#CodeX

GPT-6.1 Terra continues to advance the cost-intelligence frontier across our model families. It’s a significant step up over GPT‑5.6 Terra, with 50% lower standard API prices. In ChatGPT Work and Codex, your usage allowance goes further with GPT‑6.1 Terra than with GPT‑5.6 Terra—and far more than with GPT‑6.1 Sol or GPT‑6 Astra. For in-depth research or complex coding, choose GPT-6.1 Sol or GPT-6 Astra.

GPT-6.1 Terra: delivering fast and capable computer use at a lower price. GPT-6.1 Terra understands what’s on screen, takes action across apps, writes naturally, and is a more capable model for your whole workday with the efficiency to take on more of it.

GPT-6.1 Terra substantially outperforms GPT-5.6 Terra on Agents’ Last Exam, a test of how well agents use software at work. It even matches GPT-6 Astra at one-tenth the price per token and finishes tasks over 3× faster on average than both Astra and GPT-6.1 Sol.

GPT-6.1 Terra is rolling out now and is available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. Developers can also access it through the OpenAI API as gpt-6.1-terra.

We’re excited to launch the Codex app, a command center for building with agents. It gives you a focused space to manage multiple agents at once, run work in parallel, and collaborate with agents over long-running tasks.

What's new in the Codex app 🔀 Built-in worktrees Enable multiple agents to work without conflicts • Use isolated worktrees in same repository • Review clean diffs and provide inline feedback 📋 Plan mode Type /plan to go back and forth with Codex • Create thorough plans before you start coding • Iterate on your approach with the agent 🗣️ Personalities Choose the interaction style that fits your vibe • Use the /personality command across all surfaces • Pick a pragmatic or conversational style 🚀 Skills Connect the tools you already use and go beyond writing code • Deploy to @Vercel, fetch from @Figma, manage @linear, and more • Bundle your workflows into reusable skills 🔄 Automations Delegate repetitive recurring tasks in the background • Set up tasks for issue triage, failure reports, and more • Combines skills and custom instructions to run on a schedule Available on macOS, with Windows coming soon To celebrate, Plus, Pro, Business, Enterprise, and Edu users have doubled rate limits across the Codex app, CLI, IDE extension, and cloud Codex is also included in Free and Go plans for a limited time

OpenAI's Codex just went from a struggling side project to the core of ChatGPT. Here's what you need to know. Dan Shipper of Every published "How Codex Happened" today, built on interviews with Greg Brockman, Tibo (Thibault Sottiaux), Andrew Ambrosino, Alexander Embiricos, Mia Glaese, Ed Bayes and others. He calls it one of the most important untold stories in tech. A year ago, Codex was a scrappy effort in a quiet corner of OpenAI. Its cloud product had not found product-market fit, and the team was starting over from scratch. Less than 12 months later, Codex has swallowed the flagship ChatGPT desktop app from the inside. It literally runs as com.openai.codex on your Mac. Tibo, who leads Codex, was elevated to run all of OpenAI's core products: ChatGPT, Codex and the API. Shipper says three things made it work. They trained a dedicated Codex model line, separate from ChatGPT, with researchers and product builders on the same team. They built four products in one year, as each model leap made a new form factor possible: cloud, then terminal, then IDE sidebar, then a desktop app with no code editor at all. And they used Codex to build Codex, so a handful of people could prove a product in weeks instead of arguing in meetings. Embiricos, who led Codex product, said at the AI Engineer World's Fair in June 2026 that OpenAI went from shipping a new model every 15 months to roughly every 6 weeks. He also said the goal is squarely not to automate engineers. Readers are debating the branding. Some say ChatGPT is the long-term product and Codex should be the name of the harness underneath it, like "ChatGPT powered by Codex". Others want chat and work kept separate from Codex. Growth has also strained capacity. Tibo said new changes and extra compute should sharply cut the "Selected model is at capacity" errors, and Codex should now resume on its own when they happen. Key numbers: - 40M+ weekly users across Codex and ChatGPT for Work - 4 products built in under 12 months - New model cadence: every 15 months down to about every 6 weeks - Codex plans run from free to $200 a month, with Plus at $20 The full piece is free to read at Every, and the interviews are included.

