Correction notice — published after this print, disclosing a fact that does not change dated_siu:

3 corrections: 1 pricing, 1 attribution, 1 other

• 26 Sep 2026 — This print is correctly computed under the current methodology and passed every gate at cut-off; it is not being revised. The drop from 2026-09-25's $0.0111 to $0.007626 reflects a change in which constituents had real run records today, not market movement: gemini-3.1-pro-preview — the single most expensive constituent yesterday at $0.039236/basket — produced zero run records today because of a real Google API billing outage ("prepayment credits depleted"), while claude-haiku-4-5, excluded yesterday for the same reason (zero run records), returned with real, qualifying runs. The site's "registry changed" marker correctly did not fire for today's print: both models remain registered constituents throughout; only their per-print qualifying status changed, which computeConstituentChanges is deliberately scoped to ignore (see methodology.md's registry inclusion policy) so a bad day is never confused with an actual registry removal. That said, this print's own event exposes a real, known methodology gap: the pipeline currently treats a constituent with no run records on a given day as simply absent from the equal-weighted mean, with no link ratio bridging the day it drops out and the day it returns — a provider outage is currently indistinguishable, in the computed number, from an unannounced rebalance. A fix (separating per-print availability from registry membership, with capped carry-forward imputation and a chain-linked rebalance) is planned, not yet implemented. For reference only, and never a replacement for the signed value above: recomputing today's equal-weighted mean with gemini-3.1-pro-preview's last valid measured cost ($0.039236, from 2026-09-25) carried forward in place of its missing measurement gives $0.010787 — closer to the trend this print's own drop otherwise misrepresents as a market move.
• 26 Sep 2026 — Correction to this print's own correction note above: that note described gemini-3.1-pro-preview's zero run records as caused by "a real Google API billing outage." That mischaracterises whose failure this was. The actual cause was this project's own Google Cloud billing account lapsing — prepayment credits on our account were depleted — not any outage, degradation, or failure on Google's side. Google's API was reachable and functioning normally throughout; the requests never had valid billing behind them because of a lapse on our side. Same underlying fact the original note already quoted correctly ("prepayment credits depleted") — only the surrounding framing was wrong, and is corrected here. Separately: the methodology gap this print's own note called "planned, not yet implemented" has since been implemented — a model missing today's run records now carries its last real, published cost forward for up to 3 calendar days before being excluded, and every print now also publishes an unweighted median as a diagnostic alongside the mean. See docs/methodology.md's Aggregation section (rules-2026-09-26) for the full retro-validated evidence, including this print's own $0.010787 carry-forward figure, computed identically by the shipped implementation.
• 27 Sep 2026 — Follow-up to this print's second correction note above, which stated a carry-forward figure of /bin/bash.010787 for gemini-3.1-pro-preview — that figure reused 2026-09-25's own already-published cost (/bin/bash.039236) verbatim, which is not the method actually adopted. The shipped implementation (packages/print/src/compute/index.ts) carries forward the last valid measured USAGE and reprices it at the current print's own price snapshot, never the source day's already-computed cost. Recomputing this print under that adopted method, using the real production code path, gives gemini-3.1-pro-preview a carried-forward cost of /bin/bash.040218 (not /bin/bash.039236) and a resulting dated_siu of /bin/bash.010885 (not /bin/bash.010787) — still for reference only, never a replacement for the signed value above. The figures differ because CACHED_INPUT_PRICING_EFFECTIVE_DATE (packages/print/src/cli/historical-pricing-rules.ts) is 2026-09-26 itself, one day after the carry-forward source (2026-09-25): repricing that day's real cached-token usage under 2026-09-26's own rules correctly prices cached-input tokens the source day's own published cost never did. See docs/methodology.md's Aggregation section (rules-2026-09-26) for the corrected retro-validation table and the full explanation.

Published late — 1 earlier attempt failed before this print was produced:

• 26 Sep 2026 at 04:32 UTC — Run failed before publish.log had a diagnostic message — see the linked run's full log. (0 of 0 qualified, $0 spent)

▼ -31.3% since 25 Sep 2026 — Cheaper than the previous print. Falling prices are the expected direction for this index over time.

Equal weighting, disclosed: basket costs span 35.8× across the 9 qualifying constituents. 4 of 9 constituents (44%) sit at or above the mean basket cost and together account for 84% of the summed basket cost Dated SIU is computed from — equal weighting still gives each of them the same 1÷9 nominal share. Dated SIU therefore tracks the more expensive end of the basket more closely than a demand-weighted index would. See the methodology's Weighting section, and Frontier SIU / Commodity SIU, the tier prints published alongside this one for exactly this reason.