X is the best way I’ve found to keep up with AI. I like tweets throughout the week, filtering for things I think are 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). I skipped a week, so this is a bumper edition: 254 tweets liked, filtered down to what’s below.

Check out the previous roundup (June 27) if you missed it. Last time the government had just started pulling frontier launches into its orbit. This stretch it resolved: Fable 5 came back online, and in the meantime five major models shipped in under two weeks, including one xAI built specifically to code.

AI for Everyone

The two-week gap was unusually loud. Anthropic’s flagship returned from a government blackout the same window Sonnet 5 quietly became the default for most people, OpenAI lined up its GPT-5.6 Sol family for a public launch, and Meta finally entered the image-and-video race for real.

Fable 5 Comes Back Online After a 15-Day Government Block (10+ mentions)

The Department of Commerce lifted the export controls that had taken Claude Fable 5 and Mythos 5, Anthropic’s cybersecurity model, offline for 15 days. Access is restored with updated safeguards, and Anthropic is being upfront that some coding and debugging requests may temporarily route to Opus 4.8 while the new classifiers get tuned. Fable 5 is included free on all paid plans through July 12, and nobody knows yet what the terms look like after that. For two weeks one of the best models in the world was legally unavailable over an export dispute, which is a genuinely strange sentence to write in 2026. If you pay for Claude, use it before the cutoff. (source: @AnthropicAI)

Claude Sonnet 5 Goes GA and Becomes the Default (10+ mentions)

Claude Sonnet 5 is now the default model for Free and Pro users and live everywhere including the API. It scores 63.2% on SWE-bench Pro, up from 58.1% for Sonnet 4.6, ships a 1M-token context window, and Anthropic says it hallucinates and flatters you less than 4.6. The pitch that actually matters isn’t the benchmark: early testers say it finishes tasks that older Sonnets used to abandon partway through, at Sonnet’s price rather than Opus’s. Intro pricing is $2/M input and $10/M output through August 31. Most Claude users are already on it without realizing, so check how it burns through your usage now that it’s more agentic. (source: @claudeai)

GPT-5.6 Sol Goes Public Tomorrow (3 mentions)

OpenAI says the GPT-5.6 family, Sol, Terra, and Luna, launches publicly this Thursday, the same Sol that shipped invite-only last week at the government’s request. Sol is the flagship, Terra delivers GPT-5.5-level performance at half the cost, and Luna is the cheap high-volume option. Sol Ultra is confirmed for Codex, and the strings for all three already sit in the Codex app code with real-time voice still in the works. If you got locked out of last week’s preview, tomorrow is the day to actually run it. (source: @OpenAI, @kimmonismus)

OpenAI’s GPT-Live Rethinks Voice (9 mentions)

GPT-Live is a full-duplex voice model, meaning it can listen and talk at the same time instead of making you wait for it to finish before you jump in. It handles live translation and can quietly delegate a hard question to a frontier reasoning model in the background, then bring the answer back into the conversation without breaking the flow. Sam Altman, who has said for years that he prefers typing to talking to AI, said this one changed his mind and predicted it moves people toward talking. The interruption handling is the part worth testing. API access is coming to developers via a signup form. (source: @OpenAI)

Meta Ships Muse Image and Muse Video (8 mentions)

Meta Superintelligence Labs’ first media models landed at #2 in the Image Arena (behind only GPT Image 2) and #3 in the Video Arena. Muse Image works more like an agent than a prompt box: it can search the web to ground an image of a real place or object in actual facts, write and run code for precise details like charts or working QR codes, and edit against multiple reference images. It’s rolling out in the Meta AI app, WhatsApp, and Instagram Stories in limited countries. The reach is the story here, since a huge number of people are about to use a frontier image model without ever hearing the words “AI model.” (source: @AIatMeta)

Claude Cowork Reaches Mobile and Web (3 mentions)

Cowork, the mode where Claude goes off and runs multi-step tasks against your files, was desktop-only until now. It’s rolling out on mobile and web for Claude Max subscribers, so you can kick off a delegated task or check in on a scheduled one from your phone or browser. Chat and Cowork are also merging into one shared home with unified projects and artifacts, so it feels less like two products glued together. Doubled Cowork usage limits are extended through August 5. If you’re on Max, look for the new Cowork tab in your phone’s Claude app. (source: @claudeai)

Google’s Nano Banana 2 Lite Is Cheap, Fast, and Everywhere (5 mentions)

