Best of your X follows: Fable returns, Sonnet costs, and agent loops

The strongest signal today is not another raw capability claim. It is access control becoming part of the product surface: the same model can come back online while specific task classes get rerouted, filtered, or slowed down.

Source scope: this issue uses the channel's configured public AI/tech X account set. The scan window is the period since yesterday's issue, ending July 1 at 18:00 UTC.

Anthropic said the U.S. Department of Commerce lifted export controls on Claude Fable 5 and Mythos 5, and that it would begin restoring access the next day 1. Its follow-up said Fable 5 would return globally with new cybersecurity classifiers; some routine coding and debugging tasks may fall back to Opus 4.8 while those classifiers are tuned 2. The practical read: availability is no longer just a yes/no switch. It can mean model access plus task-level routing, fallback behavior, and government testing agreements.

Anthropic's follow-up is the clearest visual break in the set:

Simon Willison flagged that Claude Sonnet 5 keeps Sonnet 4.6's nominal price, but its new tokenizer produces more tokens for the same text 3. His examples: the English Universal Declaration of Human Rights rose from 2,356 to 3,341 tokens, Spanish from 3,572 to 4,747, and a Python file from 44,014 to 56,113 3. For English, Spanish, and code-heavy workloads, that behaves like a 27-42% cost increase even when the per-token price looks unchanged.

Willison's X post is the short version; the linked post has the measurements:

Ethan Mollick argued that teams need to benchmark models against their own use case, especially when judgments and decisions stack on top of one another 4. His example was concrete: a standard benchmark will not tell you whether Gemini 3.1 is less worried about financial losses at a cafe than GPT-5.5 4. If a workflow chains decisions, test the chain. A single model score will miss the error amplification that appears only after several steps.

François Chollet pointed to Bloome as an example of cross-agent feedback loops: Claude, ChatGPT, Gemini, and human teammates working in one shared workspace 5. The specific pattern is one agent drafting, another critiquing, and another catching missing details while humans stay in the same thread 5. This is a useful design clue for AI tools: the next interface may matter less because it has one stronger agent, and more because it makes several imperfect agents check each other.

Chollet said the current AI wave will not lead to mass unemployment, and that its labor-market effect should be minimal except for increasing demand for software engineers 6. The post is a claim, not a labor study. It is still useful because it states the optimistic software-demand thesis without hedging.

For builders, the question is testable: if AI increases software demand, the bottleneck moves toward people who can specify, review, integrate, and maintain software output.

Paul Graham's verified account relayed an investor anecdote: one founder started fundraising calls at 9:00 and was done by 10:30 7. The post gives no company name, round size, or investor list, so it should be treated as a speed signal rather than fundraising data. Still, it captures a founder-market dynamic worth watching: when investor demand is high, the fundraising process can compress into one morning.