AI Builders Digest

OpenAI reopens $200 Pro and halves its effective usage, announced the night before DevDay

Pro $200 subscriptions reopen to new subscribers with a changed usage calculation that nets out to about half the API spend of the old plan. The stated reasons: no reintroduction of the 5-hour limit so weekly usage can be spent freely, and a refusal to inflate API list prices just to make subscription value look larger. The bet is that API prices keep falling fast enough that buying usage as needed converges with subscribing, citing GPT-6 Sol and GPT-6 Luna launching this week at 50% of their prior price. Something new is also being added to the subscription that won't draw on usage, unrevealed until the announcements.

• #products
• #pricing

Box's enterprise eval: Sonnet 5.5 is 2.4x faster to a deliverable on 12% fewer tokens

Box tested Sonnet 5.5 in early access against its complex-work eval with the Box Agent and saw a 4 point overall gain on its hardest tests, reaching the finished deliverable roughly 2.4x faster while using 12% fewer tokens. Vertical gains were larger: +18 points in Financial Services, +8 in Life Sciences, +7 each in Legal and Public Sector. The task-level detail is the interesting part. On a due-diligence review it caught miscalculated interest totals and mispriced options in a deal book, and on a commercial lease review it refused to invent a 'standard' market benchmark for renewal terms and flagged what the lease was actually missing instead. Sonnet 5.5 lands in Box AI Studio shortly.

• #evals
• #agents
• #products

Sonnet 5.5 finishes ~30% more Claude Code tasks by needing fewer tokens, not more

In Claude Code, users complete roughly 30% more tasks with Sonnet 5.5 than with Sonnet 5, and the mechanism is intelligence rather than budget: it needs fewer tokens for the same work. A yard-raking demo using tool calls finished 24 seconds faster than Sonnet 5 on 6,000 fewer tokens.

• #products
• #agents

Sonnet 5.5 beats Opus 5.5 at revision, and some stacks have no room for mid-tier at all

Sonnet 5.5's writing is dramatically improved even against Opus 5.5, and it wins on revision tasks specifically, though Astra remains the favorite for writing outright. Being faster and cheaper than Opus 5.5 makes it the pick for quick iterative coding and design work on Every's team. The dissent inside that same team is the sharper signal: two of its members feel they no longer have room in their stack for mid-tier models. Separately, after attending every OpenAI DevDay since 2023, this one has by far the most launches, floated as possible recursive self-improvement.

• #evals
• #products

Nobody can just show you their prompt anymore

Prompting as a shareable artifact is dead: everything now runs on references, skills, and examples, with an agent routinely told to read three other repos, search the web for references, and call other AI APIs before starting. The token-cost objection to higher-level abstractions like projects, tags, and dynamic workflows should be fading too, since Sonnet and Opus 5.5 make that grade of intelligence cheap enough to be available by default. The specific recommendation: use Sonnet 5.5 when building workflows.

• #agents
• #products

Vercel opens domain search without auth, explicitly for agents

Vercel domain search now works without authentication, pitched as especially useful if you're an agent rather than a human. Separately, a migration run with as little intervention as possible produced roughly 70% faster builds and 75% faster paints on a mature workload full of the previous provider's idioms, finished in under a week. Two AI skills were derived from that migration and will be shared back.

• #products
• #agents

The 'distribution is a moat, so raise a lot' investors have gone quiet while incumbents flex

The investors who preached raising huge rounds to buy distribution are absent right now, precisely when incumbents are demonstrating what real distribution looks like. The counter-position: stay first principles, find your genuine unfair advantages, build a product with high retention and ideally network effects, and treat capital as a weapon for compounding rather than as destiny. Underneath it is a warning that the money spigot will dry up, and that the advice-givers will have moved on to the next portfolio company by the time it does.

• #funding

Sonnet 5.5 has the Opus 5.5 feel: clear writing, very fast, major jump over Sonnet 5

Sonnet 5.5 carries the same quality that made Opus 5.5 appealing. It writes clearly, runs very fast, and represents a major capabilities jump over Sonnet 5, making it a strong model to iterate with.

• #products
• #evals

The 'OpenAI is cooked' / 'Anthropic is cooked' cycle flips every few months and means nothing

X swings from 'OpenAI is getting mogged by Claude 5.5' to 'Anthropic is so cooked by Codex' and back within months, which says more about the audience than the labs. The useful framing is that two frontier competitors, soon more, are pushing each other and everyone downstream benefits. On the building side: a playable StarCraft level made with Sonnet 5.5 where you defend a Terran base against the Zerg with Marines, Siege Tanks, and Battlecruisers, using SC2 models from Sketchfab and music generated with Suno.

• #products
• #evals

Sonnet 5.5 debugging Claude Code: 30% faster on 30% less usage

A demo of Sonnet 5.5 fixing a bug in Claude Code itself, landing 30% faster while consuming 30% less usage than before.

• #products
• #agents

Side-by-side Sonnet 5 vs 5.5 demos: bouncing-ball physics and code-drawn pixel art

Two community builds showing the generational gap from the same prompt: bouncing-ball physics tests run on Sonnet 5 versus Sonnet 5.5, and pixel-art forest creatures where every frame is drawn in code rather than generated as an image.

• #products

A YouBike app now covers every city in Taiwan with live turn-by-turn voice navigation

The YouBike project now works in all cities in Taiwan with bike share, covering dock locations, navigation, and live turn-by-turn directions with voice.

• #products

Altman ahead of DevDay: 'We have found a new thing'

A teaser for DevDay claiming OpenAI has found a new thing, with no further detail. It lands the same evening as the Pro plan repricing disclosure.

• #products

Codegen inside WhatsApp makes a lot of sense

Endorsement of putting code generation into WhatsApp as a genuinely sensible distribution surface rather than a novelty.

• #products

Open call: builders who can't get AI-generated frontends to stop looking like slop

Looking to talk with people in three situations: building web and frontend experiences with AI but struggling to make them look good rather than like AI slop, seeing impressive demos on X built with the latest models and having no idea how to reproduce them, or building those demos and willing to share the process.

• #products