teams turn scattered AI chats into connected, transparent workflows. box by ASCII provides persistent, pre-configured Ubuntu virtual machines designed for running and scaling AI agents affordably. Kastra is a runtime authorization platform that enforces security policies on AI agents before actions execute. Thank you for reading today’s edition. Your feedback is valuable. Respond to this email and tell us how you think we could add more value to this newsletter. Interested in reaching smart readers like you? To become an AI Breakfast sponsor, reply to this email or DM us on X! Thinking of starting your own newsletter? AI Breakfast readers who sign up with Beehiiv receive a 14-day free trial and 20% off for 3 months. Keep Reading AI Breakfast Curated weekly analysis of the latest AI projects, products, and news