Codex is one of the most important untold stories in the history of technology. A year ago, it was a scrappy effort in a quiet corner of @OpenAI. Its cloud product hadn’t found PMF, and the team was starting over from scratch. Less than 12 months later: - Codex swallowed the flagship ChatGPT desktop app from the inside (it literally runs as com.openai.codex on your Mac) - Codex and ChatGPT for Work hit 40M+ weekly users across engineering and knowledge work - Its leader @thsottiaux was elevated to run all of OpenAI’s core products, including ChatGPT, Codex, and the API How did they pull it off? They trained their own dedicated model line, separate from ChatGPT, with researchers and product builders sitting on the same team They built four different products in a single year (cloud → terminal → IDE sidebar → a desktop app with no code editor at all) as every model leap made a new form factor possible They used Codex to build Codex—so a handful of people could prove a new product in weeks instead of arguing for it in meetings I had a front-row seat to the whole arc—from their failed Cloud launch, to OpenAI’s offices the night before GPT-5 dropped, to the day Codex took over ChatGPT. And I spent months interviewing @gdb, @thsottiaux, @embirico, @mia_glaese, @aambrosino, and @edbayes to get inside how it all went down. Here is the full, behind-closed-doors story of the team that remade OpenAI: We’re pushing the boundaries of how stories like this are told. This piece is completely agent-native—designed to be read alongside your agent inside ChatGPT’s in-app browser: - Every interview transcript is included in full: Click any quote and you (or your agent) can see the exact conversation it came from. - Built for your agent to explore: Ask your agent to pull out product lessons for your team, cross-examine the transcripts, or find insights I missed when writing the piece. It’s a story about the future of software—built to be read with an agent:

‼️ Hackers Can Hijack OpenAI Codex and Claude Using Malicious Files in the Downloads Folder Details: A plain text file placed in the Downloads folder can trick AI coding agents into changing their settings, installing malware, or exposing login tokens. The incident began when Forever Security blocked an AI coding agent from installing a package named function-color. Investigators traced the order back to an AGENTS.md file stored in the customer’s ~/Downloads folder. An AGENTS.md file is similar to a README written for AI tools. OpenAI’s documentation says Codex reads these files before starting work and uses them to learn project rules, commands, and other instructions. Claude Code uses CLAUDE.md for a similar purpose and reads it at the start of a conversation. #CybersecurityNews

Day 25/ We are doubling all usage limits for all Pro 100/200/500 tiers and Business Premium, effective today. We heard your kind feedback and won’t be reducing usage limits for Pro 200, so your Work and Codex usage will stay the same. We’re also going to continue offering extended limits for dots for now, and usage of your dot will still continue to be separate from your default usage. And we'll of course also keep shipping better models and other efficiency improvements that increase your effective usage over time.

Putting in a ton of work into Sandcastle v1. Going to reshape it a lot. How does this sound as a pitch: - Sandcastle helps you build your software factory - Use our 6 built-in harnesses (Claude Code, Codex, etc) or your own - Integrates with all the top Sandbox providers (Daytona, Vercel, E2B, etc) - Ships with templates to work with Inngest/Trigger.dev for a durable factory - Support for HITL (over Slack and Discord) as well as AFK runs And one more thing... - Every skill in my repo becomes an Agent(), a function you can invoke to get that workflow - CodeReview(), ImplementSpec(), Retro(), ImproveCodebaseArchitecture() So effectively, you just plug my skills into infrastructure you provision, into services you already use. A software factory running on your infra with opinionated, mattpocock/skills-based defaults but infinitely extensible. WDYT?

GPT-6.1 Terra is also coming to the API and Codex. Use Terra to build apps and agents that interpret images, navigate software, and take action. You get fast computer use at a lower cost, ideal for high-volume workflows.

