Dispatches of the past four-and-twenty hours from the laboratories of artificial intelligence
Open Weights
DeepSeek’s Flash Gets Faster, Cheaper—and Stays Open
V4.1-Flash-1017 tops every model in its own table on DeepSWE, ships MIT weights, and arrives with a $0.75 search API
By Our Open-Weights Desk
DeepSeek’s own table: the 1017 snapshot (highlighted) against the previous Flash, Kimi K3, GLM 5.3, GPT 5.6-Sol and Claude Opus 5.הטבלה של DeepSeek עצמה: גרסת 1017 (מודגשת) מול ה־Flash הקודם, Kimi K3, GLM 5.3, GPT 5.6-Sol ו־Claude Opus 5.Image: תמונה: @deepseek_ai on X
DeepSeek released DeepSeek-V4.1-Flash-1017 on Saturday, a dated snapshot of its V4.1 Flash model, and called it a Pareto improvement: stronger benchmark scores, faster task completion and shorter outputs than the version before it. It is live on DeepSeek’s website, app and API, and tuned for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy. Harness users should upgrade to v0.2.1-rc.1 or later.
For a “minor update,” DeepSeek’s own table shows big jumps. Terminal-Bench 4.0 climbs from 31.2 to 42.3, ProgramBench from 20.3 to 29.1, and DeepSWE v1.1 reaches 77.0, ahead of every model in the comparison, Claude Opus 5 included at 74.0. Opus 5 still leads on Humanity’s Last Exam, Terminal-Bench and NL2Repo. Note the rivals in the table are GPT-5.6 Sol and Opus 5, not the newest frontier releases.
The weights are on Hugging Face and ModelScope under the MIT licence: 763B parameters and a 1M-token window. Ollama had it on Ollama Cloud within hours, hosted in the US with zero data retention and launchable straight into Claude Code. The model had leaked the night before: aitrackerbot spotted the name in DeepSeek’s API model list at 9:59 PM ET, minutes before it vanished again.
Tibo resets Codex limits again, is “working on two” of a user’s wishes—and lets two dots file his taxes
By Our San Francisco Desk
Day 13 of Thibault Sottiaux’s public diary brought another Codex usage reset, promised “by EOD PST” and confirmed hours later with one word: “Done.” The post drew more than 14,000 likes. When a user posted a wish list—lift Plus’s five-hour limit, dots for Plus, a faster ultrafast mode, double Pro usage, more speed for GPT-6.1 Sol—Tibo replied that he is “working on two of these,” without saying which two.
His weekend anecdote made a better advert than any launch post. His work dot “Burger” coordinated over email with his personal dot “Dorito” to collect documents, then worked with his accountant to file some tax paperwork. It took twenty minutes. “We’re still so early,” he wrote.
Close to Astra, Opus 5.5 and Sol 6.1 on accuracy, but slower—and Alexandr Wang says Meta is on it
Benchmarks Bureau
Independent tester Morgan Linton finished running Meta’s Muse Wonder Full through VulcanBench. Medium effort looks like the sweet spot, he wrote, but time and cost per task decide the matter: Wonder Full is “quite a bit slower” than Opus 5.5, “much slower” than GPT-6.1 Sol and “way slower” than GPT-6 Astra. For now he would pick Opus 5.5 or Sol 6.1 for the same or better accuracy, faster and cheaper.
Alexandr Wang quoted the chart and did not argue: “pretty competitive overall … but definitely slower, which is something we’re working on.” Hours later he cheered a different reading, quoting yacineMTB’s “Muse wonder TPS is kind of crazy” with “it is getting HOT in the KITCHEN.”
A 61 KB Ocean, a Mech Battle Royale and a 42-Bone Crab
The weekend’s Claude Code show-and-tell, plus an official plugin that prints a receipt for your usage
From the Workshop Floor
The Claude Code crowd spent Saturday showing off. One builder made a browser ocean in about 30 minutes that fits in a single 61 KB HTML file, with WebGPU rendering FFT-based waves, whitecaps and a physically based sky. Another used an open-source Blender MCP add-on so one Claude Code session drove both Blender and Unity, producing most of the 3D assets for a fishing game, including a 40-part, 42-bone crab.
Mecha Royale, a multiplayer mech battle royale, runs in the browser and packs a 15 km procedurally generated map into 20 MB; its maker says about 99% of the code and content came from Opus. And for the bookkeepers, Anthropic’s official Receipts plugin prints a receipt of your Claude Code work: sessions, files touched, commits and PRs shipped, and each project’s share of your usage, as HTML you can save to PDF.
OpenCode Discloses a Breach of Saved Provider Keys
A “large-scale distillation attack” hit 1.4% of hosted Console users; some report charges they never made
Security Desk
OpenCode, the open-source coding agent, said a breach overnight Friday tied to “a large scale distillation attack” affected 1.4% of its Console users, the subset who had linked their own AI-provider accounts. The intruder reached a legacy system that was being decommissioned, where API keys saved in workspaces may have been exposed. Some users report unexpected usage charges. OpenCode says it is working with Anthropic, OpenAI and Google on remediation.
