AI Builders Digest

OpenAI makes free ChatGPT text chats unlimited, powered by GPT-5.6 Luna

OpenAI removed the cap on free ChatGPT text conversations, now powered by GPT-5.6 Luna. Sottiaux also described the new bar for Codex with GPT-5.6 Sol: talk to it for five minutes about something that looks like weeks of work, get up to raid the fridge and pet the dog, and come back to find it done.

• #products
• #models

Altman: GPT-5.6 Sol is much better in chat now

Altman's version of the day's announcement: GPT-5.6 Sol is now much better in chat, and free users get unlimited text conversations. Notable that OpenAI is pitching its frontier model on conversation quality, not just coding.

• #models

Anthropic cuts Fable 5 biology fallbacks by 85%, widening health questions

Anthropic updated Claude Fable 5's biology safeguards to reduce false positives, cutting biology-related fallbacks by about 85% in testing so the model can handle a much wider range of everyday health and educational questions. Dual-use territory like virology, toxicology, and molecular design still falls back to Opus 5, with trusted access pathways promised for professional researchers and drug development.

• #safety
• #models

Claude lands in Apple's Foundation Models framework via a new Swift package

A new Swift package lets Apple developers call Claude through Apple's Foundation Models framework: the on-device model handles fast local work like summarization and extraction, then hands off to Claude for multi-step reasoning, code generation, web search, or data analysis. Because the framework returns typed Swift values from @Generable annotations, Claude gets clean structured inputs instead of raw user text. It ships tomorrow across iOS, iPadOS, macOS, visionOS, and watchOS 27.

• #products

Basis open-sources behavior specs: grade agents on process, not just outcomes

Passing evals does not mean your agent generalizes: "Even if you got it right a hundred out of a hundred times, if a person is just getting it right because they're going to Wikipedia, the accounting firm wouldn't hire them, and so they shouldn't hire us either." Basis, whose accounting agents run autonomously for hours to days on work like full tax returns, grades agents on process as well as outcomes using behavior specs: plain markdown files stating how an agent should behave, checked by a judge that is itself an agent, now released as an open standard with Braintrust. Troyanovsky's larger frame is that context is runtime training data, and he thinks most builders treat their code as more precious than their context when the reverse is true, since only the context affects performance.

• #agents
• #evals
• #open-source

Levie: prompting an agent is writing a spec, and the workflow itself has to change

Quoting a post he recommends, Levie argues that "prompting an agent is closer to writing a spec than asking a question", and that the real upside arrives when companies change the underlying workflow rather than bolting an agent onto existing processes. He predicts the vast majority of enterprise token usage will be agents deployed to execute tasks inside workflows, not chat. He also read Atlassian's big quarterly beat as evidence that agents generating 100x more code make systems of record more important, not less, because enterprises still need governance, security, and safe access to data.

• #agents
• #enterprise

Masad: no code is dead because the answer was always to solve code itself

Airtable bookends the rise and fall of no code, in Masad's telling: he argued endlessly with investors that UI could never let you build arbitrary software, and that making software accessible meant solving code itself, which sounded delusional until now. He also shared some history: in 2021 and 2022 he asked Google, Meta, and everyone else in the valley to train coding-specific models with Replit, nobody thought it mattered versus NLP, so Replit trained its own Replit-code-3b before the industry got code pilled.

• #coding
• #products

Rauch: the Plugin standard makes AI coding agents universally extensible

Rauch's thesis: devtools must be open source and universally extensible, and AI coding agents are the most important devtools in the history of the industry. The Plugin standard lets anybody extend them uniformly, so building one plugin gets you distribution across the tidal wave of software creation from CLIs, IDEs, cloud agents, and personal assistants.

• #open-source
• #agents

Yang: consumer AI is ChatGPT and Google's market to lose

Consumer AI is ChatGPT and Google's market to lose: both have roughly a billion users, and the real barriers are mainstream users not trusting AI with full access to their apps and data, and not realizing what it can now do, which makes marketing and messaging as important as the product. His sharpest claim is that having the best model barely matters to normal people; his non-AI friends could not care less about Sol versus Fable as long as pricing is fair and the work gets done reliably.

• #products
• #consumer

Madhu Guru: record yourself talking through the idea, ship that as the doc

People are much clearer when they speak through a new idea than when they write a doc about it; somewhere between the mouth and the page, the core idea gets buried under context and polish. His fix for his teams: record yourself explaining the idea exactly as you would to a friend, use AI for basic cleanup while keeping the rough original structure and flow, and share that as the doc.

• #productivity

Kothari: fundraising tips no VC will tell you, starting with the 10% floor

A brain dump of fundraising mechanics founders rarely hear: get warm intros to actual GPs, ideally from founders they have already backed, know that no established lead fund will go under 10% dilution, and never anchor your valuation to what a competitor raised at, which he calls the single best way to tank a deal. Do not lie about term sheets either, because verifying one is a text away in a market where everyone talks and every deck leaks. He says this is the most consensus market he has seen in a while, so founders outside hot sectors should come armed with their contrarian case and a default-alive plan.

• #funding