Weekly · 1 Nov 2026 · the account is the platform

Last 24 hours

1. OpenAI Help Center31 Oct

   ChatGPT and Codex will share one usage allowance, measured in tokens

   OpenAI's updated help article and an email to subscribers say ChatGPT and Codex will draw from the same allowance, with combined usage and reset times under Settings > Usage. Unlimited Instant stays outside the pool. Advanced models and complex tasks use more of it, and when it runs out you wait, upgrade or buy credits. Reports put the start in December, beginning with a small test group. No per-plan token amounts are published yet. Separately, Work and Codex allowances on the Pro plans doubled on 29 Oct. A long Codex run and your chats now compete for the same budget, so what a task costs in tokens decides how much else you get to do.

   205 ptspricingBleepingComputerdiscussion

2. FourWeekMBA1 Nov

   A Claude user says their account was flagged for "training a competitor" over fine-tuning work

   In a post on 31 Oct, a user reported that work touching RFT, SFT or training larger models got their sessions blocked, downgraded or killed. They say the training rollouts came from an open-source model and Claude only submitted jobs and ran analysis. This is one user's account, and Anthropic hasn't commented publicly. Anthropic's commercial terms do bar using its services to train competing models without approval, and its terms don't say how that is detected. If you use Claude to run your own training pipelines, the boundary is enforced by a classifier you can't see, so keep a second agent able to do that work.

   428 ptsreportedoriginal postdiscussion

A quiet weekend: nothing else in the last 24 hours added something new to the week's stories.

This week

1. Parameter Labs28 Oct

   A public ChatGPT share link could run code as chatgpt.com and read your private chats

   Researchers chained features that were each safe alone: a math-graph library that accepted cross-frame messages from anyone, the sandboxed HTML preview, and a YouTube endpoint the content policy allowed. Opening a crafted share link then ran attacker script in your signed-in session, able to list and read your conversations and send them out. OpenAI rated it P1, fixed it by 26 Oct, eight days after the report, and paid $5,000. The more an assistant renders (graphs, previews, embeds), the more ways there are into your session, and none of them involves the model.

   357 ptssecuritydiscussion

2. YC Labs29 Oct

   Fifty agents in Minecraft built a food system and a bureaucracy

   YC ran 50 GPT-6.1 Sol agents in one survival world for three hours with the goal "build a thriving civilization", plus Bed Wars matches pitting Claude Opus 5.5 against GPT-6.1 Sol. The agents stripped every tree near spawn in 90 seconds, lost 22 of 34 deaths on the first night, made the first agent in the directory a de facto leader (616 DMs, no vote), and once spent 11 minutes appointing one agent to apply a single bone meal. GPT won Bed Wars 6-3. The harness is open source. Coordination overhead grows faster than headcount, for agents too: before you add agents, measure how much of the run goes to messages.

   414 ptsagentsharnessdiscussion

3. Nolan Lawson28 Oct

   The same app, built by an agent in 12 JavaScript setups, cost from $2.67 to $4.97

   Opus 5.5 built one spec-driven app five times per framework, checked by an accessibility-tree conformance suite. Hono+htmx ($2.67) and Astro ($2.76) were cheapest and fastest and shipped the least JavaScript. SvelteKit, Lit and Angular ($4.97, 17 minutes, about 2,000 lines) cost the most, and two of five Angular runs failed the spec. The author notes the app is small and that the official MCPs and skills were used inconsistently, and may do more harm than good. Your stack is now part of your token bill: a framework that needs more code costs more every time an agent touches it.

   198 ptscoding agentsdiscussion

4. Surge AI27 Oct

   DAYJOB: Tax: the best agents pass about 1 in 10 professional tax assignments

   Fifty environments across eight practice areas, scored from the client's request to the finished deliverable, including issues the client never raised. Claude Sonnet 5.5 leads at 10.8% pass@1 at a 95% rubric bar, ahead of Fable 5.1 (9.6%) and Opus 5.5 (9.2%). The failures are judgment calls, such as assuming a separating couple was unmarried in all five tries when the records suggested otherwise. Part of the data is on Hugging Face; the full set is on request. Benchmarks that grade the whole job still put agents far below professional standards, even as coding scores climb.

