Good morning AI entrepreneurs & enthusiasts,
Three labs on two continents hit the same wall this week, and every one of them hit it for the best possible reason: people wanted their product far more than anyone forecast.
Anthropic made Fable 5 permanent inside Max and Team Premium. Moonshot stopped selling Kimi K3 subscriptions 48 hours after launch because demand outran its compute. Alibaba previewed a 2.4 trillion-parameter flagship in Shanghai and priced it at 90% off. Read those three together and the story is not scarcity. It is that adoption is compounding faster than the most capitalized supply chain in history can pour concrete.
I was in the audience when Dario Amodei put Anthropic’s growth this year at 85x. Sit with that number for a second. No planning process survives it, no capacity model anticipates it, and no communications team gets the messaging perfect on the first try. What we are watching is not mismanagement. It is what genuine product-market fit looks like at a scale nobody in this industry has operated at before.
In today’s AI news:
Anthropic makes Fable 5 permanent in Max and Team Premium
Moonshot sells out of Kimi K3 in 48 hours
Alibaba previews Qwen3.8 at the World AI Conference
The revolving door between Washington and the frontier labs
Today’s Top Tools + Quick News
News: Anthropic confirmed on July 18 that Claude Fable 5 stays permanently inside Max and Team Premium, effective today. The original plan was to remove the model from subscriptions entirely. Competitive pressure and genuine consumer demand turned a removal into a permanent inclusion.
Details:
The deadline moved three times before landing here, from July 7 to July 12 to July 19, a five-week stretch that tested a lot of patience.
Fable 5 now sits permanently in Max and Team Premium at 50% of weekly usage limits. Pro and Team Standard get a one-time $100 credit, then move to usage pricing at $10 per million input tokens and $50 per million output tokens.
One wrinkle worth knowing before you plan your week: a temporary 50% Claude Code rate-limit boost expires today too, so the baseline those percentages apply to is smaller than it was during the promo window.
Anthropic cited infrastructure limits and demand it called challenging to predict, while pointing to new capacity coming online including the SpaceX Colossus 1 deal.
The competitive backdrop is real. OpenAI’s GPT-5.6 Sol runs roughly a third cheaper, and Moonshot’s Kimi K3 landed mid-decision.
Why it matters: This is what a genuinely competitive market delivers. A model that was six weeks from disappearing behind pay-per-use is now a permanent fixture in two subscription tiers, and pressure from OpenAI, Moonshot, and Alibaba is why. Be clear-eyed about the tradeoff though: permanent access at reduced caps is a better deal than losing the model, and it is not the same as the promo window you have been living in since June. Plan for less Fable 5 headroom this week, not more. The bigger lesson sits underneath the caps anyway, because at 85x annual growth, demand forecasting stops being a spreadsheet exercise and becomes an act of faith, and Anthropic found a way to keep the model in customers’ hands while it builds.
News: Moonshot AI paused new subscriptions within 48 hours of Kimi K3’s July 16 launch after demand drove its compute straight to the ceiling. The company framed the pause as protecting the experience of existing subscribers. If you have not already subscribed, you are waiting for the weights.
Details:
Kimi K3 is a 2.8 trillion-parameter sparse Mixture-of-Experts model activating 16 of 896 experts per token, with a 1 million-token context window, native vision, and an always-on thinking mode built on new architecture Moonshot calls Kimi Delta Attention.
The benchmark picture is more nuanced than the headlines suggested. K3 scored 57.1 on Artificial Analysis’s Intelligence Index, third behind GPT-5.6 Sol at 58.9 and Fable 5 at 59.9, while Arena’s blind tests had developers preferring K3 over both for front-end coding work.
On price, K3 runs about $3 per million input tokens and $15 output against Sol’s $5 and $30, undercutting the frontier by roughly 40 to 50% without being the outright cheapest option available.
Full open weights land July 27 on Hugging Face under a Modified MIT license, which would make K3 the largest open-weight model ever released.
Moonshot is also splitting memberships into separate chat and coding plans, letting it provision two very different workloads independently.
Why it matters: A model that undercuts the frontier by half and is about to be given away for free still ran out of capacity in two days. That is the whole story of this week compressed into one lab. Price was not the constraint, licensing was not the constraint, and the openness of the lab did not save it, because the binding constraint is silicon and power and it applies to everyone equally. July 27 is the date to circle, because the moment those weights hit Hugging Face, a capacity ceiling becomes a procurement question you can go solve yourself.
