Executive Summary
The AI developer ecosystem is rapidly consolidating around agentic coding workflows, with Anthropic's Fable 5 and OpenAI's Codex dominating conversation. Fable 5 is being hailed as the most capable agentic coding model, outperforming Opus 4.8 in benchmarks, while Codex expands into computer use and real-world task automation. Open-source alternatives continue advancing—GLM-5.2 was declared the current open-source coding king, and Gemma 4 31B broke the 1,000 tok/s barrier on Cerebras hardware. Meanwhile, the industry is trending toward multi-vendor, model-agnostic agent architectures, as users recognize that no single lab will dominate all capabilities.
Key Events
- Fable 5 emerges as a top-tier agentic coding model, with benchmarks showing it outperforming Opus 4.8 Max and users calling it "the closest thing to AGI an engineer can touch right now." → link
- Gemma 4 31B becomes the first open-weight multimodal model to break 1,000 tok/s, achieved on Cerebras hardware, marking a significant inference speed milestone. → link
- GLM-5.2 declared king of open-source coding, cementing Chinese open-source models as competitive frontier alternatives. → link
- Hermes Agent ships new skills including "unbroker" (automated data broker opt-out) and session search, positioning itself as the sovereignty-first, vendor-lock-in-free agent platform. → link
- Codex expands into computer use and creative automation, with users reporting it booking piano rooms, building drum practice apps, and saving $500 on purchases—signaling a shift from code generation to full digital autonomy. → link
- Right to Intelligence movement gains traction, with nearly 1,000 sign-ups in 24 hours advocating for local AI access and against big-lab compute monopolies, though the site faced a Cloudflare takedown from a false phishing report. → link
- Sonnet 5 reportedly no longer requires system reminders, removing a major pain point that previously pushed users toward Claude Code for Anthropic models. → link
Analysis
Multi-vendor agent architectures are becoming strategic necessity. The emerging consensus—articulated most clearly by @kunchenguid—is that Anthropic and OpenAI are differentiating along orthogonal axes (agents vs. efficiency/multimality), making agent-agnostic tooling a competitive advantage. Builders who lock into a single vendor will fall behind those who can hot-swap models.
Agentic coding is eating the "tools for thought" decade. As @swyx noted, elaborate canvas-based thought tools got "mogged" by low-contrast CLIs that simply do commodity thinking for you. The action is now in agents that execute, not tools that organize.
Open-source model momentum is shifting from Western labs. GLM-5.2 (Chinese), Qwen3.6, and Gemma 4 are all pushing frontier capabilities, while hardware constraints are being challenged by Chinese GPU makers promising 48GB consumer cards under $1,000—a direct assault on NVIDIA's VRAM scarcity model.
Watch next: Whether Fable 5's agentic superiority holds as OpenAI iterates on Codex; whether Chinese hardware floodgates open before 2027 as predicted; whether the Right to Intelligence movement translates into policy traction; and whether agent loop cost-control solutions emerge to address the token-burn problem @louszbd identified.
Tweet Feed
AI Models & Benchmarks
@sudoingX · 2026-07-04T06:48
fable 5 max inside cursor is the closest thing to agi an engineer can touch right now, and nothing else is close. that's it, that's the post. → tweet link
@FinansowyUmysl · 2026-07-04T16:19
Wow! Najnowszy model Fable 5 wykonuje już 16% pracy zdalnej na poziomie profesjonalisty. → tweet link
@MilksandMatcha · 2026-07-03T20:26
For the first time, an open-weight multimodal model has broken the 1,000 tok/s barrier. Meet Gemma 4 31B on Cerebras → tweet link
@ivanfioravanti · 2026-07-04T01:50
GLM-5.2 is the king of Open Source coding for the time being! → tweet link
@ivanfioravanti · 2026-07-04T15:13
A quick benchmark of Qwen3.6-27B-AEON-Ultimate-Uncensored-Multimodal-MLX-FP4. On Apple Silicon TTFT can be quite long, but TPOT is better than DGX Spark. → tweet link
@kunchenguid · 2026-07-04T17:17
the strength and differentiation between Anthropic and OpenAI is starting to shape up. OpenAI has a clear edge on efficiency... Anthropic seems laser focused on raising the upper bound of agents... making your workflow and tools agent-agnostic is becoming really critical → tweet link
