The takeaway: Claude Fable 5 was supposed to start costing money at 8am this morning. It didn't — because on Sunday evening UK time, hours before the deadline, Anthropic extended free access a second time, to July 19 (that's 07:59:59 BST on Monday 20 July for UK users). No reason was given, again. Also today: a measurement claiming Claude Code burns nearly five times the tokens of a leaner rival before you've typed a word; day 17 of the OpenAI 516 bug with zero staff replies; Apple's trade-secrets suit gets a case name; and MIT unveils a way to detect CSAM-capable models without generating a single image.

The second stay of execution

The change appeared quietly on Anthropic's support page on Sunday evening: "We've extended this promotion through July 19, 2026 at 11:59:59 PM PT." The company confirmed it on X: "We're extending Claude Fable 5 access on all paid plans, as well as keeping Claude Code's weekly rate limits 50% higher, through July 19."

This is the second last-minute extension in a week — the deadline has moved from July 7 to July 12 to July 19, a sequence BleepingComputer's reporting also lays out — though one Hacker News commenter calls it "the third extension already". What hasn't changed: eligibility (Pro, Max, Team and premium Enterprise seats), the 50%-of-weekly-limits allowance, and the total absence of an explanation. Neither the support page nor the X post gives a reason, and BleepingComputer's report carries no spokesperson quote.

The best available context comes from Simon Willison, who notes that Anthropic's original rationale was compute constraints — wanting a clearer picture of demand before committing to keeping the model cheap for subscribers. His own view has hardened: "At this point I think Anthropic should change track and keep Fable permanently available on those plans. OpenAI are winning users simply due to the uncertainty that surrounds Fable access."

That uncertainty is the theme on Hacker News, where the reaction is less relief than exasperation. "Can Anthropic please just decide on what their plan is with Fable instead of kicking the can down the road …?" asks one commenter; another had "been working non-stop all weekend to use Fable before it disappeared tomorrow."

Our Fable cliff guide — updated within the hour on Sunday — has the full decision arithmetic for whenever the meter genuinely starts. On the pattern so far, pencil in next Sunday evening for the next episode.

What your harness costs before you say hello

Fitting, then, that the weekend's most-discussed technical post is about paying for tokens you didn't know you were sending. A measurement study from Systima put a logging proxy between two coding harnesses and the API and found that Claude Code sends roughly 32,800 tokens on its first turn before your prompt — system instructions, tool definitions, machinery — against about 6,900 for the open-source OpenCode: 4.7× on Sonnet 4.5, narrowing to 3.3× on Fable 5. A 72KB instructions file added another ~20,000 tokens to every request, and mid-session rewrites of Claude Code's ~43k prefix produced up to 54× more cache-write tokens on identical tasks.

The authors are candid about the limits — one machine, three runs per task, a local gateway in the measurement path, and "a July 2026 snapshot" of prompts that change frequently — and both harnesses completed the tasks identically, so the gap is pure cost, not quality. The practical lesson is the one from our prompt-caching explainer: stable prefixes are the whole game, and a harness that rewrites its own preamble mid-session is quietly billing you for it.

Day 17 of the 516 bug — and the community's forensics get sharper

The OpenAI Codex truncation bug enters day 17 still open: 183 comments and 422 reactions as of this morning, with zero replies from OpenAI staff since it was filed on 27 June. The thread has stopped waiting and started measuring. A wire-level proxy analysis posted in-thread finds 64.4% of rounds with ≥516 reasoning tokens landing exactly on suspicious boundary values, extending the 518·n−2 pattern to n=21; a community benchmark puts an unmitigated model at 3/20 on a task it scores 18/20 on with a workaround; and turn-final boundary stops have climbed from 7.9% in April to 39% in July. All community-sourced, none of it confirmed — because confirming or denying would require OpenAI to reply.

The rest of the wire

Apple v. OpenAI has a case name: Apple Inc. v. Liu. The trade-secrets suit filed on Friday is docketed as 5:26-cv-07078 in the Northern District of California, naming engineer Chang Liu, OpenAI Foundation, OpenAI Group PBC, former Apple executive Tang Yew Tan and io Products LLC as defendants, per the court docket. The preliminary-injunction motion Apple told reporters it would file "promptly" (CNBC) hadn't appeared on the docket as of early this morning.

MIT can now spot a CSAM-tuned model without generating anything. Reports of AI-generated child sexual abuse material to the US National Center for Missing and Exploited Children hit 1.5 million in 2025, up from 67,000 in 2024 — and auditing image models for the capability has been legally impossible, since producing test outputs is itself a crime. A team from MIT, Boston University and the child-safety nonprofit Thorn today unveiled "Gaussian probing": feeding random noise through a model's fine-tuning modifications and reading the internal signature of what it was adapted to do. In tests it identified LoRA-modified CSAM-capable models with 100% accuracy — no prompt, no image. "This unlocks a new avenue for platforms that host open-source models and for law enforcement to actually test whether a model is capable of generating CSAM," says lead author Vinith Suriyakumar. The work is a spotlight paper at an ICML workshop on trustworthy AI.

On the watch list: the World AI Conference opens in Shanghai on Friday; DeepSeek's legacy endpoints retire on 24 July, with V4's formal release expected mid-month; and OpenAI's cycle-double-cover proof still has just one qualified mathematician's endorsement on record.