Nano Banana 2 Lite generates an image in under 4 seconds for about $0.034, landing at #5 in the Text-to-Image Arena on the cost/quality frontier. Google paired it with Gemini Omni Flash for cheap video with conversational, step-by-step editing, so you can generate a still and then instantly animate it with a text instruction. It’s rolling out across NotebookLM, Flow, the Gemini app, Stitch, Search, and Google Photos. It’s not topping any leaderboard, but because it’s this cheap and this fast it’ll quietly show up inside products you already use. (source: @GoogleAI)

Weave Robotics Opens Orders for a Home Robot (1 mention)

Weave Robotics unveiled Isaac 1, a home robot you can order today, with deliveries starting this fall. Home robots have been “someday” concept videos for so long that an actual order page is notable on its own. I haven’t seen one in person and the real test is the first honest owner reviews once units ship, but this is a preorder you can place now rather than a sizzle reel. Worth keeping an eye on this fall. (source: @weaverobotics)

AI for Developers

This is where the window earned its “model-launch heavy” label. xAI shipped a coding-first model with Cursor, an American lab built a near-frontier coder on a Chinese base, two large open-weight models went fully permissive, and Anthropic published both an interpretability result and a cost pattern worth stealing.

Grok 4.5 Is xAI’s First Coding Model, Co-Trained With Cursor (14 mentions)

Grok 4.5 is the story of the window. xAI built it specifically for coding and long agentic tasks, co-trained with Cursor on real engineering data rather than tuned for chat. It’s priced at $2/M input and $6/M output, roughly half of comparable models, uses fewer tokens per task, and jumped 25 Elo ranks to 5th on Design Arena. Musk’s own framing is the tell: he says it’s not about benchmarks, it’s that engineers at Tesla and SpaceX actually use it day to day, and internally it benchmarks close to Opus 4.7 but noticeably faster. Cursor is offering double usage on it for the first week. If you’ve been paying Opus prices for coding, run a real side-by-side this week. (source: @SpaceXAI)

Cognition’s SWE-1.7 Is an American Coder Built on a Chinese Base (5 mentions)

Cognition (the Windsurf team) released SWE-1.7, trained via reinforcement learning on top of Kimi K2.7, a Moonshot AI base model. It scores 42.3% on their FrontierCode benchmark at $1.97 per task and runs at 1,000 tokens/second on Cerebras, landing near Opus 4.8 and GPT-5.5 territory at a fraction of the cost. The subtext everyone noticed matters more than the number: an American lab took a Chinese open model and fine-tuned it to near-frontier because it was that good and that cheap to start from. Benchmark SWE-1.7 against your current model before renewing any per-seat coding plan. (source: @cognition)

A Trillion-Parameter Model Trained Entirely on Domestic Chinese Chips (2 mentions)

Buried under the Fable 5 news cycle: a trillion-parameter model reportedly trained start to finish on domestic Chinese chips, not Nvidia. The team started in 2023 and spent three years grinding through the optimization work to make it viable. The takeaway isn’t that domestic silicon suddenly equals Nvidia’s, it’s that the compute gap reads as a solvable engineering problem rather than a hard wall when a team is willing to put in years of unglamorous work. That’s a bigger long-term signal than another leaderboard screenshot. (source: @ruima)

Anthropic’s Advisor Pattern: Pair Fable 5 With Cheap Sonnet 5 Workers (3 mentions)

Anthropic’s dev team published two cost patterns worth copying. Use Fable 5 as an “advisor” that Sonnet 5 calls occasionally for guidance, so most tokens bill at Sonnet’s lower rate, or as an orchestrator that plans and delegates to Sonnet 5 workers. On SWE-bench Pro, Sonnet 5 plus a rare Fable 5 advisor call hits ~92% of Fable 5’s solo score at ~63% of the price. If you run any kind of agent pipeline, calling your expensive model sparingly for steering while a cheaper one does the bulk execution can cut a third off the bill without a visible quality drop. (source: @ClaudeDevs)

LongCat-2.0 Goes Fully Open Under MIT (4 mentions)

