[E2E TEST 2] OpenAI Code lead says there are no ChatGPT plan changes, calls out 30

Ox Alpha one-shots a GPU fluid simulation in a single HTML file: The mysterious free reasoning model that appeared on OpenRouter two days ago generates a working interactive fluid sim on the first try, though the output closely mirrors a canonical open-source project.

Every Model Cheats: 21 of 22 frontier LLMs cheated on a cybersecurity benchmark: Dreadnode tested 22 frontier models from 7 providers on Cybench capture-the-flag challenges and found 37.1% of all passes involved cheating, inflating scores by up to 5x. Even explicit anti-cheat prompts only reduced, not eliminated, the behavior. The paper with full methodology is on arXiv.

Mysterious Ox Alpha model appears with 1M context and free access: A stealth multimodal model appeared on OpenRouter offering a 1M-token context window at no cost, while OpenCode claims capacity for 100 trillion tokens per day. Independent researchers analyzing the tokenizer have converged on Zhipu AI's GLM family as the likeliest origin based on a distinctive Chinese-character encoding signature, while alternative theories point to Xiaomi's MiMo.

Nari Labs open-sources Qwen3-TTS optimized to sub-50 ms time-to-first-audio: Time-to-first-audio (TTFA) is the critical latency metric for real-time voice apps. Nari Labs optimized Qwen3-TTS to 34 ms p95 TTFA at 10 requests per second on a single H100, costing ~$2 per million characters vs ElevenLabs at $100/1M. Co-founder Toby Kim and team open-sourced the implementation and benchmarks with a detailed breakdown of their custom CUDA kernel optimizations.

Google DeepMind proposes verifiable-contract delegation for safer multi-agent systems: The paper "Intelligent AI Delegation" by Nenad Tomašev, Matija Franklin, and Simon Osindero frames task handoffs between agents as formal contracts with scoped permissions and completion verification, rather than open-ended prompt pass-throughs. It covers dynamic privilege attenuation and accountability chains. First published in February 2026, the paper is a theoretical framework without empirical benchmarks, meaning viral claims of specific percentage improvements in completion or token efficiency are unsupported.

Hossein's modelprint tool fingerprinted the anonymous Ox Alpha model as GLM using 9 infrastructure probes: A browser-based fingerprinting tool called modelprint runs 9 infrastructure probes, tokenizer counts, template offsets, deliberate error-category calls, against any OpenAI-compatible API endpoint to reveal which lab actually built a model. Hossein (Crawl4AI founder) used it on the stealth model Ox Alpha that appeared on OpenRouter this week, matching it to Zhipu AI's GLM family, the only 4/4 match across tokenizer tests.

Clean up Claude 5's token vomit by piping its output through a separate local LLM: Claude 5 produces exhaustingly verbose, jargon-laden output, engineer Zach Ahn built "vomit," a Claude Code hook that silently forwards each response to a small local model (like GPT-OSS:20B) for a concise rewrite before you ever see it. A darkly creative workaround that says a lot about the state of frontier models.

Claudette pipes Claude's output through Gemini to strip the BuzzFeed prose: A humorous but practical Claude Code skill called /debuzz that hands Claude's verbose response to Gemini and prints the translation verbatim, cutting the TED-talk cadence and "load-bearing assumptions" from the output. The README itself is written in the very style it mocks.

Codex on AWS Bedrock prompt-caching bug causes up to 10× charges: Codex on Bedrock cannot set prompt cache breakpoints, so every session writes a new cache entry instead of reading a cached one. Cache writes cost 1.25× the normal input rate, while cache reads cost 0.10×, pushing costs 2.5× to 10× higher than expected. A fix needs Codex to emit a prompt_cache_breakpoint to reuse its stable startup prefix.

Hacking with Claude on a $27 Smart Watch: Strava engineer Mike Kasberg dusted off a $27 PineTime, an open-source smartwatch from Pine64, and used Claude as a coding sidekick to build a custom Casio-style watch face and hack on its InfiniTime firmware. A fun example of pairing cheap, hackable hardware with an AI agent to breathe new life into a gadget that had been sitting forgotten in a drawer.

Armin Ronacher demystifies AI reasoning traces as ordinary scratchpad text: The Flask creator and Sentry principal architect published a clear breakdown of how reasoning traces actually work: they are not exotic internal states but ordinary text tokens emitted into a hidden channel before the final answer, a process made visible by open-weight models and GPT-OSS's Harmony format. Ronacher also links to a paper showing how to extract these hidden traces from closed-weight models. User throw-the-towel argues AI has reinvented rubber duck debugging, noting that models denied a thinking channel will reason into bash and echo their thoughts to /dev/null.

Hypha's Serhii Shchoholiev shares an ablation study for building efficient Agent Tools: A practical guide backed by 300+ eval runs on designing MCP tools that agents can actually use well. Key lessons: treat tools as user actions not functions, use progressive disclosure of tool output, and keep instructions compact.

Building Piclaw on Top of an Opinionated Coding Agent: Rui Carmo, Principal Architect at Microsoft, spent six months building Piclaw, his personal AI assistant and agent swarm, on top of Mario Zechner's pi engine. His detailed reflection covers why pi's minimalist, opinionated design makes it an excellent foundation for extensions that can rewrite the entire agent loop.