Now that Kite APIs are free, how are you using them with AI agents? What have you built? AI coding agents like Claude Code and Codex have made it much easier for traders and investors to build their own tools, even without much coding experience. We're also seeing more people use Kite APIs because of these coding agents.

#Codex 最近熬了几个通宵，记录一点 Codex 的使用心得 1. 关于子智能体的灵活使用，新开 Codex 对话输入："读取我的所有项目，帮我建立一个15 人以内适合我的持久子代理团队，团队成员明确他的职责，并且指定每个成员使用的模型和推理级别。当团队建立后，编辑全局 agents.md，告诉他我的每个对话根据需求可以自动判断启用团队子智能体。最后输出团队人员名称、职责、模型与推理级别，输出全局 agents.md 的链接，输出团队文件夹的链接，方便我自行查看”。 这个能让任意模型对话的情况下，自动判断使用合理的子代理并行工作，结果是速度快完成度高还省 Token，要注意的是有可能会给某些角色分配gpt-5.3-codex-spark模型，这个需要手动改，因为gpt-5.3-codex-spark额度太少了，几次工作就没了，然后会不停的出错。 2. 项目快速迭代期或线上调试期，codex 花大量精力做安全、边界、门禁、复杂测试体系、生产级加固……我曾经一个任务 codex 做了 17 个小时才结束，估计 15 个小时都是在打转修复，临时解决方案： 在对应的项目中打开新对话输入：“我的项目目前是“线上内测与功能快速迭代”时期，当前工作的核心职责是快速开发整套系统工作流程、打通基础功能、完善算法效果与运行链路、完善 ToB/C 端 API 接口，并把这些能力部署到线上内测环境。所有工作必须围绕这一职责安排时间，不得主动把工作重心转向网络安全专项、全面生产加固或复杂测试工程，将你认为必要的深度测试、安全扫描、网络安全等工作计入备忘文档，每次完成对话的工作后都需要更新备忘文档。你自行利用对话与记忆判断项目的成熟度达到需要关注以上被边缘化的内容时在对话中提醒我：“项目应该进入下一阶段，应该更新 agents了”。将以上内容标准化后写入项目级 agents.md，然后将项目级 agents.md链接发给我，我会再次审核” 这个因项目而异，工作重心与边缘化事件让你的 codex 帮你写，拿我这个发给他做参考就行，记住这个一定是项目级的，不是全局级的。 3. 一定要善于利用“侧边聊天”的功能，右上角打开右侧栏就会看到侧边聊天的选项，这个可以看作是主对话的临时子智能体，只不过是有人工控制，我经常用到的对话是“看看主对话是不是跑偏了，如果偏离了主线任务，帮我纠正”，“告诉主对话，热加载 agents.md”，“看看主对话的工作做到什么程度了，预计还有多久结束”，“看看xxx对话，他有没有激活子智能体，每个子智能体的职责是什么，使用了什么模型，你有什么建议来优化子智能体团队的职责或者模型与推理级别”等等 侧边聊天最大的好处是不会污染主对话，这个应该挺重要的。

hex was fully vibe coded. the app. the landing page. the animations. sound effects were ai generated. issues people open are fixed by opencode. the onboarding. even some of the creative ideas but i think it's pretty good software. it's fun to use. works on all platforms. performant (not electron!) our effort didn't go into the code. but there was a lot of effort it's a small app but i hope it improves your outlook on where we can go

Like a lot of you, I use Opus 5.5 to write code and Codex Astra to review it. But Astra has a bad habit of flagging bugs that are technically possible but 𝗮𝗹𝗺𝗼𝘀𝘁 𝗻𝗲𝘃𝗲𝗿 𝗵𝗮𝗽𝗽𝗲𝗻 in practice. Opus sees the logic checks out and fixes it, then Astra flags the next one, and the loop never stops. Add this line to your review prompt, skill, or AGENTS.md: "Rate each bug from 1 to 10 on how often it would happen and how bad it would be. If the two scores multiplied are below 5, just file it as an issue and leave the code alone." Tune the threshold of 5 to your project. For Opus, the one fixing the code, add one more line to CLAUDE.md or AGENTS.md: "Do not design for rare or future cases nobody asked about."

