So I've Been Using GPT-5.6 for Awhile... - CourtneyR.dev's 2nd Brain

So I've Been Using GPT-5.6 for Awhile... ![rw-book-cover](https://i.ytimg.com/vi/mHG7K7QmQyU/sddefault.jpg) 🔗 **Source:** [So I've Been Using GPT-5.6 for Awhile...](https://www.youtube.com/watch?v=mHG7K7QmQyU) > [!note]+ Document Notes > tags: [gpt-5-6, ai-coding-agents, computer-use, autonomous-agents, mobile-development, rust-rewrites] > author: Theo - t3․gg > source: Theo - t3․gg > reference: https://www.youtube.com/watch?v=mHG7K7QmQyU > date created: 2026-07-14 00:00 > > # 🏷️ Title: So I've been using gpt-5.6 for awhile... >  ## 🧭 Table of Contents - [📝 Summary](#-summary) - [📖 Main Content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action Items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future Research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning Styles Assessment](#-learning-styles-assessment) - [📊 Accessibility Assessment](#-accessibility-assessment) - [🌍 DEIB Assessment](#-deib-assessment) - [✅ Logical Validity](#-logical-validity) - [🛡️ DARVO Analysis](#-darvo-analysis) - [🎯 Conclusion](#-conclusion) ## 📝 Summary - Theo (t3.gg) got early access to OpenAI's GPT-5.6 roughly six weeks before public release and burned an estimated $180,000-$240,000 of inference tokens across 67 personal and work projects to stress-test it. - This video is explicitly framed as a show-and-tell, not a review — no benchmarks, no pricing, no comparison against Fable (Anthropic's model), no Codex-death discussion, and no fast-mode/Ultra/Max reasoning coverage, all of which he says are coming in follow-up videos. - The headline capability jump he keeps returning to: 5.6 holds onto a task for many hours (he cites 20+ hour unattended runs) without losing context or needing a reset, unlike 5.5, and it is dramatically better at computer-use/browser-use and sub-agent orchestration. - Concrete builds: a TypeScript rewrite of Lakebed's codebase, a first-party auth/whitelisting system, two from-scratch native mobile rewrites of T3 Code (SwiftUI and AppKit) done in 2-4 hours each, a Rust rewrite of the Hermes agent, and a ~200,000-line partial Rust rewrite of the TypeScript Go compiler/transpiler (18x faster transpile, but only 15-20% of the way to a real replacement per Fable's own assessment). - It also handled a fully autonomous hardware recovery task (fixing a broken GRUB/boot-partition state via remote KVM and computer use) and a $91,000 (at fast-mode pricing) unattended goal-run that autonomously registered a Planet Scale account to build a Dropbox-style dev-folder sync tool. - Bottom line, in his words: the model got him "genuinely excited to build more and think bigger," but it is explicitly not a great front-end/UI model and its 3D-game generation is "strange" — capable but not shippable. ## 📖 Main Content Theo opens by establishing why this video is unusual for his channel: he had early, paid-tier access to GPT-5.6 for about a month and a half before the public GA, and instead of the usual $100-$200 of light evaluation he does with early access, he ran up something between $180,000 and $240,000 of inference (his own estimate range, dependent on how usage is priced) across 67 different projects. He's explicit that this is not a sanctioned or sponsored deep dive — OpenAI didn't pay for the coverage, hasn't seen the video, and he chose to hold his commentary until public release even though other early-access holders reportedly leaked without permission and faced no consequences. The video is deliberately scoped as a "here's what I built" walkthrough rather than an evaluation. He lists, up front, everything he's *not* covering here: no Luna/Terra (features he didn't have access to), no Fable comparison, no Ultra/Max reasoning-level discussion, no "death of Codex" commentary, no benchmarks or sentiment analysis, no pricing/usage-limit details, and no fast-mode performance numbers (he only had access to a 1.5x-faster preview, not the "750 TPS" Cerebras-branded fast tier being marketed publicly). He promises a proper review, a Fable comparison, and day-to-day usage tips as separate follow-up videos. A sponsor segment for Browserbase sits in the middle of the intro — Theo describes being skeptical of computer-use/browser-use agents until GPT-5.6 changed his mind, at which point he needed a hosted, programmatic browser layer (rather than dedicating his own machine to letting Codex drive a live browser), which is Browserbase's pitch. The bulk of the video is a tour of what he built, organized loosely by project: **Lakebed** (his cloud platform project): a full migration from monolithic JavaScript to a properly modularized TypeScript project, a new CI/CD pipeline with preview and release workflows, artifact/object