GPT-5.6 Breaks Cover, Fable's Stay of Execution & Anthropic Sues a Customer

GPT-5.6 Breaks Cover

Theo's first-look review: "Not quite as 'smart' as Fable, but it is incredibly capable." The embargo cracked open overnight and Theo posted the first substantial hands-on report (4.1K likes, 207K views): GPT-5.6 "fixed all the problems I had with GPT-5.5," is "incredibly determined — will run for a day without even using a /goal," "understands subagents incredibly well and is great at orchestrating," and "knows iOS dev incredibly well." His verdict: "For many things, gpt-5.6-sol will become my obvious default." On instruction-following he says it's better than 5.5 — "understands intent well and hammers until it gets there. Sometimes a bit too hard" — and he didn't get to test the "ultra" tier. In a follow-up he adds that gpt-5.6-sol is "world leading in computer use": "When we lost access to 5.6, I quickly started to go insane without it."

The context around the launch. Theo notes OpenAI has been "awesome to work with despite the obvious government-related setbacks" and more transparent than expected — the only restriction being that formal content (blogs, videos, podcasts) waits for the official release. Fellow early-access tester Max Weinbach's take, RT'd by steipete: "the most impressive part is this model NEVER GIVES UP. If you throw it in Max reasoning, it will just keep working until it's done. GPT 5.6 Sol is my favorite model BY FAR." And Thorsten Ball went further on code review: "After more than a decade of required reviews in various codebases, I think there's very little a human can do better in a review than a GPT-5.6." swyx's entire commentary: it_happening.gif.

The reply-section discourse is all about the Fable comparison. The line "not quite as smart as Fable" is doing heavy lifting: "We're so cooked chat", "This is the line everyone should take note", and the sharpest question in the thread from @0xPectations_: are we comparing against "the actual Fable users can access today — the one that gets lobotomized / falls back to 4.8 for coding — or the mythical 2.5-day Fable that nobody will ever get to use again? Because 'Fable' has become two completely different things." Also floating in the replies: a mention of "Terra" as a possible 5.5-level workhorse tier, and cybersecurity folks asking about refusals — "I'm in cyber, I cannot use Fable for literally anything." Theo says it's "gonna be a fun week"; expect the full reviews Thursday per Weinbach.

Fable's Stay of Execution

Extended through July 12 — announced six hours before the deadline. After yesterday's grief-cycle around the 11:59:59pm PT cutoff, @claudeai dropped a one-liner that did colossal numbers (73K likes, 13.7M views): "We're extending access to Claude Fable 5 on all paid plans through July 12." The timing is the story — one reply computes the subtext: "From a Bayesian standpoint I calculate the probability that you are terrified of OpenAI releasing 5.6 at 100%."

The problem: everyone already torched their quota racing the old deadline. The dominant reply, in a hundred variations, is reset the limits. heather's version is the cleanest: "Unless there's a reset, this benefits no one, since anyone who actually wants to use Fable has already used their weekly limit expecting access to expire tonight." See also: "Reset then. Everyone already used it up" (732 likes), "Another 5 day extension? Really? Do we just have to keep begging for scraps each time?", and an accusation that the whipsaw is deliberate: "Make them rush to use all of their fable credits. Great, now extend fable's availability but don't reset their limits." Simon Willison got caught by it too: "Now I'm regretting having carefully blasted through 100% of my Fable allotment in time for being cut off at the end of today!" Mckay Wrigley speaks for the power users: "my life is currently completely arranged around fable rate limits right now." And LLMJunky sums up the economics: "Can't overwhelm compute if you don't have any usage left." Practical tip from the thread that got traction: run Fable at low reasoning to save ~80% — "hardly any quality drop and results are better than Opus."

Anthropic vs. Abnormal

Anthropic quietly sued Abnormal AI — and Abnormal found out from a reporter. steipete surfaced the story by quoting the most damning line from Abnormal CEO Evan Reiser's response post: "We're a very large customer of Anthropic and they still have yet to tell us about the lawsuit. I learned about it from a reporter, not our 'partner.'" The facts per Abnormal: Anthropic filed suit July 1 claiming unfair competition and trademark infringement — arguing Abnormal's "/\" logo is confusingly similar to Anthropic's "A" mark — and asks for "disgorgement of all revenues, earnings, profits, compensation, and benefits." Abnormal's defense: the company was founded in 2018 (three years before Anthropic existed), the slash-based identity was designed by agency ALINE in April 2021 (months after Anthropic's founding, before Claude existed), the wordmark is pixel-for-pixel unchanged for ~5 years, and Anthropic holds no registered trademark covering cybersecurity products. Reiser also draws a pointed product line: Abnormal's threat detection runs on its own behavioral models — Claude is an internal productivity tool, not part of the customer-facing stack.