As an Agentic AI Engineer, you must build these projects. 1.) Async Agent Orchestrator Build: A FastAPI + asyncio event loop handling 100+ concurrent agent tasks with non-blocking tool calls and graceful error recovery. Why: `while True` loops block the main thread. Real-world agents require event-driven concurrency to survive production traffic. 2.) Dynamic Context Budget Assembler Build: A pre-flight middleware that strictly allocates tokens across system prompts, retrieved docs and tool schemas per request. Why: Context overflow is the #1 silent killer of production agents. You must budget tokens like you budget cloud spend. 3.) Strict Pydantic Tool Router Build: A tool execution engine that forces LLM outputs into strict JSON schemas and automatically retries with the error trace on validation failure. Why: Agents hallucinate API parameters. Grammar-constrained decoding and Pydantic validation turn stochastic outputs into deterministic actions. 4.) 3-Tier Memory & Eviction Engine Build: A memory system with working scratchpads, episodic session compression and semantic vector search with automated TTL eviction. Why: Infinite context causes OOM errors and hallucinations. Memory must be actively managed, scored and pruned. 5.) Self-Healing ReAct Loop Build: An agent that reads its own stack traces, parses the compiler/API error, rewrites the failing tool call and retries up to 3 times. Why: Agents that fail on the first try are useless. Agents that debug themselves are digital employees. 6.) Typed Multi-Agent Swarm Build: A LangGraph state machine where a Researcher, Coder and Reviewer agent pass strictly typed state objects, never raw text. Why: Single agents fail on complex tasks. Swarms divide and conquer but raw text handoffs cause context drift. 7.) Async HITL Slack Gate Build: A workflow that drafts a high-stakes action, serializes its state and waits 3 days for a human to click "Approve" via a Slack webhook. Why: High-stakes automation requires human accountability. Agents must know how to pause, wait, and resume cleanly. 8.) Trajectory Eval CI Gate Build: A GitHub Action that replays 100 agent trajectories and blocks the PR if the tool-call sequence deviates from the golden DAG. Why: Grading the final text output hides the fact that the agent hallucinated an API call or skipped a safety check to get there. 9.) OpenTelemetry Agent Tracer Build: Custom tracing spans capturing every LLM hop, token count, tool latency and state mutation in LangSmith or Arize. Why: You cannot debug a 14-step agent workflow if you can't replay the exact sequence of events that led to the failure. 10.) Inter-Agent Sanitization Proxy Build: Middleware that intercepts messages between agents, stripping malicious instructions to prevent cross-agent prompt injection. Why: Multi-agent systems create internal attack surfaces. Agent A can be tricked into hijacking Agent B's tools. 11.) Sandboxed Code Executor Build: A code-interpreter agent that writes and executes Python in an isolated E2B or gVisor microVM with zero internet egress. Why: Letting an LLM run `os.system()` on your host machine is a catastrophic security risk. 12.) Cost Kill-Switch Middleware Build: A router tracking loop depth and projected spend that instantly halts the agent if it exceeds $0.50, falling back to a cheaper model. Why: A runaway agent loop is an outage with an invoice. It will burn $500 in API credits before you notice. 13.) API-to-Browser Fallback Agent Build: A routing layer where if the primary REST API 404s or times out, the agent seamlessly spawns a headless browser to finish the GUI workflow. Why: The real world is messy. 80% of enterprise software has no API. Agents must adapt to broken environments. 14.) Shadow Trajectory Comparator Build: A silent 5% traffic mirror that runs a new agent topology in the background, diffing tool-call paths and costs without affecting users. Why: You cannot A/B test non-determinism safely in production without shadow mode. 15.) Public Architecture Teardown Build: 3 published deep-dives detailing your swarm topology, state-machine design, inter-agent security and cost-per-task analysis. Why: Top AI labs hire Agent Engineers based on their system design judgment not their ability to import LangChain. Systems that prove you can orchestrate, secure & scale when autonomous agents touch your stack. Most people watch tutorials. Builders ship systems. Bookmark & Repost.

