The Sol Verdicts, Bun-in-Rust Finally Ships & OpenAI Retracts SWE-Bench Pro
The Sol Verdicts
Yesterday GPT-5.6 broke cover; today the considered reviews landed, and they're remarkably consistent: Sol is the new default for most work, Fable keeps a narrower crown.
Mitchell Hashimoto's review is the one to read (thread, ~730K views). After a month of early access, Sol is his default: faster, plans and judges as well as Fable, better overall work. His framing is already being quoted everywhere: "Sol is a charismatic, efficient, talented coworker you're jealous of. Fable is a genius recluse that is brilliant at its fixations but doesn't go out, doesn't date, and you don't want to hang out with them much." The carve-out: "Fable is undefeated at highly targeted debug/security/performance goals... I was never able to get Sol to push as hard in this category." Asked whether that means Fable wins on "pure determination" tasks, he agreed. He didn't get access to the rumored Luna/Terra variants.
The SST team's accidental A/B/A/B experiment (Jay's thread, 500K views) is the most interesting methodology story of the launch. They tested early 5.6 for weeks, loved it; tried Fable and were unimpressed (with a self-aware caveat about first-model bias); then the regulatory mess took both models away. The tell: when Fable came back, the team was still depressed that 5.6 wasn't available — and when 5.6 returned "it's immediately clear that it's just way better than Fable." One reply captured the principle: the real test of a better model "isn't liking it in a demo, it's the specific feeling of being annoyed you have to go back to the other one."
Theo distilled the emerging consensus into one image: "Fable is a 'wise owl' who is very thoughtful and very well spoken, GPT-5.6-Sol is like a rottweiler who will grab the problem by the throat and not let go until it is done" (post). He also argues the surge of other labs "catching up to the previous generation" is the best proof that Fable and 5.6 are genuinely next-generation, and is collecting questions for a batch of 5.6 videos.
The backlash is part of the story too. Multiple high-engagement replies to Jay's thread asked flat-out "were you paid to write a post today?" — the uniformity of 5.6 praise is itself becoming a talking point. Meanwhile the most-liked anti-Fable reply was about safeguards, not capability: one user says he'll be ecstatic to never again read Fable's "safeguards flagged this message" interstitial on routine coding work (post).
Practical notes: Peter Gostev warns that 5.6-Sol in Codex will trivially blow through a Pro plan if you run everything at
/fastwith 10x subagents — xHigh is the sane default. And steipete says 5.6 on Cerebras is "insane, 10x speed".
Bun in Rust Finally Ships
The blog post the community has been memeing about for two months ("we got this before GTA 6") is finally live.
Jarred Sumner published "Rewriting Bun in Rust" (announcement, 1.5M views; blog post). Beyond the writeup itself, the replies contain great process detail: Jarred ran the whole thing as "claude in a tmux session inside an ssh'd machine" (reply); the interactive build-timeline visualizations in the post were generated by asking Claude to download all the BuildKite annotations and build a React visualizer (reply); and the text is human-written but Claude-edited — his editing prompt was to have Claude web-search the topic and predict "what will the hacker news comments say? what should I change?" (reply).
Theo priced the rewrite at ~$165k if the tokens had been bought at API rates (post) — replies split between "that's absurdly cheap for a full runtime port" and "imagine explaining $165k in 11 days to make a fast product slightly faster."
Thariq (Anthropic) drew the bigger lesson (post): "this should be a huge update in your model of software engineering: rewrites can be good, cheap and fast." The best pushback in the replies: Bun is rewritable because every part of it is verifiable — "Most production code isn't built that way. The gap is architecture, not capability" (reply). Skeptics also note the Rust rewrite is still in canary months after the merge.
Claude Code & Anthropic Updates
New in Claude Code:
/checkup(Boris Cherny's announcement, 363K views). A spring-cleaning command that removes unused skills/MCPs/plugins to save context, dedupes your local CLAUDE.md against the checked-in one, breaks a bloated root CLAUDE.md into nested CLAUDE.mds and skills, disables slow hooks, pre-approves frequently-denied read-only commands, and more — confirming with you before every change. Reception is warm ("I was doing this manually just the other day"), with some sharp edges in the replies: several people ask how it differs from/doctor(one reports/checkupjust calls/doctorfor them), one notes the cleanup heuristic misses the worst offender — a skill that "runs every session and does nothing. still counts as used" (reply) — and the perennial "any news on AGENTS.md support?" goes unanswered.Claude Tag deep-dive webinar today: Cat Wu is hosting a live walkthrough (10am PT) of the progression "from single-player Claude Code to multi-player Claude Tag," including how the memory and proactive-work systems actually work (announcement). One grounded reply from a team that already ran Tag for a few weeks: it "churned through tokens and raced to our month max in less than 10 days" — variable cost is the adoption blocker (reply).
