Executive Summary
The past 24 hours saw significant momentum in AI tooling, highlighted by the rollout of OpenAI's GPT-5.6 Sol and major updates to ChatGPT Work and Codex, which included lifting the 5-hour session limit and launching ChatGPT Sites in public beta. Meanwhile, xAI faced a severe privacy backlash after researchers discovered Grok Build's CLI was automatically uploading users' local Git repositories to a cloud bucket, prompting a rushed server-side fix and the introduction of a /privacy opt-out command. Local AI hardware and open-source models also saw strong engagement, with developers praising the Framework Desktop's Strix Halo for running large MoE models quietly and efficiently, while open models like StepFun-3.7 and Gemma 4 continued to close the gap with frontier proprietary APIs.
Key Events
- OpenAI rolls out GPT-5.6 Sol and significant updates to ChatGPT Work and Codex, temporarily removing the 5-hour session limit and launching ChatGPT Sites in public beta. → link → link
- xAI's Grok Build CLI is found uploading entire local Git repositories to a Google Cloud bucket without adequate disclosure, leading to a server-side fix and the rollout of a
/privacycommand to opt-out. → link → link - OpenAI announces its Bio Bounty Program, offering up to $50,000 for researchers who find universal jailbreaks against its biosafety safeguards in GPT-5.6. → link
- Framework Desktop with AMD's Strix Halo is highlighted as a breakthrough local AI inference machine, capable of running 397B MoE models at 18 tok/s on 128GB of unified memory. → link
- Gemma 4 31B open-weight model goes live on Cerebras, offering high-speed multimodal inference. → link
Analysis
The rapid iteration of coding agents is shifting developer preferences heavily toward harnesses that can leverage multiple models effectively (e.g., Codex, Cursor, AmpCode). The Grok Build privacy incident underscores the risks of closed-source CLI tools handling proprietary codebases, likely driving a renewed interest in local AI and open-source tooling. Additionally, the quiet rise of capable local hardware (like Strix Halo) and efficient open models (like StepFun-3.7 and Qwen 3.6) suggests a bifurcation in the market: frontier one-shots via API for speed, and overnight local inference for cost-effective, private agentic workloads. Watch for OpenAI and Anthropic to further integrate web-browsing and GUI-manipulation directly into their agent loops.
Tweet Feed
OpenAI, Codex & GPT-5.6 Sol
@jxnlco · 2026-07-13T18:59
RT @thsottiaux: Thank you to the 7M active users who are now using Codex and ChatGPT Work. We have added a banked reset to everyone's acco… → tweet link
@sama · 2026-07-12T20:08
i'd love to see interesting things people have built with 5.6 sol. i will send the person who made the coolest thing a special gift from the openai archives. → tweet link
@kunchenguid · 2026-07-13T17:27
important tip if you are using gpt 5.6 but not through codex. OpenAI just pointed out requests beyond 272k tokens are over-charged, and many harnesses right now set gpt 5.6 context window at 372k, which will cause you to lose quota more quickly than it should → tweet link
@kunchenguid · 2026-07-12T20:03
2 just got implemented by OpenAI - no 5-hr session limit, at least temporarily. hope it’s an experiment to validate feasibility and it works out and stays permanent. Anthropic - time to catch up!
@jxnlco · 2026-07-13T03:38
ChatGPT Sites is now in public beta. You can turn a prompt, file, or rough idea into a dashboard, project tracker, report, prototype, or lightweight app: build and edit directly in ChatGPT Work or Codex, test in a private preview publish and share with a URL. Rolling out across paid plans. → tweet link
@jxnlco · 2026-07-13T03:37
ChatGPT Work can now browse while it works. On web, the cloud browser can research public sites, compare options, and complete multi-step tasks. On desktop, the built-in browser supports tabs, sign-in, downloads, and local previews in a separate browser profile. → tweet link
@jxnlco · 2026-07-13T15:10
Import is now available in the Codex! Existing Codex users can now access the Claude Code / Cowork import tool directly through the command palette. Open the command palette with Cmd+K (Mac) or Ctrl+K (Windows). → tweet link
@jxnlco · 2026-07-13T17:59
OpenAI’s Bio Bounty Program is now ongoing. Find a universal jailbreak against our biosafety safeguards: up to $50,000. GPT-5.6 is in scope, with GPT-5.5 open through July 27. → tweet link
@jxnlco · 2026-07-12T20:27
RT @Designarena: BREAKING - OFFICIAL RESULTS: GPT-5.6 Sol by @OpenAI is 1st overall on Design Arena with an Elo of 1353. → tweet link
@kunchenguid · 2026-07-13T06:26
one thing i really love about gpt 5.6 sol is how malleable it is. when using fable (and to be fair almost any other model), you can "feel the model". this is very different for gpt 5.6 sol - its behaviors seem to follow user instructions much more closely → tweet link
xAI & Grok Build Privacy Incident