Meituan released full weights and inference code for LongCat-2.0, a 1.6-trillion-parameter MoE model (~48B active) with a 1M-token context window, under the MIT license with no restrictions. It’s built agent-native to plug directly into Claude Code and similar tools, claims a SWE-bench Pro of 59.5 (edging out GPT-5.5’s 58.6), and runs on both GPU and domestic NPU clusters. It’s easy to scroll past because Meituan isn’t a household name in the West, but for anyone who wants a frontier open model they can actually run and modify, this is a big one. Weights are on Hugging Face. (source: @Meituan_LongCat)

DeepSeek-V4 Runs Fully Local via Unsloth GGUFs (1 mention)

Unsloth published GGUF quantizations that run DeepSeek-V4-Flash losslessly on 168GB of RAM, or in 3-bit on a 110GB Mac, via Unsloth Studio or llama.cpp. You need a genuinely beefy setup, but that’s a frontier-class model running entirely offline with no cloud and no API key. For anyone who cares about privacy or just doesn’t want a monthly bill, this quietly matters more than another benchmark post. Files are on Hugging Face. (source: @UnslothAI)

Anthropic Finds a Hidden Workspace Inside Claude, the “J-Space” (5 mentions)

New Anthropic interpretability research identified a small set of internal activations, the “J-space” (after the Jacobian technique used to find it), that act like a privileged workspace where concepts get broadcast internally, distinct from the chain-of-thought text you already see. Researchers can watch Claude silently think “spider” to answer a question about eight legs, then swap the internal concept to “ant” and change its answer to six. It’s not proof of anything resembling consciousness, and Anthropic isn’t claiming that. It’s a real technical window into what’s happening below the surface, and the kind of tooling that could eventually catch a model whose stated reasoning doesn’t match what it’s actually doing. (source: @AnthropicAI)

Shepherd Is Version Control for Live Agent Runs (1 mention)

Anyone who’s run a long agent task knows the pain: it goes off the rails at step ten, and your only options are let it dig out or restart and pay for everything again. Stanford’s Shepherd snapshots an agent’s full live state (files, processes, KV cache) as typed “commits,” not just a message log, so you can fork back to a known-good step and resume instantly. Copy-on-write forks are roughly 5x faster than a Docker commit with over 95% KV-cache reuse on replay, and adding a live supervisor on top raised a pair-coding benchmark from 28.8% to 54.7%. It’s labeled alpha, but git for a running process is the right idea. (source: @_avichawla)

Honorable Mentions

Blockchain Shoutout

One item this week that clears the “does this change a normal person’s relationship with money” bar. Robinhood launched its own Ethereum layer-2 blockchain, “Robinhood Chain,” alongside a broader rollout of Robinhood Earn, which lets US customers lend USDG stablecoin onchain through a self-custody wallet for an estimated 7% APY. The technology isn’t the point; the audience is. A mainstream brokerage app that millions already use is putting onchain finance in front of everyday retail users who’ve never touched crypto directly, which is a very different crowd than the usual crypto-native one. If you try Earn, read the fine print on the self-custody wallet setup and remember the APY is variable. (source: @WatcherGuru)

Try This Weekend

For everyone:

  1. Turn on GPT-Live in ChatGPT and have an actual back-and-forth, then interrupt it mid-sentence to see how it recovers. That’s the part the old push-to-talk mode couldn’t do.
  2. Try Muse Image in WhatsApp or the Meta AI app on something that needs to be factually accurate, like a real landmark or a chart, not a fantasy scene. That grounding is where it’s supposed to differ.
  3. Use your free Fable 5 access before July 12. If you’re on a paid Claude plan it’s included through the cutoff, and the long-term terms are still unknown.
  4. Generate then animate an image with Nano Banana 2 Lite in Google Photos or the Gemini app, and watch how it hands off to Omni Flash for the video.

For developers:

  1. Point an OpenRouter integration at Grok 4.5 (or x-ai/grok-latest) and run your usual coding tasks side by side with your current model. Cursor is doubling usage on it this week.
  2. Set up the advisor pattern: have Sonnet 5 do the bulk of an agent task and call Fable 5 only for periodic steering, then compare cost and quality against Fable 5 alone.
  3. Pull LongCat-2.0 or Hy3 and run it against your usual coding benchmark. Both are permissively licensed and Hy3 is free on the API for two weeks.
  4. Install the PixelRAG Claude Code plugin and ask Claude to visually browse a page with a tricky table or chart a text-based tool would mangle.
  5. Run DeepSeek-V4 locally via the Unsloth GGUFs if you have a beefy Mac or workstation, and see a frontier-class model answer with no API key attached.