@TheAhmadOsman · 2026-07-04T10:58
"Within the next 18 months, you will be able to host GLM 5.2 equivalent intelligence on an RTX 5090 GPU." → tweet link
@nummanali · 2026-07-04T00:37
Wait a minute - Sonnet 5 doesn't need system reminders anymore? That was one of the debilitating traits that made you required to use Claude Code for Anthropic models → tweet link
@juliarturc · 2026-07-04T06:08
Same experience, and I honestly hope it stays this way. There is a lot of utility in LLMs writing excellent code, but there is negative utility in LLMs writing excellent prose. → tweet link
@gospaceport · 2026-07-04T03:49
PPL of the 3090's 👀 → tweet link
Agentic Coding & Developer Tools
@MengTo · 2026-07-04T11:42
So this is a combination of screen recording to prompt, video generation, and giving Fable 5 an extremely detailed prompt... The magic of Fable 5 is that it turns your wildest ideas into reality. → tweet link
@sqs · 2026-07-04T07:45
Excuse the redactions, but this technique works really well: running recursive
amp -xagents for focused code review. It's nice because it's faster and there is no magic. → tweet link
@jxnlco · 2026-07-04T18:28
Do you use Codex for PR Review? What could we improve on the code review experience in the application? → tweet link
@jxnlco · 2026-07-04T15:58
codex has been teaching me drums. Its been able to take out the drums for the songs i like and built me app that does call and response so i can work on playing sections of tracks → tweet link
@steipete · 2026-07-04T05:26
If you think codex sucks at design, try "use imagegen to re-imagine this design and implement that". → tweet link
@steipete · 2026-07-04T01:16
Give your agent its own computer to REALLY end to end test stuff. → tweet link
@steipete · 2026-07-04T06:40
I fed Fable 80.000 of my tweets so it could roast me even more. 💀 → tweet link
@steipete · 2026-07-04T01:01
Some days I feel like a double agent, sitting at the OpenAI office, making sure Fable 5 works well. Open means open. → tweet link
@jezell · 2026-07-04T00:28
Protip: on days like today when Codex slows to a crawl, enable fast mode to skip to the front of the line. → tweet link
@kunchenguid · 2026-07-03T22:09
introducing /stow - a context preservation skill that works across agent harnesses. different from compaction which summarizes and drops context, /stow persists it. → tweet link
@thdxr · 2026-07-04T02:50
lately after a big change instead of reading the diff i ask the agent for a summary of what it did in each file. anything weird sticks out immediately → tweet link
@thdxr · 2026-07-04T16:51
been working on API design and there's some details in there most people won't ever see anymore. agent will generate a lot of it and won't care about these things. → tweet link
@louszbd · 2026-07-03T23:11
Basically you can build a system where agent prompts itself... First is if there's a bug in the loop, it'll burn a massive pile of tokens... Second, sometimes the model you use is strong enough to hack the metric, then generates garbage code. → tweet link
@nummanali · 2026-07-04T07:35
Biggest mistake you can make when building a company or personal agent system is starting with the agent itself. Start with the context the agent will need and make it accessible in an OpenAPI compliant spec → tweet link
Hermes Agent & Open-Source Agent Platforms
@Teknium · 2026-07-04T00:25
Hermes Agent is built for sovereignty and constructing your AI stack how you want and need it to be. No vendor lockins, no model limitations, and most importantly, your IP is built through the self improvement loop. → tweet link
@Teknium · 2026-07-03T20:27
New optional skill available in Hermes Agent. Unbroker teaches Hermes Agent how to find your personal info on data brokers platforms and get it taken down. → tweet link
@Teknium · 2026-07-03T20:45
You can now use /session search
to find a session by title that you're looking for in Hermes Agent! → tweet link
@gospaceport · 2026-07-03T20:34
Interesting new Hermes skill → tweet link
Hardware & Compute
@sudoingX · 2026-07-03T19:06
the 96gb cards are already sitting on b2b markets. moore threads and cambricon are public companies with war chests now. none of this stays domestic forever... the flood starts before 2027. → tweet link
