X is still the best way I’ve found to keep up with AI. I like tweets throughout the week, filtering for what’s actually worth knowing, then use Claude Code to pull those likes automatically and help me turn them into this post (here’s how the pipeline works). It’s been two weeks again, so this is another bumper edition: 285 tweets liked, filtered down to what’s below.

Check out the previous roundup (July 8) if you missed it. Last time the story was five frontier models shipping in a fortnight. This stretch the frontier stopped being the story: an open Chinese model reached the top of a coding leaderboard, an American lab open-sourced a trillion-parameter multimodal model, and Washington started drafting a response.

AI for Everyone

The models kept coming, but the shape of the news changed. The most talked-about model of the two weeks is one anyone can download, it comes from a Chinese lab, and its arrival is now a policy question in the US. Meanwhile the American labs spent the window reshuffling what you get for your subscription.

Moonshot’s Kimi K3 Tops a Coding Leaderboard and Lands in Washington (15+ mentions)

Kimi K3, from Chinese lab Moonshot AI, is the story of the fortnight. It launched July 16, topped the Frontend Code Arena with a 76% pairwise win rate, and knocked Claude Fable 5 off the number-one spot. On the Artificial Analysis Intelligence Index it scores 57, in the same neighborhood as Opus 4.8 and GPT-5.5, and Moonshot says it plans to release the weights of the full 2.8-trillion-parameter model, which would make it the strongest open-weights model available. Demand was heavy enough that Moonshot paused new subscriptions for a stretch. The part that turns this from a leaderboard post into news: the Trump administration is reportedly weighing an executive order to restrict Chinese open-source models in the US, with momentum reviving specifically after K3, per Axios. When a free model you can download is good enough to trigger an export fight, open weights have stopped being a hobbyist story and become a strategic one. (source: @Kimi_Moonshot)

GPT-5.6 Sol Goes Public and ChatGPT Absorbs Codex (5 mentions)

The GPT-5.6 family, Sol, Terra, and Luna, rolled out in ChatGPT, Codex, and the API, and Sol landed joint number one in the Frontend Code Arena alongside Fable 5. The bigger everyday change is packaging: OpenAI is merging Codex and ChatGPT into one desktop app, so the coding tool and the chat app stop being two separate things, and the combined product now reports 8 million active users. If you already pay for ChatGPT, Sol is the model you’ll get by default now, and the desktop app is where OpenAI clearly wants you living. I’ve been using Sol for a week and the speed is the part that stands out more than the benchmark. (source: @OpenAI)

Anthropic Puts Fable 5 in Max and Team Plans (5 mentions)

After Fable 5 spent June coming back from a government block, Anthropic settled the terms: starting July 20, Fable 5 is included in all Max and Team Premium plans at 50% of the normal limits, and Claude Code’s weekly rate limits stay 50% higher. The timing wasn’t subtle. As one observer put it, Anthropic waited until OpenAI shipped GPT-5.6 and its combined app before announcing the reset. Every lab this window was reacting to the others, and even Elon Musk conceded that Anthropic is currently the model to beat. If you’re on Max or Team, Fable 5 is now part of what you already pay for, so check your model picker. (source: @claudeai)

Grok Bundles Premium+ Into SuperGrok Heavy (3 mentions)

xAI made its subscription math simpler: SuperGrok Heavy now includes X Premium+ at no extra cost, so if you were paying for both, you can cancel the standalone Premium+ and keep it as a bundled benefit. It’s a small thing, but it’s the kind of packaging move that tells you xAI wants Grok and X treated as one product rather than two subscriptions. If you pay for both today, this is money back. (source: @premium)

Claude for Teachers Gives K-12 Educators Free Premium Access (2 mentions)

Anthropic launched Claude for Teachers, free premium Claude access for verified K-12 educators in the US, with a library of teaching skills and curricula mapped to academic standards in all 50 states. Free-for-a-vertical launches are usually a land grab, and this one is aimed squarely at getting Claude into classrooms before the school year. If you teach, or know someone who does, the verification is the only hoop and the tier is the paid one. (source: @claudeai)

AI for Developers

If you build with these models, the open-weights side is where the fortnight got genuinely interesting. A trillion-parameter multimodal model went open, Grok’s coding CLI got open-sourced, and two more open-weight releases are queued right behind Kimi K3.