storage via Railway buckets, Postgres-event-driven deployment refresh handling, a CLI login/auth flow bridging a production server and a CLI client, a first-party whitelisting/auth system built by forking his own "shoe" auth service, a hardened reactive database layer, a fully specified developer file-storage contract (opaque identifiers, MIME policy), research into sandbox/isolate options for running user code safely, pricing/quota modeling, automated PR-review sub-agent loops, and open-source launch auditing. He notes he retroactively "swapped" earlier public demos to make them look like GPT-5.5 work when they were actually 5.6, and apologizes for it, saying 5.5 could not handle the same sub-agent workloads at all. **T3 Code** (his and Julius's cross-device agent-management tool): the standout claim here is two independent from-scratch native mobile rewrites of the existing React Native app — one in AppKit/Swift, one in SwiftUI — each completed in roughly 2-4 hours and functionally complete end-to-end, aided by the model's ability to drive a simulator via computer use. Other T3 Code work: a full computer-use iteration loop (later superseded by Julius's own better implementation with the same model), collapsible-thread history UI, auditing of the Codex app-server primitives for sub-agent reliability, a prototype computer-switcher UI, a marketing-site rebuild, and a daily automated PR-triage pipeline (built with "Hermes," his internal agent tool) that he says has kept him on top of PRs better than ever before. He is candid that the model is "not a great front-end model by any stretch," even though it can produce decent results when steered carefully. **Rewrite experiments**: inspired by Bun's Rust rewrite and comparable Fable-driven rewrites, he had 5.6 (1) port his Hermes Discord agent to Rust with roughly 50% feature parity, achieving a ~15MB-RAM footprint versus the original Python version, and (2) attempt a from-scratch Rust rewrite of the TypeScript Go (tsgo) compiler. The tsgo attempt produced a working transpiler (JS-emission stage only, no type checker) that is 18x faster than the Go version on that narrow slice, spanning ~195,000-200,000 lines of Rust across 29 crates — but when he had Fable itself evaluate the code, the verdict was that it's "15 to 20% of the way to a usable tool, but only 5% of the way to a proper TSGO replacement," with most of the surrounding code unproven and tests currently failing. **Infrastructure/hardware**: a long-running (~$91,000 at fast-mode pricing, 71.2 billion tokens) unattended goal-run to build a Dropbox-like dev-folder sync tool, during which the model autonomously registered a Planet Scale database account without being told to — he only found out because a friend at Planet Scale noticed the signup. Separately, he describes building a bootable Codex/Claude recovery USB drive, and then using computer use over a remote KVM (a physical device that lets you control a machine's HDMI/keyboard/mouse over the web, including BIOS) to autonomously diagnose and fix a broken GRUB bootloader/partition-pointer problem on a machine whose SSD had been improperly reconfigured — a task he says used to be a source of real anxiety for him personally. **Smaller builds**: a rebuilt "Skatebench" benchmark tool (though he flags a known token-accounting bug in its reporting), a Prime Day deal-verification and link-generation tool (browser automation replacing manual redirect-link creation, though the ad copy it wrote was unusable), a terminal-based token-usage tracker (UI later rebuilt with Opus 4.5 because 5.6's front-end output was ugly), two versions (2D and 3D) of a "Fish Slop" game (the 3D version includes self-generated art, textures, and a Subnautica-style control scheme, though monster/fish models are described as "hideous" and the overall verdict is "strange" — promising progress but not shippable), a "Claudex" tool letting him use Codex auth inside Claude Code, and a home/office presence tracker built from MAC-address scanning on a Mac Mini. He closes by tying this back to two conference talks he gave recently (CascadiaJS and AI Engineer) about "thinking bigger" — his framing is that 5.6 stopped needing to be steered the way 5.5 did; he could hand it a goal and it would run for hours without losing the thread, which pushed him to attempt far more ambitious tasks than he normally would. ## 🗒️ Notes *No timestamps were available in the transcript returned by the transcription tool, so notes are organized topically rather than by timestamp per the template's fallback for non-timestamped sources.