Benchmarks & Token Economics
OpenAI retracted its own recommendation of SWE-Bench Pro (announcement, 1.2M views): an audit found 30% of tasks broken, and it "no longer reliably measures frontier coding capability." Two reply threads dominate: the timing — "Why did you only bring this up right after xAI showed their results?" (Grok 4.5 had topped GPT-5.5 on it hours earlier) — and the deeper indictment: "the fact that nobody caught 30% broken until now is the real story. how long was everyone just trusting the numbers".
Armin Ronacher asks the question under all of this (post): is the cost of solving a task considered during RL — are models punished for wasting tokens? He's reacting to Matei Zaharia's finding that the minimal Pi harness matched vendor harnesses' success rates with Opus and GPT-5.5 at half the cost. "Models are getting better in their native harnesses, but at what price." Florian Brand (who works on this) confirms token efficiency is a factor in RL; a reply notes the Mythos/Fable model card repeatedly touts reduced cost-per-task without saying whether it was optimized for. One reply nails the stakes: "If token burn isn't in the reward, you get models that ace native harnesses and torch prod budgets."
swyx on Cognition's SWE-1.7 (post): the interesting part isn't the score, it's that "most agent labs are shy about acknowledging chinese model use because they need to sell to gov/defense" — Cognition did the hard part to productionize a Chinese base model: built a multilingual propaganda & censorship eval, corrected for it in post-training, and serves the result at 1000 tok/s (launch on Cerebras).
Grok 4.5 & the Catch-up Wave
Theo pinned "Grok 4.5 is a good model" (post) after realizing he'd been testing it extensively without knowing — "Pretty damn good and REALLY well priced." He's planning a video. One reply describes the workflow that's emerging: Fable as orchestrator, cheaper models like Grok 4.5 for delegated tasks.
Cursor partnered with SpaceXAI to train Grok 4.5 (announcement, RT'd by leerob): "our most powerful model yet and the first we've built for more than software engineering" — the first big fruit of the SpaceX/Cursor acquisition from June.
Theo's broader point (post): GLM-5.2 is close to Opus, Grok 4.5 allegedly beats Opus and GPT-5.5, Meta has something near 5.5 — the previous generation is commoditizing fast. "Feels like it barely matters now."
Agentic Coding & Craft
Matt Pocock vs. "code is cheap" (post): the phrase correctly communicates that code is cheaper to produce but wrongly implies it's disposable — "Code is the environment the agent operates in. Better code = better output." This is becoming a recurring theme (see Monday's "better models, worse tools" debate). He also shipped skills v1.1 (changelog):
/wayfinderfor planning ambitious work,/to-specand/to-ticketsreplacing/to-prdand/to-issues, plus/implementand/code-reviewcompleting the lifecycle. The skills are also no longer GitHub-specific — point the setup skill at any programmatically-accessible tracker.Simon Willison on AI commit messages (post): he's been letting Claude and GLT-5.5 write almost all of his commit messages "but I don't feel great about it." He's quoting a team lead who declared a moratorium on AI-written change descriptions: they describe details visible in the diff while "omitting the higher-level framing needed to understand broadly what the code is doing" — and the memorable line, "I'd rather see your prompt than your output." Simon's mitigation: link commits back to human-written issues, which carry the actual rationale. Worse than no rationale, he notes, is when the model guesses the rationale incorrectly.
Quick Hits
The OpenClaw Foundation launches — and steipete clarifies the org chart: "OpenAI hired me, not OpenClaw" (post). The foundation is an independent nonprofit with sponsors rather than owners, and for the first time a full-time team (led by Dave Morin) keeping the claw alive. 🦞
Prime Intellect raised $130M at $1B for "the open superintelligence stack" (Vincent Weisser) — thesis: pre-training concentrated frontier AI in a handful of labs, RL changes who gets to build it.
antirez pushes back on local-inference pessimism (post): responding to Theo's video on the hard limits of local models, he agrees with every word — then reports 50 tok/s on DeepSeek v4 Flash Q5 with two M5 Max 128GB systems and RDMA. LLMJunky is on the same crusade: a near-full-precision GLM 5.2 at home starts at 4x RTX 6000 Pro, and someone is running 8 clustered DGX Sparks with 1TB unified memory on vLLM.
LLMJunky × Cerebras: his first project with the Cerebras DevX team — "the world's first natively multimodal model to crack 2,300 toks/s" (post).
NerdSnipePod covers the whiplash week (episode): Fable was supposed to give 14 days' notice and gave 3 before the export ban; now it's back at half the rate limits. Chapters on the FDE revolution, AI Engineer Conference, Anthropic vs. Alibaba, and the OpenAI government stake.
Mysterious teaser watch: "Sol" wasn't the only thing being teased — jpschroeder's perfect non-announcement: "I can finally talk about Sol... I have nothing to say. Until tomorrow."