Thinking Machines Ships Inkling, the First Open 1T Multimodal Model (4 mentions)

Mira Murati’s Thinking Machines released Inkling, which it calls the first open 1-trillion-parameter multimodal model handling text, audio, and image in one set of weights. Audio is the standout: it ranks among the strongest open-weights models on VoiceBench, MMAU, and AudioMC. It also ships a “continuous thinking effort” dial that lets you trade tokens for accuracy, reaching the same score with a fraction of the compute when you want it cheap. It’s live on OpenRouter and there are already GGUF quants from Unsloth for running it locally, plus a smaller Inkling-Small variant with weights to follow. For a lab’s first public model, shipping an open 1T multimodal with real audio strength is a serious opening move. (source: @thinkymachines)

Meta’s Muse Spark 1.1 Undercuts the Frontier on Price (4 mentions)

Meta released Muse Spark 1.1, an agentic coding model it says is SOTA on Harvey’s Legal Bench, TaxEval, and MedScribe, taking the top TaxEval spot from Fable 5 while being 10x cheaper and twice as fast. On the Vals Index it ranks #4 and is the fastest model in the top 10, roughly 3x faster than the three models above it. The Meta Model API also lets developers build on Muse Spark for the first time, so it’s no longer locked inside the Meta AI app. The pitch is unglamorous and effective: near-frontier quality on domain benchmarks at a price that makes high-volume agent work actually affordable. Worth a real cost comparison if you run anything at scale. (source: @finkd)

xAI Open-Sources Grok Build (3 mentions)

xAI open-sourced Grok Build, its coding CLI, and reset usage limits for everyone. This lands right after Grok 4.5, the coding model xAI co-trained with Cursor, shipped at $2/M input and $6/M output, roughly half of comparable models, with users reporting Fable-level output at a fraction of the cost. xAI also addressed a privacy question directly, stating Grok Build has honored zero data retention since launch with an opt-out in the CLI. An agentic CLI whose source you can actually read and fork is a much more concrete thing to offer developers than a hosted product. Try it against your current setup on a throwaway repo. (source: @SpaceXAI)

Qwen3.8 Is in Preview and Going Open-Weight Soon (4 mentions)

Alibaba’s Qwen3.8 is in public preview, improving day by day with a notable step up on web frontend work, and Alibaba says it’s launching and going open-weight soon. Between Kimi K3, Inkling, and Qwen3.8, the open-weight release calendar is stacked right now. Separately, Unsloth shipped new Qwen3.6 quants that run 2.5x faster on a single GPU if you want the current generation running locally today. Test the preview on your own frontend tasks before the weights drop, then switch to the open build when it lands. (source: @Alibaba_Qwen)

OpenAI Drops a Codex Plugin for Claude Code (3 mentions)

In a genuinely funny bit of cross-pollination, OpenAI shipped a Codex plugin for Claude Code, so you can drive Codex from inside Anthropic’s CLI. Codex itself also became a full video editor this window, which is a stranger sentence than it sounds. On the Anthropic side, Claude Code added an in-app browser on desktop and a new /checkup command for diagnosing your setup. The tooling war is producing weird, useful crossovers faster than anyone can keep up with. (source: @Saboo_Shubham_)

Honorable Mentions

Try This Weekend

For everyone:

  1. Open GPT-5.6 Sol in ChatGPT and, if you’re on the new combined desktop app, notice how the Codex coding tools now live in the same window as chat.
  2. If you teach K-12 in the US, verify for Claude for Teachers and pull down a teaching skill mapped to your state’s standards before the term starts.
  3. Check your SuperGrok Heavy billing and cancel a redundant standalone X Premium+ if you’re paying for both.
  4. If you’re on Claude Max or Team, switch your model picker to Fable 5 now that it’s included, and see how far the 50%-of-limits allowance gets you.

For developers:

  1. Call Kimi K3 through Cloudflare AI Gateway and run it against your usual coding tasks. It’s the model that just topped the Frontend Code Arena.
  2. Run Inkling via OpenRouter or the Unsloth GGUFs and lean on its audio understanding, which is where it’s strongest.
  3. Clone the open-sourced Grok Build CLI, point it at a throwaway repo, and compare it to your current agentic setup.
  4. Try the Qwen3.8 preview on a frontend task now, and grab the Unsloth Qwen3.6 quants if you’d rather run something open locally today.
  5. Install the Codex plugin for Claude Code and drive an OpenAI model from inside Anthropic’s CLI, just to see the two ecosystems talk.