Touchstone Assay
Dated SIU — SIU-2026a
$0.0049
08 Sep 2026 provisional
Correction notice — published after this print, disclosing a fact that does not change dated_siu:
  • 08 Sep 2026 — claude-sonnet-5 and claude-haiku-4-5 (both frontier-tier) produced zero run records for this print — every attempt across all three task classes failed with the same Anthropic API error, "Your credit balance is too low to access the Anthropic API," a provider-side billing failure, not a quality-gate failure. Both models were reachable and qualified normally in the immediately preceding print (2026-09-07). This print's basket_costs and exchange_rate_table did not disclose this exclusion with a gap row when first published, because the print pipeline (packages/print/src/cli/load-inputs.ts) silently dropped zero-run-record models before the existing excluded_reason mechanism could apply to them — fixed for every print from this one onward. The published $0.0049 figure is the correct equal-weighted average of the seven constituents that did produce real data; it is not a market move. The apparent 33% drop from 2026-09-07's $0.0073 is a mechanical consequence of losing the two most expensive constituents to a provider outage, not a change in inference cost. This incident also exposed a gap in gate design, not only disclosure: MINIMUM_QUALIFYING_MODELS gated this print's overall qualifying count (7 of 9) but not its tier composition, so it published while frontier collapsed to 3 of its 5 registered constituents — below the same minimum Frontier SIU's own standalone gate correctly applied that same day (Frontier SIU declined to publish, 3 of 5, for this exact reason). Under a per-tier version of this gate, applied retroactively for confirmation only: this print would have been refused outright, not merely disclosed after the fact. The gate now applies per-tier as well as overall for every print from this one onward — see methodology.md's Index governance for the fix.
  • 08 Sep 2026 — A reasoning-token cost gap affected this print: gemini-3.1-pro-preview's reasoning tokens (RunRecord.usage.reasoning — real, captured in every run record for this print) were not priced. packages/print/src/compute/class-cost.ts computed cost from usage.output alone. Reasoning tokens are billed by the provider at the output rate — confirmed live, 2026-09-08: Google's own published pricing states "Output price (including thinking tokens)", and a separate live check (xAI's real per-call billed cost) independently confirmed the same billing convention for reasoning models generally. This is not a methodology refinement — the index measures realised cost from executed runs, and the run records already carried this field. Corrected figures, computed from this print's own real run records and price snapshot with the fix applied, not from new data: this print's dated_siu was originally reported as $0.0049; the corrected figure is $0.0060. gemini-3.1-pro-preview's own basket cost was originally reported as $0.013990; the corrected figure is $0.021777. This print itself is not recomputed or replaced — the original stands, per the revision policy's unconditional never-edit guarantee. The fix (packages/print/src/compute/class-cost.ts, cost-of-production.ts) applies forward from the next published print onward, which will show a real, one-time jump in gemini-3.1-pro-preview's cost for this methodological reason, not a market move — see docs/methodology.md's Index governance.
  • 09 Sep 2026 — The reasoning-token pricing fix's effect, anticipated in this print's own correction note above, became visible in the very next print (2026-09-09): gemini-3.1-pro-preview's real basket cost, computed under the fixed formula against that day's own run records and price snapshot, was $0.020406 — close to, but not identical to, this note's own $0.021777 same-day estimate. The difference is real day-over-day price movement (a new price snapshot), not measurement error; both figures are correct for the day they describe. The full three-print picture, stated plainly: 2026-09-07's $0.0073, this print's $0.0049, and 2026-09-09's $0.0089 are each fully explained by an infrastructure failure (the Anthropic billing outage disclosed above), a pricing bug fix (this note's own subject), or a registry composition change (claude-sonnet-5 and claude-haiku-4-5 recovering, grok-4.6 being admitted) — none of it is a change in the underlying market price of inference, and the series does not yet show a usable market trend. A live-computed, ongoing version of this comparison — splitting any future print's move into a compositional share (admitted/removed constituents) and a continuous share (movement among constituents present in both prints) — now renders on every print with a registry change; see docs/methodology.md's Registry inclusion policy.
▼ -32.9% since 07 Sep 2026 — Cheaper than the previous print. Falling prices are the expected direction for this index over time.
Equal weighting, disclosed: basket costs span 21.4× across the 7 qualifying constituents. 3 of 7 constituents (43%) sit at or above the mean basket cost and together account for 83% of the summed basket cost Dated SIU is computed from — equal weighting still gives each of them the same 1÷7 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.

Exchange-rate table

ModelUSD / SIUSpread to indexSIU per $1
deepseek-v3.2 $0.0018 -63% 551.1
llama-3.3-70b-deepinfra $0.0007 -86% 1413.1
mistral-small-3.2-24b-instruct $0.0007 -87% 1531.1
qwen-2.5-72b-instruct $0.0025 -49% 398.8
gpt-5.1 $0.0092 +88% 109.0
gemini-3.1-pro-preview $0.0140 +187% 71.4
gpt-5.4-mini $0.0053 +9% 188.1

Floor & market spread

Not yet published — the floor needs a measured hardware-rental rate for the reference configuration, which this print does not yet include.

Sensitivity

The print recomputed under alternative cache and batch-discount policies, so the effect of those policy choices is stated rather than hidden inside one number.

Policy variantDated SIUDeltaApplies to
cache-enabled: 40% of T2 input at 10% of input price $0.0033 -32.0% deepseek-v3.2, llama-3.3-70b-deepinfra, mistral-small-3.2-24b-instruct, qwen-2.5-72b-instruct, gpt-5.1, gemini-3.1-pro-preview, gpt-5.4-mini
batch-discount: 50% off input and output $0.0024 -50.0% deepseek-v3.2, llama-3.3-70b-deepinfra, mistral-small-3.2-24b-instruct, qwen-2.5-72b-instruct, gpt-5.1, gemini-3.1-pro-preview, gpt-5.4-mini

Methodology, signature & anchor

Methodology
v0-draft
Cost of production
$0.5539
Signature
0x50048f88…9620f0aa
Public key
0x030c598e…e25cb299
Anchor tx
0x24d48777…3ceb0b5a (base-sepolia)
Raw runs
https://github.com/petrkrulis2022/dated-siu-inference/tree/main/data/runs/2026-09-08

Dated SIU is an assessment of qualifying inference work under this basket version, model/provider configuration and quality gate — not a prediction of production agent costs, which vary with workflow design, tool reliability, retry policy and context handling.

Verify this yourself

Don't take this page's word for it — every value below is checkable independently.

Body hash
0x9ef3ab0d…cb2ef788
Recovered signer
0x284ff2f8…d48f845a — matches the TouchstoneAttestation publisher below
TouchstoneAttestation
0xF6070179…93297190 (base-sepolia)
  1. Recompute this print's body hash from the published JSON (JCS-canonicalised, signature fields excluded) and confirm it matches the body hash shown above.
  2. Open the contract link above on Blockscout, go to "Read Contract", call publisher(), and confirm it matches the recovered signer shown above.
  3. On the same "Read Contract" tab, call postedAt(bytes32) with the body hash shown above and confirm it returns a non-zero timestamp.

All prints