Guillermo Rauch
Guillermo Rauch shared a website migration performed with as little manual intervention as possible. He reported roughly 70% faster builds and 75% faster paints. The team completed it in under a week and derived two AI skills it plans to share. In another update, Vercel domain search no longer requires authentication. He specifically highlighted easier domain discovery for agents.
Peter Yang
Peter Yang built a playable StarCraft level with Sonnet 5.5. Players control Terran, building Marines, Siege Tanks, and Battlecruisers to defend against the Zerg. The project uses Catholomew’s SC2 models on Sketchfab and music generated with Suno. He also reflected on X’s rapidly shifting winner-and-loser narratives about OpenAI and Anthropic. He values competition pushing capabilities forward more than declaring either company finished.
Dan Shipper
Dan Shipper evaluated Sonnet 5.5 using a personal benchmark drawn from real work. He found substantially better writing than Opus 5.5; Astra remains his favorite for writing, but Sonnet did better at revision tasks. Its lower cost and greater speed make it Kieran Klaassen’s preference for rapid coding and design iterations. Two other colleagues feel they no longer have room for mid-tier models. The differing reactions preserve the role of tasks and personal workflows in model choice.
Thariq
Thariq argues that asking to see someone’s prompt is no longer enough to reproduce a result. Actual workflows include references, skills, and examples. He often asks an agent to inspect three of his repositories, search the web for references, and use other AI APIs. When discussing projects, Claude Tag, and dynamic workflows, he frequently hears concerns about token cost. He recommends trying Sonnet 5.5 and believes it and Opus 5.5 make those higher-level workflows more affordable.
Aaron Levie
Aaron Levie reported Box Agent’s early Sonnet 5.5 results: a four-percentage-point overall gain on its hardest tests, roughly 2.4 times faster delivery, and 12% fewer tokens. Financial services, legal, life sciences, and public-sector results improved by 18, 7, 8, and 7 points respectively. A due-diligence task gained 48 points, catching incorrect interest totals and option pricing while finishing about 61% faster. Commercial lease review gained 14 points without inventing market “standards” for renewal and notice terms. A field-trial report gained 22 points and finished about 41% faster, correctly calculating sample standard deviations in both groups. A math-intervention report gained 11 points and finished about 71% faster with the correct student progress figures. These are Box’s evaluations; Sonnet 5.5 is expected in Box AI Studio shortly.
Thibault Sottiaux
Thibault Sottiaux previewed usage changes accompanying the reopening of the $200-per-month Pro subscription. At API list prices, the new allowance works out to about half the old plan’s amount. He also committed not to restore the five-hour limit, letting users allocate their weekly usage. The team expects more efficient models and cheaper APIs to deliver more high-quality work for the same cost; he noted that GPT-6 Sol and GPT-6 Luna had fallen to 50% of their previous prices that week. He does not want artificially inflated API prices to make subscriptions appear more generous. Additional features that do not consume usage were planned for the next day but had not yet been disclosed.
Nikunj Kothari
Nikunj Kothari challenged advice that treats fundraising for distribution as a moat. Incumbents are also exercising their distribution advantages. Founders still need distinctive strengths and a great product with strong retention and, ideally, network effects. He advocates using capital to compound existing advantages rather than treating it as destiny. The business ultimately needs to sustain itself because abundant financing will not last forever.
Cat Wu
Cat Wu said Claude Code users complete roughly 30% more tasks with Sonnet 5.5 than with Sonnet 5. She attributed the improvement to using fewer tokens for the same work. Her demonstration involved clearing leaves from a yard through tool calls. Sonnet 5.5 finished that example 24 seconds faster. It also used 6,000 fewer tokens.
Boris Cherny
Boris Cherny demonstrated Sonnet 5.5 fixing a bug with Claude Code. The example was 30% faster and used 30% less usage.
Ryo Lu
Ryo Lu shared a tool he built for YouBike. It works in all Taiwan cities with YouBike service. Features include docking-station information and navigation. Riders can also get live turn-by-turn directions with voice. Ryo provided a link to try the experience, turning his enthusiasm for YouBike into a practical tool.
Alex Albert
Alex Albert shared his initial impressions of Sonnet 5.5. He said it retains the feel he liked in Opus 5.5. Its writing is clear. It is fast and represents a substantial capability improvement over Sonnet 5. He particularly likes it for iteration.
Zara Zhang
Zara Zhang invited followers building web and frontend experiences with AI to talk. One group can build functionality but struggles to make the interface look polished. A second sees new-model demos but does not know how to reproduce them. A third already creates impressive demos and wants to share processes and tips. She invited contact through direct messages or comments.