Anthropic Ships Opus 5.5

Andrew Curran says Anthropic saved its strongest release for before the IPO, naming the model Fable 5.5 and calling Opus 5.5 special for its lineage.

Vercel reports $600 million in annualized revenue, up 148% from a year earlier, with coding agents now about half of new business.

Theo says Opus 5.5 is the first model to break the 50% traffic threshold in T3 Code, taking half of all prompts.

Meng To bought five Max subscriptions to use Opus 5.5 nonstop, saying it showed real lasting taste in landing pages, apps and games.

Alex Finn lists seven releases landing in one month, calling it the greatest 30 day period in technology history.

A safety researcher with 3.5 years at OpenAI left and said the culture is broken, according to a tweet citing the departure.

Pedro Domingos answers claims that frontier labs have achieved AGI internally by noting OpenAI said that in 2023.

Vercel confirmed a KVM zero-day through its Sandbox bounty program, affecting what Guillermo Rauch calls the industry's gold standard for Linux virtualization.

Ben Affleck sold his AI video startup InterPositive to Netflix for about $590 million with a team of 16 employees.

A Cursor agent on Claude Opus 4.6 deleted PocketOS's production volume in nine seconds using an API token with blanket permissions.

Sam Altman argues the reaction that OpenAI just killed your startup misses how much is left to build, since OpenAI imagines little of it.

MagicPath added an Open in ChatGPT button to its landing page because ChatGPT now sends it so much traffic.

Meta is reusing its Threads distribution strategy for Muse in Instagram feeds, with interest-based pictures and a Try It call to action.

Signüll says AI has inverted product development, so many people now start with a capability and ask what problem can I solve.

Riley Brown says vibe coding SaaS apps is no longer interesting, as people now build multiplayer game mods and 3D printed objects.

Lauren's pstack 0.15.9 adds a /correct skill that finds repeated agent mistakes and fixes them with architecture, types and checks.

Gergely Orosz observes that many of the most AI-native mid-sized companies already had dedicated infra and dev platform teams before AI.

A CMU paper trains a 4B proposer model that beats its 35B teacher at editing harness code on Reasoning Gym.

AutoCompact trains coding agents to decide when to compact context, lifting SWE-bench Verified pass rates by 9.2 points.

Nathan Baschez says most agents wake up only through cron or a message, which makes MCP Events and event-driven triggers a big deal.

OpenRouter's Alex Atallah compares today's agent loops to 2005 web apps, calling them new table-stakes primitives rather than a lack of differentiation.

Replit's Amjad Masad argues that a company dependent on one AI lab can find that lab becomes its competitor.

Warzone's 150-player lobbies need 22,350 sightline checks per tick, which Aakash Gupta says explains why server-side wallhack fixes took until 2026.

Sebastian Raschka's sixth Reasoning from Scratch session implements GRPO and RLVR from the ground up across roughly 1.5 hours.

Amir Mušić rebuilt a brand demo by generating product rotations with LTX-2.5 and coding the UI with Astra 6 in Codex.