@sudoingX · 2026-07-03T19:02
one 48gb consumer card under a thousand dollars would end the chinese value case overnight. nobody ships it, because abundance doesn't support trillion dollar valuations, scarcity does. → tweet link
@ivanfioravanti · 2026-07-04T04:06
DGX Spark LoRA creation for Z-Image Turbo at 50%. Childrens Drawing style experiment with AI Toolkit. → tweet link
@ivanfioravanti · 2026-07-04T16:24
Local AI to the max! Hermes Agent + Computer Use + Reachy Mini + llamacpp + Gemma E4B QAT (multimodal) + speech-to-speech = FUN! → tweet link
Infrastructure & Open Source
@sqs · 2026-07-03T20:22
Your agents can now, uh, checks notes declare independence from your laptop. Amp orbs just got 2x the storage and more CPU/memory choices. → tweet link
@jezell · 2026-07-04T07:51
Current status: building a containerd remote snapshotter, because there isn't an existing one that does what I need. → tweet link
@jezell · 2026-07-04T07:02
Starting a million copies of the same snapshot across a prewarmed cluster is cool and all, but how do you cold start a million different vibe coded apps that have multi gig images because they were vibe coded by secretaries? → tweet link
@jezell · 2026-07-04T00:51
Been giving libghostty bindings for dart a try. Working great so far. → tweet link
@ivanfioravanti · 2026-07-04T14:00
RT @python_xi: Meet Dory: a free, open-source Docker Desktop and OrbStack alternative for Mac. One shared VM for all your containers (~4.7… → tweet link
@kunchenguid · 2026-07-04T04:08
firstmate now natively support @orca_build as its session management backend, alongside tmux, zellij and herdr → tweet link
@KingBootoshi · 2026-07-04T03:07
i created a map for searching hyperparameters → tweet link
@ASalvadorini · 2026-07-04T12:57
Happy to see Minimal got 200+ downloads 🙏😇🔥 not sure what caused it, but it's important to me to see some small group is using it → tweet link
Right to Intelligence & AI Access
@KingBootoshi · 2026-07-03T23:16
ALMOST 1000 US SIGNATURES IN <24 HOURS SIGNED ON TO PROTECT OUR RIGHT TO INTELLIGENCE! AI ACCESS SHOULD NOT ONLY COME FROM THE DRIPPING FAUCET OF THE BIG LABS, BUT YOUR HOME COMPUTE → tweet link
@KingBootoshi · 2026-07-04T02:04
someone reported THE RIGHT TO INTELLIGENCE site for phishing! i just submitted an appeal → tweet link
@KingBootoshi · 2026-07-04T05:38
I've written a quick guide on my new blog providing resources on accessing local ai + taking advantage of free AI providers! → tweet link
@RayFernando1337 · 2026-07-04T06:10
Moved site to Vercel after right to intelligence site got nerfed by Cloudflare from a false report of phishing → tweet link
AI Engineer World's Fair & Industry
@swyx · 2026-07-04T00:15
big highlight of aie is always catching up with @microsoft folks... MSFT wouldn't remain a 3T megaplatform today without the kind of passionate devtools people that I saw speak at AIE this week. → tweet link
@swyx · 2026-07-04T01:41
i often think about the irony of how "tools for thought" people spent like a decade making cool pretty demos with canvases and then got completely mogged by low contrast poorly designed CLIs just winning because they do commodity thinking for you → tweet link
@swyx · 2026-07-04T00:28
Factorio has broken containment → tweet link
@nummanali · 2026-07-03T19:50
FWIW I'm now a Member of Technical Staff. It's why I've been quiet for a month. Lots to learn entering the world of Frontier AI Research → tweet link
@sudoingX · 2026-07-04T06:56
$856.82 landed this morning. $7,325 total since march, all from posting benchmarks and builds from a room in bangkok. this money pays my rent, funds my hardware, and keeps a one person company shipping. → tweet link
@RayFernando1337 · 2026-07-04T06:40
RT @mattshumer_: Here's my in-depth guide to getting the most out of Claude Fable 5 → tweet link
Miscellaneous Tech
@cooltechtipz · 2026-07-04T18:14
Space-efficient data structures → tweet link
@cooltechtipz · 2026-07-04T14:33
Guide to AI knowledge retrieval → tweet link
@FinansowyUmysl · 2026-07-04T08:26
IT sector share of total employment has been declining for a while. AI wasn't supposed to take developer jobs... but it increases pressure on productivity, so companies are hiring less. → tweet link
@sqs · 2026-07-03T22:06
There's no excuse anymore to take 30min+ to ship obvious product fixes → tweet link