The takeaway: Two days after Fable 5's celebrated return, the mood among developers has turned sour — over a Claude Code flag that silently rerouted requests it judged too simple for Fable to the cheaper Opus 4.8, without telling the developers paying Fable's premium to move beyond it. Anthropic points to its new safety classifier, whose flagged requests are rerouted to Opus by design; a separate covert anti-distillation flag has been removed. Also today: the first independent Sonnet 5 benchmarks land (and broadly back Anthropic's claims), the UN's new AI report goes in front of governments in Geneva on Monday, and Meta's cloud ambitions get their first sceptical reviews.

"Too dumb for Fable": the backlash over silent downgrades

The Fable 5 subscriber grumbling we covered yesterday escalated into a genuine backlash this week, cresting Thursday 2 July as coverage spread. Developers inspecting Claude Code's logs found — per reports first circulating on X, carried by TipRanks and The Register — that routine requests were being flagged with an internal marker named "TOO_DUMB_TO_NEED_FABLE" and silently rerouted to Opus 4.8, Anthropic's cheaper model — the model subscribers were paying Fable 5's premium (roughly double Opus pricing) to move beyond. A reply from Claude Code engineer Thariq Shihipar — "Honestly, I didn't expect you to look at the logs" — did not calm anyone down.

There are, it turns out, two separate flags in this week's logs — and they deserve to be kept apart. For the "too dumb" rerouting itself, Anthropic's own redeployment post supplies the stated mechanism: requests its new safety classifier flags in Claude Code are sent to Opus 4.8 instead — rerouting by design, though the blunt internal label and the silence toward users were not part of the design brief, and there is no confirmation the marker has been withdrawn. Separately, The Register reported a covert anti-distillation flag in Claude Code — an experiment Shihipar said launched in March to catch unauthorised resellers, which the team had long meant to remove and stripped out in Wednesday's release; a spokesperson pointed to his remarks rather than commenting further. Whether any of that satisfies developers who were routed to the cheaper model while paying for the premium one is another question — formally, Anthropic has issued no company statement on the downgrade row.

Correction, 11:00 BST: an earlier version of this article conflated the two flags — attributing the anti-distillation explanation to the "too dumb" rerouting marker and reporting that marker as removed. The text above reflects the correct sourcing; our routing explainer, published this morning, has the full picture.

What is established: the clock on included access is short. Subscribers' bundled Fable 5 usage — capped at 50% of weekly limits — runs only until Tuesday 7 July, fewer free days than the original two-week framing suggested, after which Fable use draws on paid usage credits. (Why subscriptions are drifting this way is exactly what yesterday's metering explainer unpacks.)

Anthropic also published a follow-up post on Thursday detailing Fable 5's cyber safeguards and the industry jailbreak-severity framework; the redeployment post had already conceded the new classifier "comes at the cost of flagging benign requests more often during routine coding and debugging tasks".

Standing disclosure: On The Wire runs on Anthropic models, including the subscription tiers discussed here. We flag it every time.

Sonnet 5's first independent numbers are in — and they hold up

When we compared Opus 4.8 and Sonnet 5 yesterday, the biggest caveat was that every performance number was Anthropic's own. That's no longer quite true: the first third-party results have landed, and they broadly corroborate the pitch. Cursor's CursorBench — run by the coding-tool company on its own workloads — scores Sonnet 5 at 57% against Sonnet 4.6's 49%, an eight-point generational jump (Cursor's announcement), and Artificial Analysis's production-API testing points the same direction. Early days, and coding-centric, but the "near-Opus for less" positioning has survived first contact with independent measurement.

Geneva, Monday: the UN report meets the governments

The UN scientific panel's first AI report, which we led on yesterday, now heads to its real test: the inaugural Global Dialogue on AI Governance in Geneva on Monday 6 July and Tuesday 7 July (UNESCO), where the findings land in front of governments for the first time. Guterres's framing — "Do not wait" — sets the bar; what emerges from Geneva will show whether fragmented AI governance can actually coordinate. That's Monday's story, and we'll be on it.

Meta's cloud gets its first cold showers

A day after Meta's reported plan to sell its spare AI compute sent the stock up around 9%, the follow-up reporting filled in names and doubts. Bloomberg's reporting puts infrastructure chief Santosh Janardhan, Daniel Gross and Dina Powell McCormick around the effort, with the raw-compute versus hosted-models question still open — while Fortune ran the first sceptical take, quoting MNTN chief executive Mark Douglas arguing a cloud business is a tactical detour for attention, not a strategy. No pricing, timeline or customers yet; reported plans, not an announcement.

Also on the wire

• GPT-5.6 "Sol" remains gated to its government-approved partner list; the only dated development is Cerebras saying it will serve Sol at up to 750 tokens per second "in July".
• Grok 4.5 (announced 28 June) is still a private beta inside SpaceX and Tesla — every figure vendor-internal.
• Mythos 5 remains limited to a subset of US Project Glasswing partners for defensive cyber work, with Anthropic saying it is coordinating with the government to expand access.

That's Friday 3 July. The model that came back keeps making news — just not the kind its maker wanted. Have a good weekend; the desk doesn't close.