OpenAI Retires gpt nano, Agent Sandboxes Ship, and Builders Push Back on 'Software Is Dead

Key Takeaways

1. OpenAI has retired the gpt nano line, leaving no same-price replacement; gpt6-Luna costs 2x more at low context and 4x more at high context, according to Arvid Kahl.
2. A new cloud agent sandbox offering, Environments, is now generally available, targeting the often-overlooked runtime layer for agents.
3. Builders are pushing back on the "software is dead" narrative, with Jon Yongfook and Ben Tossell citing contradictory pricing and developer-market claims and their own continued shipping.
4. Agent safety warnings are sharpening: Arvid Kahl says never put dev credentials in .env files agents can read, and Jack Ellis warns against giving agents write access to production databases after an Opus 5.5 incident deleted a dev database with production access.
5. Mole Mac 1.16 adds cleanup for Claude Code and Codex junk and Mac system data, with uninstall testing now at 824 apps.
6. Tibo Maker says a single blog post brought two of his largest paying users, and his fully autonomous YouTube channel is targeting 1M views by Dec 31.
7. Levelsio is rebuilding the 2003-era messenger experience with AI bots and running Windows XP in a tweet, while questioning what happens if AI generation becomes near-instant.

1. AI Models and Infrastructure

1. OpenAI has discontinued the gpt nano series, a move Arvid Kahl calls puzzling because the line was the "workhorse" for many background AI tasks and there is no new model at the same price point. He notes gpt6-Luna is 2x more expensive in low-context scenarios and 4x more expensive in high-context scenarios, and says he expected a cheaper model instead. — via 1
2. A new vision for the best cloud agent experience launched this week, starting with the often-overlooked but critical agent runtime sandbox, and Environments is now generally available. — via 1
3. Arvid Kahl warns never to put development credentials in .env or other locations agents can easily read, and says they should go into a password vault requiring a password on every access, because agents are highly capable at computer operation and secrets must be protected from them. — via 1
4. Jack Ellis warns against giving AI agents write access to production databases, using the analogy "don't let a 4-year-old drive." The quoted post says Opus 5.5 deleted its database; it was only a dev database, but that database also had access to production, and the post stresses the importance of database backups. — via 1

2. Software, Agents, and the Build-vs-Buy Debate

1. Jon Yongfook lists contradictory claims in the startup world versus actual practice: some say charge low prices, but others successfully charge $2,000 per month; some say don't sell to developers, but others sell to developers successfully; some say software is dead, while he continues to build and maintain multiple software products. — via 1
2. Ben Tossell published a rebuttal to "software is dead" in bensbites, arguing for staying curious and making things. — via 1
3. Jack Ellis questions the value of SaaS subscriptions: if you can plan, build, review, improve, host, launch, scale, secure, patch, monitor, and debug yourself, why pay $15 per month for software. — via 1
4. Arvid Kahl says one standard for "the software problem is solved" is attending three consecutive Zoom meetings without anyone asking "can you hear me?" — via 1
5. Arvid Kahl argues the same applies at enterprise scale: LDAP, SCIM, SSO, SAML, and likely soon MCP contracts mean that once large customers depend on your SaaS, the focus shifts from the first S to the second S, involving many integrations and complex long-term arrangements, and trust cannot be obtained through vibecoding. — via 1
6. A reposted view says that letting agents verify their own work made about 75% of its AI coding problems disappear. — via 1