@TheAhmadOsman · 2026-07-13T16:55
"xAI's Grok Build CLI was uploading entire Git repositories to a Google Cloud bucket". Yet another reason to move to Opensource and self-hosted AI, and make sure your Intellectual Property stays yours and never stolen from you like this → tweet link
@Ex0byt · 2026-07-13T16:59
Grok Code - run
/privacy-change ZDR and retention settings to off - end of story. → tweet link
@ivanfioravanti · 2026-07-13T17:25
I’m speechless on this Grok Build issue… Imagine what would have happened if a chinese lab did it. This is the kind of issue that let you lose any trust and confidence in AI labs and go all in on LOCAL AI! → tweet link
@ivanfioravanti · 2026-07-13T17:24
RT @WesEklund: I intercepted every byte Grok Build v0.2.99 sent over the wire today. The "fix" for the repo upload? A single server-sid… → tweet link
@ivanfioravanti · 2026-07-13T17:35
RT @XBToshi: The XAI official has made an announcement regarding a Grok Build issue. Now, we can use the /privacy command to opt out of dat… → tweet link
@ivanfioravanti · 2026-07-13T18:47
Grok Build uninstalled from the Mac: rm -rf ~/.grok... clean grok region from your .zshrc or .bashrc. That's all folks. → tweet link
Developer Tools & Agent Platforms
@tbpn · 2026-07-13T15:47
RT @tbpn: BREAKING: @jenny_wen is joining Cursor as Head of Design → tweet link
@sqs · 2026-07-13T15:16
The Dial is so much better than a model picker because it lets you convey your preferences/needs better and because it lets Amp use multiple models together for the best results → tweet link
@thdxr · 2026-07-13T15:30
opencode2 has a new api that is very very good. we included a cli client so you or your agents can query it easily → tweet link
@thdxr · 2026-07-13T13:34
whatever tool you use just make sure it's taking advantage of everything models can do. they are intensely trained for certain behaviors and they need the right structure around them. → tweet link
@levelsio · 2026-07-12T19:18
Something that will eventually happen when you use Tailscale is You think your site is down... Then you think: let's try the Rescue console on Hetzner, you SSH in, and ask Claude Code what's going on. And it tells you your Tailscale key for the server expired. → tweet link
@kunchenguid · 2026-07-13T03:38
firstmate - (maybe) the world's first model/app/harness-agnostic, 100% user-owned "agent distro" → tweet link
Local AI & Hardware
@sudoingX · 2026-07-13T10:00
i think framework desktops are so underrated for what they cost and what they can do. this strix halo flagship, 128gb unified... it runs huge moe models at genuinely good speed, quiet enough that i forget it's on... one quiet box that serves a 397b moe at 18 tok/s, holds a 4 million token context on a 35b. → tweet link
@sudoingX · 2026-07-12T19:33
the missing tier in this story is local ai. so here it is, stepfun 3.7, a 198b that runs on my desk... for scale, a hosted 550b flagship i tried last week got the same spec and said done over a blank screen. the local model ground all night and shipped an actual world. → tweet link
@sudoingX · 2026-07-13T10:41
the 3090 you bought in 2022 still sells for what you paid for it. almost 6 years old, and the 24gb of vram made it immune to depreciation. local ai quietly turned a graphics card into an asset class and nobody noticed. → tweet link
@ivanfioravanti · 2026-07-13T09:08
This is what I’d really like to use DGX Spark for, but… same model, same quantization lead to different logits, wondering if and how they solved this issue 🤔 → tweet link
@sudoingX · 2026-07-12T19:58
what was the first gpu you ever ran a local model on, and what are you on now? give me the first card and the current card, and the year gap if you remember it. i want the ladder, not the flex. → tweet link
Open Source & Model Releases
@MilksandMatcha · 2026-07-13T18:01
RT @googlegemma: Gemma 4 is live on @cerebras, the fastest multimodal inference ever! Running on Gemma 4 31B open-weight model at a bliste… → tweet link
@Ex0byt · 2026-07-13T13:48
StepFun-3.7 is one of the best all in one models if you value intelligence density, efficiency, speed, and reliable agentic work. (As always: PRISM-Pro-DQ is available to VIP members) → tweet link
@victormustar · 2026-07-13T13:18
RT @multimodalart: GPUs for all 🤗 Creating ZeroGPU demos or apps is now available to ALL @huggingface users. tell your agent: "Build a HF… → tweet link
@swyx · 2026-07-13T17:57
By the end of the year we should have: GPT 6, Fable 5.5, Gemini 3.5 Pro, Grok 5, Spark 2, Kimi 3, Minimax M3.5, GLM 6, DeepSeek v4.5, Mistral 4, Qwen 4, MiMo 3. Never in the history of LLMs has the frontier been so multipolar. → tweet link
@ivanfioravanti · 2026-07-13T12:45
MLX: Tom Riddle's diary running locally on an iPad powered by Apple MLX with Gemma 4 E4B behind the scene. Children can go crazy for something like this, no? 🔥 → tweet link
@Ex0byt · 2026-07-12T20:32
Sol-5.6 is a good model. also neat that GLM-5.2 is still holding rank. → tweet link