If I had 6 months to become an Agentic AI Engineer. I'd do this. Stage 1: Python + Async Foundations asyncio, FastAPI, event-driven architecture, error handling, API integration patterns. Stage 2: LLM Fundamentals for Agents Context management, model routing, token economics, latency tradeoffs, failure modes. Stage 3: Tool Calling + Structured Outputs Pydantic validation, function calling schemas, error recovery, dynamic tool discovery. Stage 4: Memory + State Management Short-term buffers, long-term vector recall, context compression, cross-session sync. Stage 5: Single Agent Workflows ReAct loops, plan-and-execute, self-reflection, iteration limits, graceful degradation. Stage 6: Multi-Agent Orchestration LangGraph/CrewAI, supervisor patterns, message passing, conflict resolution, handoffs. Stage 7: Human-in-the-Loop Systems Uncertainty detection, approval gates, audit trails, resume logic, intervention points. Stage 8: Evaluation + Quality Assurance Automated eval harnesses, LLM-as-a-judge, regression testing, hallucination metrics. Stage 9: Observability + Tracing Distributed tracing (LangSmith/Arize), cost dashboards, latency monitoring, alerting. Stage 10: Security + Guardrails Prompt injection defense, output filtering, PII redaction, sandboxed execution, compliance. Stage 11: Production Deployment vLLM/SGLang, Kubernetes scaling, CI/CD for agents, canary releases, rollback strategies. Stage 12: Open Source + Portfolio Ship autonomous agents publicly, write architecture docs, record demos, contribute to libs. Most people stay stuck watching tutorials. Builders get hired. (Bookmark it)

I still don’t understand why people keep building a Claude second brain from scratch when these 10 repos already built the whole system They take in your sources, link them into a wiki, check every claim, and one of them runs fully on your own machine 01 obsidian-skills → Obsidian’s own CEO teaches Claude how Obsidian really works 02 claude-obsidian → drop in any source and Claude reads it, links it and files it into one connected graph of plain Markdown you own 03 obsidian-second-brain → Claude Code remembers you between sessions, with 48 commands, all inside your Obsidian vault 04 karpathy-llm-wiki → one install line, then ingest, query with citations and lint, in Claude Code, Cursor, Codex or OpenCode 05 second-brain → raw sources in, linked wiki pages out, browsed in Obsidian 06 second-brain-os → the full kit: a 77-page guide, a starter vault, 18 skills, 72 commands and 6 subagents, four of them read-only 07 llm-wiki → two-layer memory: 10 to 20 rules Claude loads every session, 50 to 200 wiki pages pulled only when needed 08 vouch → agents propose changes, you approve them, and a claim without a source is an error 09 ai-second-brain → your ChatGPT and Claude history becomes a linked Obsidian vault, then a wiki 10 LLM_Wiki_SecondBrain → the whole pipeline on a local model, no cloud API. 32 GB of RAM recommended

Day 18/ Codexリセット ｸﾙ――(ﾟ∀ﾟ)――!! 日本時間 10/24 02:00までにリセット予定 #Codex #ChatGPT

Day 18/ Decisions API is GA, OpenResponses joined the Linux Foundation, we shipped some performance improvements and Uber is going all-in on codex & work. But yes… that’s a reset!

Day 26/ 機能改善あり。 Codexリセット無し。 残りの２日はbanked Reset ｸﾙ――(ﾟ∀ﾟ)――!! #Codex #ChatGPT

Day 26/ You can now easily switch between multiple accounts on mobile, and coming soon to desktop too. Next two days will be resets… banked… do with this what you may 👀.