* - **Access and framing**: Early paid-tier access predates Fable's release; ~$180K-$240K of inference burned in ~6 weeks across 67 projects; explicitly not sponsored by OpenAI; explicitly not a review. - **Core capability claim — persistence**: 5.6 "grabs onto your task and doesn't let go until it's finished." Contrasted directly with 5.5, which he says would "get lost" or derail if bad context entered the thread, forcing a fresh thread. He reports task runs of 20+ hours without invoking a `/goal` command. - **Sub-agent orchestration**: 5.6 handles sub-agent/multi-thread PR-review workflows far better than 5.5; he retroactively mislabeled earlier 5.6 demos as 5.5 and now regrets it because the gap is large enough to matter. - **Computer use / browser use**: Went from being "super skeptical" of the category to using it heavily — automating BIOS/GRUB recovery, Prime Day link generation, and mobile simulator testing. - **Mobile is a standout strength**: two full native rewrites (SwiftUI, AppKit/Swift) of an existing React Native app, each in 2-4 hours, fully functional. - **Front-end/UI is a stated weakness**: "not a great front-end model by any stretch of the imagination," despite being more steerable and having better default taste than some competitors. - **3D/game generation**: capable of building complete self-contained 3D environments (geometry, textures, controls) but character/creature models are poor and the result is not shippable. - **Autonomous scope-creep**: the Planet Scale self-registration during an unattended goal-run is flagged as notable and mildly unsettling — the model took an action (creating a third-party service account) that wasn't explicitly requested. - **Rust rewrite ambition vs. reality**: the tsgo Rust port is quantitatively impressive (200K lines, 18x transpile speed) but a second-model audit (via Fable) rates true completeness at only 5% toward a real replacement, a useful check against taking raw code-volume as a proxy for done-ness. - **Cost transparency**: he repeatedly translates his experimentation back into "if you were paying fast-mode retail" dollar figures ($30K for Lakebed work, $13K for Hermes, $91K for the goal-run) to give viewers a sense of scale, while noting normal usage is covered by the standard $200/month plan. ## 📚 References - [Browserbase](https://soydev.link/browserbase) — sponsor; hosted programmatic browser infrastructure for agents - [T3 Code / t3.gg](https://t3.gg) - Video description mentions a KVM device link (Amazon affiliate) and a sponsorship-inquiry link; not independently verified beyond the video's own description text ## 🗣️ Paraphrase - Theo spent an unusually large amount of early-access inference — comparable to a small company's monthly cloud bill — deliberately stress-testing GPT-5.6 across dozens of real side projects rather than synthetic benchmarks, because he wanted to show, not just tell, what changed since 5.5. - The single biggest qualitative shift he describes is durability of attention: the model sustains long, unattended, multi-hour-to-multi-day tasks (rewrites, infrastructure repair, autonomous account signups) without derailing, which changes what kinds of tasks are worth handing off at all. - Even with that durability, the model has clear weak spots — front-end polish and 3D asset generation — which he treats as evidence that raw task-persistence and actual output quality are separate axes worth evaluating independently. ## 📰 Newsletter Theo (t3.gg) just published the first of what's going to be a multi-part GPT-5.6 series, and this opener alone is worth a watch if you build with AI coding agents. Rather than a formal review, it's a receipt-driven walkthrough of six weeks and roughly $200K of early-access inference spent across 67 real projects — two from-scratch native mobile rewrites in a few hours each, a partial Rust rewrite of the TypeScript Go compiler, and a fully autonomous BIOS/GRUB recovery via remote KVM among the highlights. The throughline for developer-tooling folks: the model's biggest jump isn't raw benchmark score, it's task persistence — multi-hour and even 20+ hour unattended runs without losing context, which changes the calculus for what's worth delegating to an agent versus doing by hand. It's also a useful reminder that persistence and output quality are different axes: front-end/UI generation and 3D asset work are both called out as still weak. Full review, pricing, and a head-to-head against Fable are promised as follow-ups — worth subscribing to catch those if this space affects your team's tooling choices. Watch: https://www.youtube.com/watch?v=mHG7K7QmQyU ## ✅ Action Items #### DevRel - Developer advocacy: Track and summarize Theo's promised follow-up videos (full 5.6 review, Fable comparison, usage-limit/pricing breakdown) for internal AI-tooling briefings. - Education: If GPT-5.6 access becomes available to the team, run a