Anthropic Backtracks and Restores Fable 5 to Max Plans at Half the Limits

01Anthropic reversed its plan to pull Fable 5 from subscriptions; Max plans get it today at 50% of limits

Claude Fable 5 is included in all Max and Team Premium plans beginning July 20, at 50% of limits, according to a post from Anthropic's @claudeai account. Pro and Team Standard users keep access through usage credits and receive a one-time $100 credit. Simon Willison, who flagged the announcement, writes that the $20 Pro plan still does not bundle the model. Subscription-included access starts at the $100 Max tier.

The post reverses Anthropic's plan to remove Fable 5 from subscription accounts, a change Willison had covered the week before. He attributes the reversal to competition, naming GPT-5.6 Sol and, to a lesser degree, Kimi 3. Subscribers had started asking why they should pay for a plan that excludes the vendor's strongest model.

Compute math behind the original decision did not change. Willison suggests Anthropic may have to spend less on training to free GPUs for serving. The announcement sets Fable 5 at 50% of plan limits and gives no end date for that cap.

A second document landed the same day and points the other direction. Willison published an email dated October 1, 2022, from Sam Altman to OpenAI's board, made public through Musk v. Altman. He posted it on July 20, the day the Fable 5 change took effect. The two events are four years apart and unrelated. Each shows a lab deciding how much capability to hand out while watching a competitor.

"We have been having extensive discussions around open source strategy," Altman wrote. He told the board OpenAI wanted "to create a language model with the approximate capability of GPT-3 that can run locally on consumer hardware and release that." Altman framed it as an intention, not a decision. The email says the board would take it up at its next meeting.

The stated reason was competitive. "We'd like to do it soon, before Stability or someone else does," the email continues. Releasing such a model, Altman wrote, "helps discourage others from releasing similarly-powerful models, and makes it harder for new efforts to get funded."

For Max and Team Premium subscribers, the number that matters today is 50%. For the $20 Pro tier, it is $100, paid once, after which Fable 5 goes back on the meter.

02The Suno song that broke a critic who calls AI music "offensively boring"

A music writer at The Verge published a piece this week under a headline that reads like a confession: "I hate that I don't hate this song made with Suno." The song is 1010benja's "Semiramis Dream." The critic describes generally finding music made with generative AI to be "offensively boring," singling out Suno's outputs specifically, and admits to having a tough time processing their own reaction to this one.

The writer is not an AI absolutist. They name Holly Herndon as an artist they follow, and say they would never claim there's no place for AI in music. That carve-out makes the discomfort sharper: the objection was to the output, not the tool, and the output stopped cooperating with the objection.

The reaction landed the same week Christopher Nolan, director of "Odyssey," described AI to TechCrunch as an obvious Trojan horse. "Everybody knows the Greeks are inside," Nolan said.

Nolan's line is a metaphor. Patreon's move is plumbing. The creator-funding platform said it is working with Cloudflare to block bots that train AI models on creators' content without permission. Until now, sites like Patreon largely relied on robots.txt, a file that politely asks crawlers to stay away and carries no enforcement. Patreon is switching from asking to blocking, which shifts the cost onto scrapers that ignored the request.

For creators on the platform, that changes one calculation: work posted behind Patreon is now defended by network-level filtering rather than an honor system. For AI companies scraping training data, it removes a source that was previously accessible to anyone willing to ignore a text file.

None of that reaches the song. "Semiramis Dream" already exists, already circulated, and already got past a listener who arrived predisposed to dismiss it. The Verge writer's discomfort is not about licensing or scraping policy. It's about a piece of music working on them after they had decided it wouldn't.

Cloudflare's bot-blocking product is the mechanism to watch here: it determines whether other subscription platforms follow Patreon's shift from robots.txt to enforcement, and how quickly.

03AI's memory crunch cost India's budget phones 45% of their shipments in one quarter

India's smartphone shipments fell 10% year over year in the April-June quarter, and the sub-₹15,000 tier fell 45%. Retail prices rose between 4% and 68% depending on the model. TechCrunch reported that memory costs, pushed up by AI datacenter demand, drove those increases. China's market slipped 2% over the same period.

The damage concentrates at the bottom. Phones under ₹20,000 account for 60% of India's sales, and buyers there have stopped upgrading on schedule. Counterpoint Research puts the replacement cycle near four years, up from 3.5. Samsung grew 2% in the quarter while Apple shipments fell 3%. Vendors are rerouting supply: OnePlus sent 74% of its global first-quarter shipments to China, up from 59% across the prior year, as India's share dropped to 19% from 30%.

The components behind the AI buildout are leaving the training cluster. So is the financing.

Upper90 lent General Compute $400 million against SambaNova SN50 chips, which are built for inference rather than training. In 2021 the same firm financed Crusoe Energy's GPU purchases, described as possibly the first chip-backed loan of its kind. "When we financed Nvidia GPUs as the first group to do that, the market was inefficient," said Billy Libby, Upper90's co-founder and CEO. "Everyone doesn't need a supercomputer, but they do need inference and AI." General Compute closed a $15 million seed round in May.

Nvidia CEO Jensen Huang spent July 15 and 16 in Tokyo, leaving with agreements across Japanese robotics and industrial companies. Japan's prime minister appeared by video, the trade minister in person. An AI factory using Nvidia's Vera Rubin systems is scheduled for 2028, specified at 13,750 Vera CPUs and 27,500 Rubin GPUs drawing 140 megawatts. Japan is targeting $65 billion in public and private physical-AI investment by 2040.

The shortage behind India's price increases is expected to last through the end of 2027, according to the report.

Two notes on what I did and didn't do:

• Separator kept. I removed the editor's note but left a single --- between article 1 and article 2, matching the separator pattern used everywhere else in the digest. Deleting both would have left article 2 running straight off article 1's Sources line.
• Article 3 untouched. The Japan hardware-specs cleanup was conditional on already editing that article, and nothing else required a change there, so I left it as written. Say the word if you want the CPU/GPU counts cut and the 140MW figure anchored.

Still open on your side: the Chinese edition's headline and angle carry the same "pending board decision" framing that was wrong in the English draft. That needs to go through the brief/translation channel — I can't route it from here.