▶ YouTube Summaries

• https://www.youtube.com/watch?v=mHG7K7QmQyU
• Original title: So I've been using gpt-5.6 for awhile...

Theo got early access to GPT-5.6 before Fable dropped and spent between $180K and $240K of inference over roughly six weeks, working across 67 projects to stress-test what the model can actually build. This video is explicitly NOT a review (that comes later) — it is a tour of the concrete things he built and how the model behaved, deliberately recorded before he read any external coverage or sentiment. His headline takeaway: 5.6 is a workhorse that understands intent far better than 5.5, latches onto a task and won't let go, and can run autonomously for many hours (even 20+) without getting lost or needing context babysitting.

Before the project tour, Theo lays out what he deliberately will NOT cover: Luna and Terra (no access during testing), a Fable comparison (its own dedicated video), ultra/max reasoning levels, the "death of Codex", benchmarks and sentiment (hasn't seen them), pricing and usage limits (didn't have them), and the fast version (his was only 1.5x faster, not the 750 TPS Cerebras is promising). He stresses OpenAI did not pay for this coverage, he is still paying full price for his own $200 plan, and the $200K figure is him pushing limits — not a realistic monthly use case. He credits the model's much-improved computer/browser use as the single most immediately impressive trait, and pitches sponsor Browserbase (hosted puppeteer/browser infra for agents) as the way to give agents the web without managing Playwright yourself. Two recurring themes throughout: (1) 5.6 is a genuine workhorse that grabs a goal and runs for hours, unlike 5.5 which would get lost, stop, or need a fresh thread; and (2) it is far better at sub-agent orchestration — he admits he had to retroactively swap earlier PR-review loop demos to look like 5.5 because 5.5 actually couldn't do that work.

The project tour spans a lot of ground. On Lakebed (his cloud app-building platform he's eager to ship), 5.6 migrated giant monolithic 5.5-era JavaScript files into a properly structured TypeScript project, built CI/preview/release pipelines, artifact storage and Railway buckets, CLI login (a non-trivial production-server-to-CLI auth relationship), whitelist access requiring a centralized first-party auth service (forking his OAuth service), hardened the capsule DB reactivity layer, researched runtime isolate options, modeled pricing/quotas, and landed dozens of PRs — turning a small side project into something he's excited to open-source. On T3 Code (a unified agent-management layer with desktop, web, and mobile apps), the standout was mobile: he had 5.6 do two separate from-scratch native rewrites of the React Native app — one in AppKit/Swift, one in SwiftUI — each fully feature-complete end-to-end in just 2–4 hours, aided by the model spinning up the iOS simulator via computer use to verify itself. He also built a full computer-use loop into T3 Code, rebuilt the marketing site, and set up a daily Hermes automation to triage all T3 Code PRs. The rewrite experiments are his favorites: a Rust port of his Hermes agent (a ~15MB-RAM memory-safe reimplementation that actually runs, replies, uses skills — built after having 5.6 mine his and Ben's chat history to find the ~50% of features actually used), and a wildly ambitious Rust rewrite of the TypeScript Go compiler. The latter produced a genuinely working narrow no-check transpiler (up to 18x faster than the Go version) across ~195K lines of Rust in 29 crates — but per Fable's own analysis it's a broad prototype with a tiny verified slice, ~15–20% toward a usable tool and only ~5% toward a real TSGo replacement, with tests currently failing.

Other builds pile on: a Dropbox-like dev-folder sync (FS2) launched via a single /goal that burned 71.2 billion tokens (~$91K at fast-mode pricing) and autonomously registered for PlanetScale on its own to build what it needed — Theo hasn't even opened the result yet. A bootable Codex/Claude recovery USB drive he can plug into any Linux machine on his network for pre-configured remote access. A dramatic story where 5.6, after initially hallucinating nonexistent HP BIOS options, used computer use through his remote KVM to reboot into a broken GRUB state, drop to a shell, boot correctly, open a terminal, and fix the boot partitions fully autonomously — the kind of high-risk sysadmin task he calls an "oh, we're really there" moment. Plus a "fleet" repo that auto-configures new machines from context, a SkateBench overhaul, a network-presence tracker (who's home/at the office via MAC addresses), a Presto-alternative experiment, "claudex" (running his Codex auth inside Claude Code), a terminal token tracker, Prime Day deal automation via browser use, and a Fish Slop 3D game where the model generated its own environment, rocks, textures, and monster models. The consistent caveat: 5.6 is NOT a strong front-end model — it's more steerable and has less bad default taste, but he repeatedly had to rebuild UIs (e.g. with Opus 4.8), and its 3D understanding is "strange" and not yet shippable, though progressing fast. His closing note: this model got him genuinely excited to build bigger — it directly inspired his recent CascadiaJS and AI Engineer "think bigger" talks — and he's curious how everyone else feels, with a proper review, a Fable comparison, and usage-tips videos all coming soon.