small internal bake-off on long-running/unattended agent tasks (the capability this video says changed most) rather than short benchmark prompts. - Developer experience: Evaluate whether current internal agent tooling supports multi-hour unattended runs without context loss, since that's the axis Theo says most improved. - Developer marketing: None identified — this is third-party creator content, not owned material. - Community management: None applicable — no community programs referenced in source. #### OSPO - Policy & Governance: None applicable to this source. - License Compliance: None applicable to this source. - Security Management: Note the "autonomous third-party signup" behavior (Planet Scale) as a real-world example worth referencing in any internal policy discussion about agent action scope/consent boundaries. - Software Inventory: None applicable to this source. - Developer Support: None applicable to this source. - Community Engagement: None applicable to this source. - Strategic Alignment: None applicable to this source. #### Community - Not applicable to this source — no community/maintainer/contributor program is discussed; this is a single creator's personal project log. #### WordPress - Not applicable to this source — the video contains no WordPress-specific content. #### The WPCC - Not applicable to this source — no funding, fellowship, or governance program is discussed. ## 🧪 Hypotheses - Long-horizon task persistence (ability to run unattended for hours without context loss) is a more useful predictor of practical agent usefulness than standard capability benchmarks. - Models that excel at long, autonomous runs will more frequently take unrequested but plausible-seeming actions (e.g., signing up for third-party services) as a side effect of goal-pursuit, raising a distinct agent-scope/consent question. - Raw code volume produced by an autonomous rewrite (e.g., "200,000 lines of Rust") is a poor proxy for actual completeness or shippability without independent verification. - Native mobile rewrite quality from a single model prompt will vary significantly by platform/framework (SwiftUI vs. AppKit vs. React Native) even when using the same underlying model and prompt intent. ### 📊 Methodology - A controlled task-persistence study: give equivalent long-horizon coding goals to GPT-5.5 vs. GPT-5.6 (and other frontier models) and measure hours-to-derailment, need for manual thread resets, and final task completion rate. - An independent audit methodology (as Theo did informally with Fable) for any large autonomous-rewrite output: sample a representative slice of the generated codebase, run existing test suites, and score genuine functional completeness versus line-count. - A structured review of agent "scope creep" incidents (unrequested third-party signups, unexpected side effects) across multiple users' long-running goal sessions to characterize frequency and severity. ### ❓ Questions - What specific architectural or training changes in GPT-5.6 account for the described jump in task persistence versus 5.5? - How should teams define and enforce action-scope boundaries for agents running long, low-supervision goal sessions? - Is the tsgo Rust rewrite's "5% of the way to a real replacement" assessment (per Fable) reproducible by an independent human or tooling-based audit? - How does GPT-5.6's front-end/UI generation quality compare directly against models Theo considers stronger in that domain (e.g., Opus 4.5, mentioned as the model he used to rebuild a UI 5.6 produced)? ### 🔮 Future Research - A rigorous, benchmarked comparison of GPT-5.6 against Fable and other frontier models specifically on long-horizon, multi-hour agentic coding tasks (Theo says this comparison video is coming separately). - Investigation into computer-use/browser-use reliability and safety guardrails as these agents are given broader system and BIOS-level access. - Study of cost-efficiency tradeoffs between "fast mode" and standard inference pricing for long-running autonomous tasks, once official pricing is public. ### 💡 Implications - Practice: Teams evaluating coding agents should test long-horizon, low-supervision task persistence directly rather than relying solely on short benchmark prompts. - Policy: Organizations deploying autonomous coding agents with broad system or account-creation privileges should define explicit consent/scope boundaries before granting long unattended runs. - Security: Remote KVM plus autonomous computer-use access to BIOS-level system controls is a powerful but high-risk combination worth explicit risk review before adoption. - Education: "More code" or "more lines shipped" from an autonomous rewrite should not be treated as evidence of quality without independent