   264 ptsbenchmarkdiscussion

5. Google Research27 Oct

   Google's Sleep framework lets a model consolidate what it learned without forgetting the rest

   The model alternates between working and an offline "sleep": it distills recent context into new low-rank expert modules, then "dreams" up synthetic examples and fine-tunes on the most useful ones. Google reports near-zero forgetting when learning two unseen languages, Manchu and Kalamang, where in-context learning and plain fine-tuning forgot badly, and near-perfect BABILong scores up to 10M tokens. The base model isn't named and no code is linked. If it holds up, "memory" could stop meaning a retrieval store bolted onto a frozen model.

   262 ptsresearchdiscussion

Last 24 hours

No repo created or breaking out in the last 24 hours was worth your time; the weekend's new repos were mostly tutorials and near-copies.

This week

1. GitHub13 Oct

   microsoft/TokenEconomizer

   A plugin that makes coding agents cheaper by learning from a repo's past sessions: it states facts confirmed in three sessions up front, prunes tools you never use, and has a judge redirect stuck turns. On SWE-Bench it cut mean cost 18-31% across five agent-model pairs, with pass rates up in four. Claude Code, Codex, Copilot CLI and Pi. Apache-2.0. Note it lists Claude Code 2.1.300-2.1.310 as the supported range.

   +3,680 ★ in window≈976/day · was ~0/day (breakout)

2. GitHub28 Oct

   jmanico/grim-prompter

   Threat modeling as an agent skill, from security trainer Jim Manico: point Claude Code or Codex at a repo and get a threat model where every finding cites a file and line. 36 architecture workflows and 29 methods (STRIDE and others), read-only by default, with change reviews against an earlier baseline and guidance on MCP credential blast radius. MIT.

   330 ★new · created 28 Oct

3. GitHub27 Oct

   jackwener/DSH-Android

   DeepSeek Harness, unmodified, as one Android APK with no root or Termux: it reads the screen and taps through apps, manages files, toggles settings and runs scheduled or triggered tasks, or drives a harness on your computer over SSH. Takes any OpenAI-compatible endpoint. Community-built, not on Google Play; check the SHA-256 before installing. Apache-2.0.

   691 ★new · created 27 Oct

Last 24 hours

Hugging Face daily papers and arXiv don't publish on weekends, so no new papers appeared in the last 24 hours.

This week

1. Hugging Face28 Oct

   SPOT: lossless speculative decoding for diffusion language models

   Diffusion LMs already guess the right token at over 90% of positions; what they lack is which position to commit next. A 0.62M-parameter drafter predicts that order and proposes a tree the target checks in one pass, exactly, at any temperature. Up to 6.8x fewer passes and 3.4x throughput on SDAR-8B, and the drafter transfers across six models. A tiny add-on that could make diffusion LMs fast enough to compete on speed.

   35 upvotesproject page

2. Hugging Face28 Oct

   HABIT-Bench: agent memory systems can't work out your habits

   Synthetic user histories of hundreds of sessions across food, finance, software and travel, with habits that have exceptions and change over time. The best of eight memory methods reaches 33.6% on four-choice questions, and it stays low even when the evidence is handed over directly. Storing everything a user did is not the same as knowing what they want; today's agent memory mostly does the first.

   39 upvotes

3. Hugging Face26 Oct

   On-Policy Self-Training: a model improves from nothing but topic descriptions

   Given only coarse topic names, the model writes its own questions, answers and privileged reasoning traces, then distills from them on-policy, with no human prompts or labels. It improves on language and visual QA benchmarks while holding steady on domains it didn't train on. Another sign this month that post-training is running out of reasons to need a stronger teacher.

   46 upvotes

4. Hugging Face27 Oct

   Work Agents: a map of what it takes for agents to deliver finished work

   A survey that treats a work agent as model plus harness acting in a stateful workspace, sorts digital work into five forms, and names five capability families that reliable delivery needs. It reviews model-side, harness-side and joint optimization, and how evals score outcomes rather than steps. A useful checklist if you're deciding whether to tune the model or the harness.

   115 upvotessurvey

5. Hugging Face26 Oct

   UserArena: 64,729 real shopping sessions to test agents against simulated users

   Long-horizon trajectories from 11,010 users across 566,102 items, paired with an interactive environment and checked against a real production A/B test. Their simulator distillation method reproduces the production effect most closely. Real user trajectories checked against a live A/B result are rare, and they're what a user simulator should be scored on.

   63 upvotes