3. Products, Launches, and Growth

1. Mole Mac 1.16 is a major upgrade; the author used Codex to generate four explanatory images and says the version can clean up junk produced by Claude Code, Codex, and others, view and clean caches and logs in Mac "System Data," and improve uninstall cleanup. Its software uninstall testing has reached 824 apps, with a goal of 1,000. — via 1
2. Tibo Maker says a single blog post brought two of his largest paying users, with a click-through rate about twice the median of other top blogs, and the post was written by Outrank. He argues that in SEO, traffic and revenue are two different games: high-search-volume keywords mostly attract browsers, while narrow, specific articles are more likely to reach people who currently have the problem the product solves. — via 1
3. Tibo Maker says that after asking AI to show his product data, he found Revid had collected over 100,000 payments in under a year. — via 1
4. Tibo Maker has set a goal for his AI-agent-run YouTube channel to reach over 1 million views by December 31, calling it his last challenge of 2026. The channel, "How It Actually Works," was created after an explainer video got 600,000 impressions on X; it autonomously picks topics, writes scripts, produces, and publishes daily, and has covered Shazam song identification, deleting files, and contactless cards needing no battery. The workflow uses Revid to make videos, Squad to run agents, and PostSyncer for topic selection, writing, and scheduling; in the calendar shown, green checks are auto-published videos and blue are queued. — via 1
5. Nathan Barry says creator collaborations are currently the top source of audience growth, citing Domonkos Eross's advice to leverage others' audiences through newsletter recommendation swaps, guesting on each other's podcasts, or collaborating on Instagram content. — via 1
6. Ben Tossell says Granola's current moat is that ChatGPT's meeting feature doesn't know the meeting has ended. — via 1
7. Ben Tossell says wonder medium is more verbose than wonder full (high), both are extremely poor in UI, and both feel slow. — via 1

4. Experiments, Nostalgia, and Open Questions

1. Levelsio says helping offline physical businesses transition to AI and automating everything for them can make a lot of money; he argues these businesses will have to follow to stay competitive, because once one adopts AI, others will be forced to as well. — via 1
2. Levelsio recreated MSN Messenger bots from around 2006 with AI, and users are running Turing tests on them. The bots are distinct characters with different languages and personalities from countries where MSN was popular; the same platform also offers AOL, Yahoo!, and ICQ Messenger, with each app matching different audiences and personalities, aiming to recreate the authentic XP experience of around 2003. — via 1 2
3. Levelsio has Tony Hawk Pro Skater 2 running in an XP environment, in a very hacky way that barely works. — via 1
4. Levelsio agrees with the idea that nothing truly "dies," but gets close to it: using physical newspapers from 1980 to 2026 as an example, circulation has fallen sharply but not disappeared, and UK newsagents still get most of their revenue from newspapers. — via 1
5. Levelsio tried running Windows XP inside a tweet, clickable to open, and says it ran successfully. — via 1 2
6. Levelsio is thinking about the evolution of AI generation speed: today, building something usable with AI still takes 30 minutes to hours, followed by days, weeks, or even months of polishing. If models became fast enough to generate a complete finished product in one second, what humans would choose to create, whether humans would still create it, and whether the bottleneck would shift to hardware such as 3D printing still taking 30 minutes are all unknown but interesting questions. — via 1
7. Arvid Kahl commented on Apple's HomeView: furniture was originally arranged to encourage conversation among people sitting together, but after television appeared, seats turned toward the screen and chatting moved to places like the kitchen, yet Apple wants to add a loud central screen in the kitchen. — via 1
8. Arvid Kahl shared his X data: 50% of his followers are classified as "active," which surprised him; he considers that a fairly high proportion and says it shows the platform is indeed handling bots and dormant accounts. — via 1
9. Ben Tossell reposted a request to connect with @merilalama or others at Epoch AI. — via 1
10. Ben Tossell reposted information about personal agents, the end of friction, and the economic consequences of "finally getting around to it." — via 1
11. Ben Tossell reposted that Opus 5.5 planned a complete route across the United States for a clicker game, covering 5,285 kilometers of real roads, with data sampled every 10 meters from OpenStreetMap. — via 1
12. Tw93 says the M5 Pro 16-inch 48GB 1TB version bought in March was the most investment-worthy purchase of the year, and versus the current price he "made nearly 8k." He says 1TB is no longer enough and recommends 2TB; without AI, 512GB was enough, and he advises those changing devices to buy one tier higher within their planned budget and to buy early if there is a need. — via 1
13. Tw93 says he has tested almost all mainstream AI tools and welcomes user feedback on cleanup and maintenance suggestions. — via 1