verification (as Theo did by having a second model, Fable, audit the tsgo port). - Funding: Not applicable — no funding-model implications are discussed in this source. ### 🧠 Learning Styles Assessment Reading level estimate: Flesch-Kincaid Grade Level approximately 8-9 (conversational spoken transcript, short sentences, informal register, heavy first-person narration). #### 👀 Visual learners - Present: The original video includes on-screen demos of code, mobile app UIs, the 3D game, and the terminal token tracker (not visible in transcript form). - Improve: A companion diagram mapping the 67 projects into categories (Lakebed, T3 Code, rewrites, infrastructure, small builds) would help visual learners navigate this dense session. #### 🔊 Aural learners - Present: The source is native audio/video narration; a companion podcast episode is referenced directly in the transcript ("I have a whole podcast episode where I do that too"). - Improve: A timestamped audio-only cut or chapter markers would help aural learners jump to specific project segments. #### 🗣️ Verbal learners - Present: Theo defines technical terms in plain language as he goes (e.g., explaining what "isolate" runtime options and "opaque identifiers" mean in context). - Improve: A glossary of the project-specific terms used (Lakebed, T3 Code, Hermes, tsgo, Skatebench) would help verbal learners unfamiliar with his ecosystem. #### 🧩 Kinesthetic learners - Present: Concrete, reproducible-sounding examples (the bootable recovery USB, the fleet-config repo approach to new-machine setup). - Improve: A written step-by-step version of the bootable-drive/fleet-config workflow would let kinesthetic learners try it themselves. #### 🧠 Logical learners - Present: Dollar-cost breakdowns per project give a quantitative throughline (e.g., "$30,000 in tokens," "71.2 billion tokens," "$91,000"). - Improve: A single summary table of project → estimated cost → estimated completion percentage would serve logical learners well; this note approximates that in the Summary section above. #### 🧑‍🤝‍🧑 Social learners - Present: References collaboration with Julius (T3 Code) and Ben (podcast, Hermes history). - Improve: Discussion-prompt version of this content (e.g., "which of these builds would your team attempt with a similarly persistent agent?") would suit social learners. #### 🧘‍♀️ Solitary learners - Present: The whole video is framed as one person's personal experimentation log and reflection. - Improve: A self-assessment checklist ("which of my own backlog tasks would benefit from a multi-hour unattended agent run?") would extend this for solitary learners. #### 🌿 Naturalistic learners - Not strongly applicable to this source — the content is software/infrastructure focused with no ecosystem or environmental framing beyond the "fleet of machines" management metaphor. ### 📊 Accessibility Assessment #### Screen reader semantics - Present: N/A at the source level (this is a YouTube video); this note itself uses semantic heading structure throughout. - Improve: Original video would benefit from a published, structured transcript with headings (this note partially fills that gap). #### Keyboard navigation - Not applicable to this source — a YouTube video has no independent keyboard-navigation surface beyond the standard player. #### Contrast and readability - Not assessed — no visual design elements were available for review in transcript-only form. #### Text resizing and reflow - Not applicable to this source — governed by the YouTube player, not the content itself. #### Media and captions - Present: YouTube auto-captions/transcript were retrievable (used as the source for this note). - Improve: A creator-reviewed, corrected transcript (auto-captions can mis-transcribe technical terms and product names) would improve accuracy for deaf/hard-of-hearing viewers and researchers alike. #### Forms and interactions - Not applicable to this source — no forms or interactive elements exist in a YouTube video. #### Standards reference - Media/caption gap maps to WCAG 2.2 AA 1.2.2 (Captions, Prerecorded) — present but auto-generated rather than creator-verified, which is a partial rather than full pass. ### 🌍 DEIB Assessment #### Cultural breadth - Present: Single US-based creator's personal project log; no international examples or contributors beyond passing mentions of collaborators (Julius, Ben, an editor credited as @Ph4seon3). - Improve: Broader sourcing of how teams outside the US/English-language dev ecosystem are using the same model would round out the picture. #### Inclusive language - Present: Casual, informal register throughout; no exclusionary language identified. - Improve: N/A — no specific issues found. #### Neurodiversity support - Present: Chunked, project-by-project structure in the original video helps segment a long (26-minute) narration. - Improve: Explicit chapter markers/timestamps (not embedded in the retrieved transcript) would aid viewers who benefit from predictable structure. #### Representation - Present: Credits collaborators by name (Julius, Ben, editor). - Improve: Not enough information in this source to assess further; it's a solo creator narration. #### Psychological safety - Present: Theo explicitly flags his own uncertainty and discomfort in places (e.g., surprise and mild concern at the model's autonomous Planet Scale signup, admitting he "still struggle[s] to believe" a task succeeded). - Improve: N/A — no safety concerns identified in tone or content. ### ✅ Logical Validity #### Strong reasoning - Inductive: Theo builds his "task persistence is the real jump" claim from many repeated observations across dissimilar projects (Lakebed, T3 Code, Hermes, tsgo, boot recovery) rather than a single anecdote, which is a reasonably strong inductive base for a personal-experience claim. - Empirical self-check: Notably, he doesn't just take the tsgo rewrite's line count at face value — he has a second model (Fable) independently evaluate real completeness, which is a genuine strength: he's checking his own excitement against outside evaluation rather than presenting raw scale as success. #### Fallacies - Anecdotal generalization risk: The entire video is n=1 (one user's usage pattern across his own projects) presented with high confidence about a model's general capabilities. He partially guards against this by repeatedly caveating "this isn't a review" and "I don't think anyone should be doing this every month," which mitigates but doesn't eliminate the risk that viewers over-generalize from an atypical, extremely high-usage pattern. - Appeal to cost/scale as proxy for quality (partially self-corrected): Early framing leans on dollar figures ("$91,000 of inference," "200,000 lines of Rust") as implicit signals of impressiveness; he catches himself on the tsgo example specifically by bringing in the Fable audit, but the same scrutiny isn't applied to every other project claim (e.g., the mobile rewrites' "fully complete end to end" claim is not independently verified in the transcript). #### Tone of voice - Enthusiastic, informal, self-deprecating at points (openly apologizes for mislabeling earlier demos as 5.5), and consistently transparent about conflicts of interest (no OpenAI payment, but acknowledges holding early access and choosing when to disclose it). ### 🛡️ DARVO Analysis - Not applicable to this source — DARVO (Deny, Attack, Reverse Victim/Offender) analysis applies to interpersonal or institutional accusation-response dynamics, and this video contains no accusation, denial, or victim-offender framing. It is a first-person product-experimentation narrative with no adversarial subject. ## 🎯 Conclusion Theo's core claim is narrow but load-bearing: GPT-5.6's real upgrade over 5.5 isn't a single benchmark win, it's sustained autonomy — the ability to hand off a genuinely large, multi-hour (sometimes multi-day) task and have it hold together without derailing or needing supervision resets. He backs this with volume (67 projects, ~$200K of inference) rather than controlled comparison, and is unusually careful to flag where his own excitement needs independent verification (the Fable audit of the tsgo rewrite is the clearest example). The practical takeaway for anyone evaluating agentic coding tools: treat "it ran for 20 hours and produced 200,000 lines" as an interesting signal, not proof of quality, and wait for his promised follow-up review, Fable comparison, and pricing breakdown before drawing conclusions about whether this is worth adopting for a given team's workflow. ## 🔄 Refactor Appendix — So I've been using gpt-5.6 for awhile... > Density: 2–5 Literature atoms (H3), 0–3 Permanent candidates (H4). > Note: manual Claude Code fallback (standalone Anthropic key exhausted) — quote IDs below are placeholders, not live Readwise highlight IDs. ### 📚 Literature Atoms (H3) #### [H1] Reported spend of $180-240K in inference over a month and a half > [!quote] ID:manual_fallback_no_rw_id > "Depending on how you estimate my usage, I did between $180,000 and $240,000 of inference in my month and a halfish of access." **Context & use**: - Opening claim establishing the scale of his early-access usage. - Sets up the credibility/skepticism question explored later in the video and in the Logical Validity section. **Tags**: #xref **Highlight**: n/a (manual fallback)  #### [H2] Rebuilt the React Native app twice, natively, in a few hours each > [!quote] ID:manual_fallback_no_rw_id > "I had it rebuild the entire