News: Alibaba previewed Qwen3.8-Max-Preview on July 19 at the World Artificial Intelligence Conference in Shanghai, a 2.4 trillion-parameter multimodal model the Qwen team claims ranks second only to Claude Fable 5. The preview is already live and running in production coding harnesses. Full open weights are promised, with benchmarks and the model card still to come.
Details:
Qwen3.8 is the first model in the family above 1 trillion parameters to handle images, video, and documents alongside text, built as a sparse Mixture-of-Experts stack.
Access runs through Token Plan, Qoder, and QoderWork at a promotional 10% of standard pricing, with an off-peak rate reaching 98% off.
Open weights are coming, which would break Alibaba’s long-standing practice of keeping Max-tier models closed.
The second-only-to-Fable-5 line is Alibaba’s own claim with no data attached. No benchmark scores, no activated-parameter count, no model card. For contrast, Qwen3.7-Max shipped in May with full published results including a 56.6 on the Artificial Analysis Intelligence Index.
Markets liked it modestly. Alibaba’s Hong Kong shares closed up 3.73% at HK$116.80 while the Hang Seng Tech Index added 2.79%.
Why it matters: Qwen has quietly become the default for anyone doing serious fine-tuning work, and that community has spent two years building on a family that ships reliably and prices aggressively. Qwen3.8 is the next jump in that lineage, and an open-weight release at 2.4 trillion parameters would hand those teams the largest base model anyone has ever put in public hands. The missing benchmark table is a real gap and we would rather have it today, but Alibaba’s track record earns some patience, and we expect the full release shortly. We will report the numbers the moment they land.
Ps, enjoy this very satirical image to make the political story a little bit more lighthearted.
News: Dean Ball, a co-author of the administration’s AI Action Plan, joined OpenAI as Head of Strategic Futures effective July 6, reporting to Chief Strategy Officer Jason Kwon. Within two weeks he drew public fire from David Sacks, the former White House AI and crypto czar who now cochairs the President’s Council of Advisors on Science and Technology.
Details:
Last fall, Sacks accused Anthropic of running a regulatory capture strategy built on fear-mongering, following an essay by policy chief Jack Clark on technological optimism and appropriate fear.
Five weeks ago, Anthropic proposed that the government hold standing authority to block frontier model releases that fail independent safety tests, with penalties tied to global revenue.
On July 6, Ball left his White House-adjacent policy fellowship for OpenAI.
This past weekend, Ball wrote that the administration would eventually create regulatory risk around Chinese open-weight models, arguing no formal ban is needed if agencies issue soft law that creates uncertainty.
Sacks fired back that weaponizing regulatory uncertainty is unacceptable, and that the leading closed labs, which he described as a revenue duopoly, want the government to remove their open-source competition. A senior Pentagon official separately disputed Ball’s framing, and Ball then clarified he was not advocating the approach he had described.
Why it matters: We are not here to talk about politics, and honestly we would be delighted to never talk about politics again. But politics is welded to this industry now, and pretending otherwise does not make it go away. What you are looking at is the leadership of one of the world’s top AI labs and the US government arguing in public, in real time, about who gets to compete. Most newsletters will skip this because it does not fit a product roundup, and skipping it is exactly how you get surprised when the rules change underneath you.
🌝 Kimi K3 Moonshot’s 2.8T open-weight model, top of Arena front-end code, weights land July 27
⚙️ Canva Code 2.0 vibe-coding now free for all 265M+ Canva users, with HTML import and drag-and-drop editing of AI-generated sites
🧠 Inkling Thinking Machines’ first open-weights model, 975B total and 41B active under Apache 2.0, reasoning natively over text, images, and audio
🗣️ GPT-Live OpenAI’s full-duplex voice model that listens and speaks at once, and hands hard questions to a frontier model mid-conversation
Chamath Palihapitiya weighed into the open-source debate with three words: the future is open source.
Gavin Baker of Atreides calls Kimi K3 a potential inflection point for AI, negative for Anthropic and OpenAI and net positive for nearly every other company in the world.
SpaceX is negotiating a multibillion-dollar compute deal with the Department of Defense, deepening a relationship that already spans launch, satellite communications, and missile tracking.
The MLB has cut off the custom tab on dugout iPads, after a June 11 memo found teams using it for AI recommendations on substitutions and pitch calling. The block took effect as the second half of the season opened.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.