Poteto teases a verification skill for pstack as Addy Osmani argues agents must check their own output
Developer Notes
“Working on a new verification skill for pstack,” Lauren Tan (poteto) posted late Saturday, and 1,400 people liked a one-line teaser. She also reposted Addy Osmani’s case: give your agent a way to verify its work; with that in place, Claude Code wrote more than 90% of its own code, and when software arrives faster than humans can review it, the bottleneck shifts to proving it works. Her earlier question, “is chat the final form of agent UI?”, drew about 1,200 replies.
Meta Edits Its Terms, FLUX 3 Klein & Android Skills for Agents
Training clause dropped; a 5.6B open image model; 30 official skills from Google; OpenAI’s weekend-project gallery
Roundup from the Wire
Meta removed language from its commercial terms that let it use content from a business account’s domains, web pages or URLs to train its products, after legal experts raised scraping and AI-training concerns. Black Forest Labs released FLUX 3 Image Klein, a 5.6-billion-parameter image model with downloadable weights, text-to-image, editing with up to 10 reference images and bounding-box layout control; the company says an INT8 version runs on consumer hardware.
Google’s Android team published 30 official skills that teach coding agents to build Android apps its way, aimed at the spots where Google’s own tests found models struggle, such as moving old XML layouts to Jetpack Compose and upgrading Play Billing. And OpenAI Devs, “looking for a weekend project?”, posted a gallery of community builds with guides to replicate each one.
Open-source AI projects worth an afternoonפרויקטי AI בקוד פתוח ששווים אחר צהריים
Model Switchboardמרכזיית מודלים
magpie: Any Model for Any Coding Agent, From the Menu Barmagpie: כל מודל לכל סוכן קוד, משורת התפריטים
Claude Code on Kimi, Codex on DeepSeek: one local gateway that also moves on when a quota runs out.Claude Code על Kimi, Codex על DeepSeek: שער מקומי אחד, שגם עובר הלאה כשהמכסה נגמרת.
magpie’s Agents page: one switch and one model picker per coding agent.עמוד הסוכנים של magpie: מתג אחד ובורר מודל אחד לכל סוכן קוד.Image: תמונה: yetone/magpie on GitHubGame Moddingמודים למשחקים
universal-modder: Point Your Coding Agent at Any PC Gameuniversal-modder: לכוון את סוכן הקוד לכל משחק מחשב
Skills, tools and a fal MCP server that let Claude Code, Codex and friends mod games you own, from recon to showcase video.כישורים, כלים ושרת MCP של fal שמאפשרים ל־Claude Code, Codex וחבריהם לבנות מודים למשחקים שלכם, מסיור ראשוני ועד סרטון תצוגה.
The repo’s teaser reel of mods made with universal-modder.סרטון הטיזר של המאגר: מודים שנבנו עם universal-modder.Image: תמונה: rehan-remade/universal-modder on GitHubNews Engineמנוע חדשות
AIHOT: A News Site That Finds Its Own Stories and Writes Its Own DailyAIHOT: אתר חדשות שמוצא את הסיפורים שלו וכותב לעצמו עיתון יומי
The open engine behind a Chinese AI-news site: two-pass LLM scoring, event clustering and a morning edition.המנוע הפתוח מאחורי אתר חדשות בינה מלאכותית סיני: דירוג כפול במודל שפה, איחוד אירועים ומהדורת בוקר.
AIHOT’s daily picks feed and its about page, where sources flow into one daily report.פיד הבחירות היומי של AIHOT ועמוד האודות, שבו המקורות זורמים לדוח יומי אחד.Image: תמונה: KKKKhazix/AIHOT on GitHub
Wren & Robin read you this morning's dispatches, with banter.
Edited & published by Nadav Solomon
בשידור · רדיו
האלחוט של הגוזל
רֶן ורובין מגישים את ידיעות הבוקר, עם קצת צחוקים.
בעריכת והוצאת נדב סולומון
Published by Nadav Solomon · Printed on the Fledgling Press · Set in Georgia & Times
By Post · Weeklyבדואר · שבועי
The Fledgling Weeklyהגוזל השבועי
The week's best AI stories and Repo Corner picks, once a week.מיטב ידיעות ה-AI של השבוע ובחירות פינת הריפו, פעם בשבוע.
Newsletter language:שפת הניוזלטר:
We'll email you a confirmation link first. Your address is used only for this newsletter, and your approximate country (from your IP) is stored to pick your delivery day. Never shared; every issue has a one-click unsubscribe.קודם נשלח לך קישור לאישור. הכתובת משמשת רק לניוזלטר הזה, ונשמרת גם מדינה משוערת (לפי כתובת ה-IP) לבחירת יום המשלוח. לא מועבר לאף אחד, ובכל גיליון יש הסרה בלחיצה אחת.
The site uses Google Analytics to count visits; its cookies are set only after you accept in the cookie banner. Privacy policy.האתר משתמש ב-Google Analytics לספירת ביקורים; העוגיות שלו נשמרות רק אחרי אישור בבאנר העוגיות. מדיניות פרטיות.