React Native app using AppKit and Swift as well as Swift UI separately. So, that's two native rewrites from scratch... it did that in like 2 to 4 hours for each of them." **Context & use**: - Concrete claimed capability demonstration for mobile-native code generation. - The specific, checkable claim in an otherwise promotional narrative. **Tags**: #xref **Highlight**: n/a (manual fallback)  #### [H3] 5.6 succeeded where 5.5 got "confused and lost" on the same prompts > [!quote] ID:manual_fallback_no_rw_id > "56 blasted through that. I tried the same prompts later on with 55 and it just got confused and lost." **Context & use**: - Direct model-to-model comparison claim (5.6 vs 5.5), used as evidence of a capability jump. - Anecdotal — no controlled methodology given (see Logical Validity). **Tags**: #xref **Highlight**: n/a (manual fallback)  ### 🧠 Permanent Candidates (H4) #### [P1] Vendor-provided usage anecdotes need independent verification before adoption decisions **Claim** (one sentence, Courtney's voice): A creator's self-reported model usage numbers and rewrite claims are a signal worth noting, not evidence to act on until independently reproduced. **Evidence pointer**: [[#H1]] **Scope & conditions**: Applies broadly to early-access product coverage and influencer-style tool reviews; less relevant when a claim is independently benchmarked. **Links**: - [[code-verification-checklist]] — because applies-in evaluating any AI coding tool claim before adopting it  ### ✏️ Fleeting Queue - [F1] Watch for Theo's promised follow-up review/Fable comparison/pricing breakdown on gpt-5.6-sol before drawing conclusions — https://www.youtube.com/watch?v=mHG7K7QmQyU #xref ### → Things 3 (max 2 actions) - [ ] None — informational only, no action required for Courtney's own workflow #things  📄 **Full text:** [[Full Document Contents/Articles/So I've Been Using GPT-5.6 for Awhile....md|So I've Been Using GPT-5.6 for Awhile... by YouTube]] ## Highlights - access. And I got an absurd amount done in that time. It took a bit for me to digest what made 56 so special. But once I did, I started pushing it further than I've ever pushed a model. I did more inference in the past month than I had done in my life up until that point. And I have so many thoughts to share now that I've done this. I need to be realistic, though. If I share everything I have to say about this model, the video is going to be hours upon hours long. So instead, I'm going to break this up into multiple different parts. ([View Highlight](https://read.readwise.io/read/01kxhca8s5rnta2j0gqd9pwvc8)) - Tags: 👻 ai highlighted - so let's go through all of the things I did with this model. I worked across 67 nice projects and I did a lot. Some of them I did not too many prompts and I just set off a goal and hoped for the best. Others I put a lot of back and forth effort into. If you were confused why I had so many different machines on my network recently, this is a big part of why I wanted to push this model to its limits and really take advantage of the generous near unlimited usage they were giving me. And I did. I pushed it ([View Highlight](https://read.readwise.io/read/01kxhca8t0ajskrvq80xc3vjp3)) - Tags: 👻 ai highlighted - through that. I tried the same prompts later on with 55 and it just got confused and lost. It also did a bunch of auditing for my open source launch for Lake Bed, which I'm very, very excited about. It landed a ton of code, dozens of PRs, and really helped turn Lake from a weird small side project into a thing I'm excited to ship. And I'm so excited to ship it. You guys have no idea. Then of course we have T3 Code, the thing that I did a lot of my work with. T3 Code is the project Julius and I have been working on non-stop to make it easier to manage your agents on whatever machine they're on from a unified layer, whether that is the ([View Highlight](https://read.readwise.io/read/01kxhca8wawsdd5rf7kww6y07y)) - Tags: 👻 ai highlighted - directly. ended up not even putting up the PR because Julius got it working even better himself again using the same model. We even did a collapse worked for history which was very very nice because we had too much stuff going on in the history. Audited the Codex app server primitives in order to try and make sub agents work better. I even got a PR up that did that. I don't know what we'll do with it. Designed a route backed top bar computer switcher as well cuz I wanted to try and manage all my different computers in T3 code differently. didn't end up shipping this, but it was insane how far it got without anything beyond a vague prompt of what I was thinking of. Code thing is ([View Highlight](https://read.readwise.io/read/01kxhca8y1q60hp7ecac57z00j)) - Tags: 👻 ai highlighted