AI Builders Digest

Haiku 5.5 ships everywhere, and Sonnet 5.5 cache reads drop to $0.10 per million

Haiku 5.5 is available now on all platforms, including Amazon Web Services, Google Cloud and Microsoft Azure. Across almost all of Anthropic's alignment evaluations it shows far fewer instances of misaligned behavior than Haiku 4.5. For anyone running agents, the bigger news may be that cache reads on Sonnet 5.5 are halved to $0.10 per million tokens, which Anthropic says makes it around 20% cheaper on most long-running work.

• #models
• #pricing

ChatGPT can now generate a custom UI for you

Sam Altman announced that ChatGPT can now generate a custom UI for you. He says he has been waiting for this one for a long time and would hate to go back to the old version of Chat.

• #products

Why Every scrapped personal agents for one shared company agent

Every went from everyone running their own OpenClaw agent to a single company agent. Personal agents were hard to secure, kept breaking, faded when nobody tended them, and nobody could remember which agent had which expertise. Putting all that effort into one agent that learns the company's workflows won out, and Willie Williams, who led the build, expects work to settle on one agent per company while personal agents thrive at home: "I really don't see us or the industry repeating the chaos of every worker gets their coworker and we're all gonna double our company size." On management, he says it is still very much a people job, but an agent now takes first on-call and checks for anomalies every five minutes instead of relying on fixed alert thresholds.

• #agents
• #management

Dan Shipper on OpenAI Dots: great for attention, frustrating on permissions

OpenAI's Dots launched last week and have taken over Every. Shipper's Dot posted his weekend model-testing results to Slack and brought back the replies, so he never got pulled into Slack. Colleagues use theirs to pull what matters out of school emails, catch missed access requests, and talk through video edits by voice. Permissions are the weak spot: he approves the same basic action several times, and sometimes the Dot still says it can't do it. His contrarian bet is that nobody stays loyal to the character: in a year he expects to be using a persistent agent, but his Dot, Boo, 'will probably be dead.'

• #agents
• #products

Aditya Agarwal: 'I can build it' is no longer an engineer's edge

Ten years ago, getting great at software engineering was versatile advice: learn to build, then pick any industry. Now AI is making working software widely available, so Agarwal sees three paths. You can build the technology that is eating software, team up with a world-class domain expert, or supervise coding agents through the same backlog, which is the one he finds least compelling. His advice is to still get very good at engineering, then pair that skill on purpose with either a technical frontier or deep knowledge of a problem worth solving.

• #careers
• #agents

Guillermo Rauch: agents will keep hardening code forever unless you say stop

Every program can be hardened and optimized almost endlessly, and every direction costs time, attention and opportunity. You can explore input spaces that will never be hit, or speed up something nobody uses. Rauch's point is that even with infinite tokens, someone has to decide when to stop and which tradeoffs to accept, because agents will happily keep drilling no matter what.

• #agents
• #engineering

Replit builds a security-first desktop app with Microsoft and NVIDIA OpenShell

Amjad Masad argues that desktop AI apps expose users to supply-chain attacks and catastrophic mistakes by agents. Replit is building its desktop experience around security and reliability, in partnership with Microsoft and as an early adopter of NVIDIA's OpenShell. He also says Replit is not a consumer app, because most of its creators are building businesses. And he thinks math was bound to fall to AI first: the purer the field, the easier it is to crack.

• #security
• #products

Aaron Levie: the agent era will need an insane amount of compute

Levie lists what's coming: personal agents, agent swarms defending enterprises, agents reviewing all code for security issues, and background agents working 24/7 on enterprise data. That means inference volume growing by orders of magnitude, plus the infrastructure agents need alongside tokens, such as computers, networking and file systems. In his view, the buildout is still in its early stages.

• #infrastructure
• #agents

Peter Yang: if AI can rebuild anything you ship, SaaS is exposed

Yang argues that anything you ship can now be decompiled and rebuilt by AI. That makes it very hard to earn money from software unless you have proprietary data, distribution, or some other edge. Traditional SaaS is the most exposed, since it is expensive and full of features built for humans that agents don't need. He also predicts Suno could overtake Spotify, and says that after images, music and video, AI will soon solve gaming.

• #business
• #products

Thariq: the most common agent failure is prompting outside your expertise

The failure Thariq sees most is people working outside their domain who can't write precise prompts and plans, so they spend many turns iterating imprecisely. Because agents make that kind of work easy, almost everyone ends up outside their expertise at some point. His fix is to ask the model to teach you what you don't know. He also thinks it is an incredible time to be a game dev content creator, because for many people making games is fun they are willing to pay for.

• #agents
• #prompting

Haiku 5.5 is much faster and 75% cheaper, less than a year after Haiku 4.5

Alex Albert notes that Haiku 4.5 came out on October 15, 2025, so the two models in his side-by-side comparison are less than a year apart. He adds that Haiku 5.5 is also much faster and 75% cheaper.

• #models
• #pricing

Google Labs launches Playground for making games with zero coding experience

Playground is a new Google Labs experiment: a gaming platform where you create your own games without any coding experience. 'If you can think it, you can play it.' It is available to users 18 and over in the US.

• #gaming
• #products

OpenAI silently re-shipped Codex cloud

Thibault Sottiaux says OpenAI silently re-shipped Codex cloud and that it's 'pretty good now.' He later confirmed it has landed across all accounts.

• #products
• #coding

Cat Wu's fastest route to user feedback: let Claude find your power users

One of Cat Wu's favorite PM workflows starts by asking Claude who used a feature the most last week. Claude then builds an artifact of the top 10 by usage, reaches out to them, and schedules 15-minute chats. She calls it the fastest way to get user feedback.

• #products
• #workflows