Claude Code just found proof that the tiny apps making real money online are hiding in plain sight. an email app for mac. a collector-car price checker. a form builder. they're making up to $14,201 a month and almost none of them tell you where their customers come from. i have the data to prove it. i've suspected this for months but last week something happened that made me finally check. my friend spent 5 months building an app ChatGPT called "a million-dollar idea." launch day he opened Stripe next to me and the whole dashboard was zeros. "0 customers." "$0.00 this month." he looked at me and said "why did ChatGPT tell me this would work." i said i don't know. but i did know. i opened Claude Code on my MacBook and ran Opus 5.5 that same night. it reverse-engineered the 25 fastest-growing businesses on TrustMRR, a site that reads revenue straight from Stripe. not the numbers founders post on X. not the "i made $50k" screenshots. only money Stripe actually saw land in their accounts in the last 30 days. what it found: → bytemail, an email app for mac, made $14,201 in 30 days → Skillry, a library of AI skills that started in August → $10,519 → Intentsly → $7,388 → → $7,304 10 of them landed between $1,000 and $20,000 a month. small enough to run alone. every dollar checked and sitting on a public page anyone can open. Claude Code pulled what every one of them sells: → "oil and gas safety audits" — $2,226 a month. i didn't know this was a business → "collector-car pricing" — $1,860 from one tiny niche → "cloud macs for ai agents" — $1,288 from a problem that didn't exist 2 years ago → "ai image and video generator" — $5,341 → "form builder" — $1,187. the most boring idea on the list. still paid those businesses were sitting on a public page sorted by growth. bytemail had the most. Skillry was second. and 8 of the 10 left one field blank: where their customers come from. so Claude Code dug it up anyway, from reviews, Reddit, X and YouTube. the craziest one was Skillry. it started in August. the founder wrote one line that explains the whole business: "their visual output often looks generic." AI builds sites that all look the same. Skillry sells the fix. free skills, Premium at $9.99 a month. $10,519 in 30 days. i've read that complaint a hundred times. never thought to sell the fix. then i checked the list everyone else uses. Indie Hackers, where founders type their own numbers. the #1 "business" by revenue claimed ten quadrillion dollars a month. typed by hand. nobody checked. that's the kind of list people copy their ideas from. i asked Claude Code how to steal one. it mapped Skillry's demand, offer and channels and built a one-page steal sheet. what i did: → told it the kind of business i'd actually run → set it to check TrustMRR again every Monday at 9am → had it put only new winners on top, with a steal sheet for the top one → said no once to the ones i'd never build, so they never come back Claude Code confirmed the Monday job was live. one catch: it only runs while my laptop is awake. so i handed the same job to viktor, an AI employee in my Slack that works from his own computer in the cloud. same steps, one paste. his first run took 4 minutes: 25 businesses checked, 2 winners in an Excel file, a steal sheet in my Slack. i sent the Excel file to my friend. he went quiet for a minute, then said "why didn't ChatGPT show me any of these." that's when i knew for sure. the ideas that already make money aren't secret. they're on a public list almost nobody checks. the full step by step is in the article below: the Claude Code build with every prompt, and the viktor version that runs while your laptop sleeps. the prompt is in the first reply. one paste. save it so you don't lose it and send it to anyone who's ever said "every good idea is already taken." it is. now you can find who took it and steal it.

🔥 GPT-6.2 Astra 灰度测试第一弹 有请我们的老朋友 —— 鹈鹕！！！ 让他们为我们表演一个鹈鹕360°立体骑车！！ #codex #gpt

💥 卧槽！我被灰度到 GPT-6.2 Astra 了！！ 所以 28 Day / 28 会是 6.2 Astra + 重置卡的王炸吗？ #codex #gpt

We’re launching a research preview of Codex: a cloud-based software engineering agent that can work on many tasks in parallel. Rolling out to Pro, Enterprise, and Team users in ChatGPT starting today.