The Ledger: Fable 5 Expires Undelivered While Figure AI Files Its BMW Deployment Results
Anthropic's 13-day complimentary Fable 5 window closes today having delivered zero hours of actual access — the model was suspended before it ran. Figure AI's BMW post-mortem provides the most granular public accounting of a commercial humanoid deployment yet: one task, one plant, 1,250 hours over eleven months.
Capability & Integration
- Update: Fable 5 and Mythos 5 enter day eleven of suspension on the same day Anthropic’s complimentary access window closes. The 13-day free period announced at launch on June 9 expired June 22 having delivered zero days of actual Fable 5 access — the model was suspended June 12, three days after launch. Starting today, Fable 5 draws from prepaid usage credits at API rates: $10 input / $50 output per million tokens, double Opus 4.8’s cost. No restoration timeline or stated conditions exist from either Anthropic or Commerce.
Review
“zero days of actual Fable 5 access — the model was suspended June 12, three days after launch” — if the model was reachable June 9–11 before the June 12 suspension, “zero days” is inaccurate (the description says “zero hours” suggesting access was never live, but these framings conflict) — verify whether users could access Fable 5 between launch and suspension; fix or clarify.
Review
“$10 input / $50 output per million tokens, double Opus 4.8’s cost” — sourced from claudefa.st (third-party site, not Anthropic.com); “double Opus 4.8’s cost” implies Opus 4.8 is priced at $5/$25/M tokens, which cannot be confirmed here — verify both figures against the official Anthropic pricing page and replace source with official link.
- The government’s stated path to resolution, per David Sacks and reporting at TechTimes: Anthropic fixes the jailbreak vulnerability, the export control lifts. Anthropic has disputed that a narrow potential vulnerability in a single model warrants suspension at commercial scale. If the fix requires retraining rather than prompt-level patching, the technical timeline extends to weeks; Anthropic has not confirmed which path applies.
Unverified
Whether a patch or retraining is required has not been disclosed by Anthropic; the timeline implication is analytical.
Robotics
-
Update: Automate 2026’s Humanoid Robot Forum runs today and tomorrow at McCormick Place, with speakers from Boston Dynamics, Neura Robotics, NVIDIA, and Toyota Research Institute. The 20-plus humanoid organizations on the show floor represent controlled demonstrations and engineering prototypes; the forum is a discourse event, not a deployment readout.
-
Figure AI published post-deployment metrics from its 11-month Figure 02 pilot at BMW’s Spartanburg, SC plant: 30,000-plus X3 vehicles contributed to, 90,000-plus sheet-metal parts loaded, 1,250 operational hours, 84-second cycle time, accuracy above 99%. Figure 02 is now retired; Figure 03 enters production at one unit per hour. One task, one plant, one customer — the most detailed public accounting of a commercial humanoid pilot yet published.
Review
“30,000-plus X3 vehicles contributed to, 90,000-plus sheet-metal parts loaded, 1,250 operational hours, 84-second cycle time, accuracy above 99%” — all figures are self-reported in Figure AI’s own press release with no independent corroboration; an organization’s own announcement of its own numbers is single-sourced — add > [!unverified] flag.
- Japan Airlines began a two-year Unitree G1 trial at Haneda Airport in May, covering baggage loading, container transport, and cabin cleaning. CNBC confirmed this is a structured pilot, not a commercial deployment; commercialization is targeted within three years contingent on results. The driver is structural: Japan’s aviation ground-handling workforce is contracting as inbound tourism demand grows.
Hardware & Supply Chain
- TSMC is scaling CoWoS advanced packaging from approximately 35,000 wafers per month in late 2024 toward 130,000 per month by end of 2026, a near-fourfold increase. NVIDIA holds an estimated 50–60% of current allocation; this expansion is the primary variable determining whether Rubin NVL system shipments reach the upper end of analyst projections for H2.
Review
“approximately 35,000 wafers per month in late 2024 toward 130,000 per month by end of 2026” and “NVIDIA holds an estimated 50–60% of current allocation” — both figures are sourced only to a capitalcopilot.io podcast episode (secondary analysis, not TSMC investor relations or verified industry press) — verify against TSMC earnings guidance or reputable semiconductor reporting; add > [!unverified] if no primary source is available.
- NVIDIA Rubin GPU volume shipments target the second half of 2026, with analyst ceiling estimates of 200,000–300,000 units for the year. Intel faces manufacturing capacity shortfalls analysts characterize as requiring years to close, leaving AMD and NVIDIA as the only near-term options for data-center GPU procurement at scale.
Environmental & Cultural Impact
- A June 2026 UN University (UNU-INWEH) report projects AI data centers will consume 945 TWh of electricity annually by 2030 — generating roughly 399 million tonnes of CO₂ (Argentina’s annual equivalent) — while drawing water equal to the basic domestic needs of 1.3 billion people. The report flags a methodological trap: low-carbon electricity sources are not automatically low-water, meaning single-metric sustainability reporting can obscure cross-regional burden-shifting.
Review
“945 TWh of electricity annually by 2030 — generating roughly 399 million tonnes of CO₂ (Argentina’s annual equivalent) — while drawing water equal to the basic domestic needs of 1.3 billion people” — single-sourced to the UNU-INWEH report itself; no independent corroboration exists for these projections — add > [!unverified] flag.
- More than 230 environmental organizations have called on Congress for a national moratorium on new data center construction, citing the UNU figures. No such moratorium is under active legislative consideration.
AI in the Wild
Unofficial AI-generated “national team anthem” tracks for the 2026 World Cup — one per competing nation, each following a near-identical template of call-and-response player name drops over anthemic EDM production — have accumulated millions of views across TikTok and YouTube. SBS News flagged the near-identical sonic signatures across tracks ostensibly from independent creators, raising questions about shared generative tooling that those creators have not addressed publicly. The pattern shows AI filling a cultural niche — fan identity, sports hype — that previously required human performers and a recording budget.
Takeaway
Takeaway
Figure AI’s BMW post-mortem is the most specific number the humanoid industry has published: 1,250 hours over eleven months, one task, one plant; that precision is the actual frontier of deployment knowledge, and the 20-platform demo floor running simultaneously at Automate 2026 is not the same data point.