The court of X leans one way. "Filing a trademark suit against your own large customer without giving them a heads up first is a bad look regardless of who's technically right on the logo", and "a $10M/year customer finding out from a reporter says more about the relationship than the trademark claim does." Coming the same week as the Fable rate-limit whiplash, the thread doubles as a referendum on Anthropic's customer relations — with a side of gallows humor: "They're going to make poor Thariq respond to this, aren't they? 🤣"

Agentic Coding & Agent Harnesses

Matt Pocock's system-prompt bloat purge: proxy it, gasp, delete. His step-by-step guide (1.7K likes, 397K views) to killing Claude Code system-prompt cruft: (1) run a proxy so you can see exactly what gets sent to Anthropic, (2) "Fuck, there is so much cruft in there," (3) use his settings.json to strip it — down from ~25K to a clean 13K tokens per session, full process on aihero.dev. The results people report are dramatic: one user measured startup context dropping from 96,602 to 28,928 real input tokens (70.1% saved) with tools cut from 115 to 20, and another trimmed 21,654 → 13,644 tokens per turn. To the inevitable "the model is fit to the harness, just use it as built" pushback, Matt's reply: "I see you have not run a proxy. Please follow step 1 and see if you still agree." Useful tooling from the thread: Datadog's free local proxy Lapdog traces reasoning and tool calls across Codex, Claude Code, and Pi — and one observer notes this is exactly the journey that led Mario Zechner to build pi.dev in fall 2025. Asked why he stays on Claude Code at all: "Buy vs build. I'm buying for now."

Thariq live-tweets Fable editing 60GB of conference footage. swyx sent him four raw videos (~59GB) of his AIE keynote, and Thariq threw them plus his HTML slide deck at Fable inside a lightweight video-editing harness — "just a bunch of scripts and remotion assets" he'll open source if there's interest. The pipeline: Whisper transcribes the whole talk, Claude renders low-res drafts to dodge the real bottleneck (render time), and because the deck is HTML, Claude turns static slides into animated section intros and even tracks his position on camera to lay out slides-plus-speaker. It also cut YouTube Shorts unprompted and added flourishes he suspects human editors skip only because they're tedious. A reply corroborates at scale: someone's doing the same thing with 2TB of footage. (The thread's replies are also ground zero for rate-limit anger — "mogging us while we're out of fable creds" — see above.)

sqlite-utils 4.0 ships — with Fable as release manager. Simon Willison released sqlite-utils 4.0, the 124th release but first major bump since 2020 (annotated release notes). The AI angle: after Fable found five release blockers in the weekend's $149 review session, he ran one final Fable review before shipping and it kicked up another FOUR blockers. He also wrote an upgrade guide explicitly designed to be fed to a coding agent so it can apply the breaking changes for you — a small sign of docs being written for agents as first-class readers.

The Local-Model Fight

Theo crashes out about local models; the local-AI crowd crashes back. His pinned thread (960 likes) plus the GLM 5.2 flashpoint: "GLM 5.2 is an incredible model. I wish people would stop pretending it is self-hostable and that it compares to Fable." The rebuttals are substantive. Sentdex's point-by-point takedown: batching is how you run subagents locally ("massive throughput with tensor parallelism"), idle GPUs draw ~12W so the $2K/year power math assumes 24/7 agents ("which is literally the strongest case to run local"), DeepSeek V4 Flash hits 350 tok/s single-stream and 1,500 tok/s throughput at native precision, and he cancelled his OpenAI and Claude subs when GLM 5.2 dropped. will brown notes a researcher self-hosting GLM-5.2 on an NVL72 GB300 rack with great daily-driver results — the catch being that's data-center hardware. The best steelman is 0xSero's, which Theo himself RT'd: frontier subs are "absolutely a better deal from a cost, intelligence, and ease standpoint" today, but the subsidies won't last forever, and nine months after Sonnet-4.5 had orgs spending hundreds of millions, a comparable model now runs on a $4,000 box: "Is it perfect? Is it easy? No. But it'd be yours."

Meanwhile, the actual numbers from LLMJunky's garage. Receipts for what local looks like right now on two RTX 6000 Pro Blackwells: DeepSeek V4 Flash (284B MoE, 13B active, 1M native context) decoding at 137 tok/s single-stream — "this model is no slouch" — and Hy3-NVFP4 (295B MoE) at 25 tok/s, memory-constrained to 90K context. On the hardware front: leaks point to an RTX 5080 SUPER with 24GB DDR7 at ~$1,300 as the budget 3090 replacement, and NVIDIA is openly recruiting for open-weights work — LLMJunky's read: "NVIDIA knows the US has been LACKING in the open weights space... they are going to take up the mantle and push US open source forward."

Quick Hits