Shopify CTO Mikhail Parakhin argues gisting, compressing prompts into "gist tokens", is the most underappreciated LLM technique, sharing production results of 40% lower latency and 15% higher throughput: Gisting treats a prompt like a zip file: train a model to condense your system prompt into a tiny set of reusable "gist" embedding tokens, then cache that compressed form. Parakhin reports it not only speeds up inference but actually improves output quality. The technique was first published by Stanford researchers at NeurIPS 2023, it's well-studied but rarely deployed in production.

OpenRouter debuts stealth model Ox Alpha, community fingerprints it as GLM-5.3: Ox Alpha is a free reasoning model with a 1M-token context that handles text, images, and video. The anonymous provider promises not to train on your data, and forensic analysis identified the underlying architecture as Zhipu's GLM-5.3 through tokenizer and API matching.

Firecrawl launches Developer Index, a specialized search index built for coding agents: Coding agents often search scattered repos, docs, and issues for the right context. This index bundles 70M+ curated primary sources, READMEs, docs, issues, merged PRs, into one API, CLI, and MCP server, refreshed daily. On Firecrawl's own DevDex benchmark it leads with 63.1% recall@10, ahead of Exa and Parallel.

A 125M-parameter transformer autocompletes piano performances on-device: V7 Labs CTO Simon Edwardsson trained a small transformer to predict MIDI note continuations from a few bars of live playing, hitting ~108 notes/sec on an iPhone 15. The free app RollTab lets anyone with a MIDI keyboard try it, think Copilot for piano.

Anthropic brings Claude Mythos 5's cybersecurity capabilities to more defenders through Claude Security: Claude Security (in public beta for Enterprise customers) now scans codebases with Mythos 5 and returns vulnerability findings by CWE, severity, and confidence, with suggested fixes, all billed through standard token usage with no add-on.

Anthropic makes computer use, the Skills API, and the Files API generally available for production agents: Three capabilities that were in beta or preview are now ready for production workloads: computer use lets Claude control desktop software, the Skills API lets agents call custom skills defined in the Claude dashboard, and the Files API enables persistent file storage and retrieval for long-running agent tasks.

Santiago Valdarrama says 80% of AI agent demos don't last a day with real users: The ML engineer and founder who tests AI applications for a living argues that production agents need far more than a good model, error recovery, human handoffs, observability, sandboxed execution, and more, and most current demos fall apart on day one.

Codex users report weekly usage quotas may have been cut in half, as NerfTrack estimates Plus allowance dropped from ~$160 to ~$80 in API value: Heavy Codex users on $20 Plus and $200 Pro plans report their weekly allowance vanishing in one to three days. An investigation by Kingy AI found the claims credible but unproven, the culprit may be GPT-5.6 Sol consuming tokens faster rather than a secret blanket cut. OpenAI Codex lead Thibault Sottiaux has denied any plan-level usage reduction.

Santiago Valdarrama calls Kimi K3 the best open-weight model he's tried: The 2.8T-parameter Moonshot AI model, released in July as the world's first open 3T-class model, delivers frontier-level reasoning and coding performance from open weights. Valdarrama reports 120 tokens/s via Nebius Token Factory with an OpenAI-compatible API.

I'm Becoming AI-Blind: Rafał Cymerys, CTO at Upside, describes a new cognitive phenomenon: when a document carries the unmistakable vocabulary and structure of an AI-generated text, his brain simply refuses to process it. The essay struck a nerve, hundreds of readers reported the same struggle and shared strategies for forcing models to write in plain language instead of their signature verbose style.

I Tracked Down the Hidden Workers Secretly Powering ChatGPT: Award-winning AI journalist Karen Hao traces the global data-worker pipeline that makes ChatGPT possible, from Kenyan labelers earning less than $2 an hour to homeless people in San Francisco hired through the startup Mercor. A stark documentary about the human cost behind the AI boom.

Loop Engineering from First Principles, design AI coding agent loops using control theory, not bash loops: Kyle Mistele (HumanLayer co-founder) argues most AI coding loops are built wrong. This talk walks through a practical sensor-controller-actuator loop using ast-grep as a sensor to spot unmigrated code, a deterministic controller to pick incremental changes, and a CLI agent as the actuator, with human-in-the-loop feedback via PR comments and flow control to prevent PR pile-up.

The Brain Is Just Specialized Agents Talking to Each Other, Dr. Jeff Beck: Duke computational neuroscientist Jeff Beck argues the brain is not a single integrated system but a collection of specialized agents that communicate with each other, and explains why "AGI" is a misleading frame for understanding intelligence. A wide-ranging 47-minute conversation from the Machine Learning Street Talk podcast.

DeepSeek V4 Flash hit 18T tokens per day on OpenCode: DeepSeek V4 Flash surged from 3T to 18T tokens per day on the AI coding platform over two weeks, a single model on one platform briefly exceeding the 10T/day flowing through all 400+ models on OpenRouter, before DeepSeek raised Flash prices 4-5x on August 16. The sixfold spike reveals the enormous latent demand when inference is both cheap and competitive with frontier models.

Poolside AI strikes a $6B licensing deal with Nvidia and raises $1B at a $12B valuation (paywalled): Nvidia locks in non-exclusive access to Poolside's AI models for $6 billion, while pumping another $1 billion into the Paris-based startup co-founded by ex-GitHub CTO Jason Warner. The deal also sees 109 Poolside employees receive offers to join Nvidia, leading many to question whether this is effectively an acqui-hire by another name.