Price Index Delta: Seven Rows Revert in the Basket's Biggest Reset

The sixth DeAI Price Index observation is dated 2026-10-26 UTC, two weeks after issue five. Comparing OpenRouter's public catalog against the 2026-10-12 snapshot, seven of seventeen tracked rows moved — and every single mover reverted toward a level the index had already recorded, making this the basket's largest synchronized reset in six issues. The headline: openai/gpt-oss-120b returned to $0.15/$0.60 per million tokens, the exact row of 2026-09-21, resolving a three-issue low-price artifact cluster. The sharpest mover: deepseek/deepseek-v4-flash-0731, input up 2,650% to $0.44 as its two-week-old low row disappeared.

Key facts

• gpt-oss-120b on OpenRouter: $0.037/$0.17 → $0.15/$0.60 (input +305%, output +253%) — the exact 2026-09-21 row, observed 2026-10-26 on the OpenRouter models API. The two low readings (issue 4, issue 5) are now a confirmed artifact cluster, not a price history.
• DeepSeek V4 Flash 0731: $0.016/$0.64 → $0.44/$1.32 (input +2,650%, output +106%) — back to the $0.44/$1.32 row the basket recorded on 2026-09-21 and 2026-10-05. Issue 5's halved output ($0.64) lasted exactly one observation.
• GLM-5.3 Flash scissored back: $0.0232/$1.20 → $0.15/$0.50 (input +547%, output -58%) — the precise row of 2026-10-05. Three weekly readings, three different shapes; the row has now returned to where the scissoring started.
• DeepSeek V4.1 Flash: $0.30/$1.20 → $0.02/$1.00 (input -93%) — after six Mondays of peak pricing, the row flipped. The $0.02 input sits below the $0.30 peak rate DeepSeek's official pricing page still lists (re-verified 2026-10-05); the catalog row is the observation, not a confirmed official cut.
• Kimi K3: $0.73/$10.00 → $0.152/$20.00 (input -79%, output +100%) — the input/output spread blew out from 14× to 131×, the widest in the basket. Moonshot AI has published no first-party price change.
• GLM-5.3 (full-size): $0.079/$4.80 → $0.0699/$1.40 (output -71%); sibling deepseek-v4-flash (0423) output -86% to $0.182. Ten rows held flat, including both closed anchors (claude-sonnet-4.6 $3.00/$15.00, gpt-5.5 $5.00/$30.00). Hermes 4 70B is absent a sixth consecutive observation.

What happened

Six issues in, the index has accumulated enough history to distinguish a price change from a catalog reset, and this week's diff is almost entirely the latter. Four of the seven movers — gpt-oss-120b, glm-5.3-flash, deepseek-v4-flash-0731, and glm-5.3's output leg — returned to levels the index recorded on 2026-09-21 or 2026-10-05. That is the same signature the index documented on deepseek-v4-pro in issue five: a value returning exactly to a prior observation marks the intervening readings as artifacts. The difference is scale. Issue five resolved one anomalous row; this issue resolved a cluster, with the two-issue gpt-oss-120b low ($0.037/$0.17) closing out at once.

The rows that did not revert are the informative ones. deepseek-v4.1-flash had listed DeepSeek's official $0.30/$1.20 peak rate for six straight Mondays, long past the point where the off-peak schedule test from issue four was expected to flip. This Monday it finally moved — to $0.02/$1.00, a shape that matches no official DeepSeek card: input 93% below peak, output below peak by 17%. A new deepseek-v4.1-flash-1017 checkpoint also appeared in the catalog this cycle at $0.15/$0.60 — DeepSeek's official off-peak rate — so the family's rows are being restructured around the fresh checkpoint. Whether the legacy row's new shape is a real repricing, an off-peak reflection, or another drift artifact is exactly what the two-observation rule is for: re-check next Monday before treating any of it as a cut.

Kimi K3's scissor is the week's genuine widening. Input fell 79% to $0.152 — cheaper than gpt-oss-120b, a closed-source model at 1/13th the input price of Claude Sonnet 4.6 — while output doubled to $20.00, a 131× spread. A summary-heavy agent workload now pays 2× what it did two weeks ago on this row for the same output; a RAG workload pays 1/5th. None of these rows carries a first-party explanation; the index records catalog rows, not intents.

Why it matters

• The "real price" is the mode across observations, not any single Monday. After six issues, gpt-oss-120b's true OpenRouter row is $0.15/$0.60 with three confirming observations — 4× the number a builder would have paid per the last two issues' tables. Anyone who re-priced workloads off issues 4–5 was reading drift. The index's flag-on-first-observation, confirm-on-second discipline now has a third confirmed catch.
• Scissor rows make per-workload pricing decisions unstable at weekly resolution. GLM-5.3 Flash has now posted three different input/output shapes in three weeks ($0.15/$0.50 → $0.0232/$1.20 → $0.15/$0.50); Kimi K3's spread moved from 14× to 131× in one hop. For input-heavy workloads (retrieval, short answers) and output-heavy workloads (long generation, agents) the same model row changed relative economics by an order of magnitude in fourteen days — on an aggregator whose rows the tracker treats as unverified until a first-party page confirms.
• The deepseek-v4.1-flash watch item finally moved — off the official rate card. Six Mondays of peak pricing broke with a row that matches no published DeepSeek card, on the same cycle a new 1017 checkpoint entered the catalog. If the $0.02/$1.00 shape holds next Monday, it is the first sustained below-peak listing in the DeepSeek family this index has observed; if it reverts, it joins the artifact file.

Background

The DeAI Price Index is a weekly, dated, append-only observation of list pricing — it records what a catalog listed on a date, and never rewrites prior issues. Issue one established the basket and methodology: OpenRouter's GET /api/v1/models prompt/completion fields converted to dollars per million tokens, batch rows excluded, plus first-party anchors. Issue three caught the catalog moving down; issues four and five caught increases and resolved single-row anomalies. This issue's cluster reversion — four rows snapping back in one diff — is the strongest evidence yet that OpenRouter catalog drift, not provider pricing policy, is the dominant signal in this series, and that single-week snapshots of this surface need the two-observation rule applied to every row, not just the outliers.

A process note, per the standing gap report: the automated DIPT pricing fetcher again produced no d_pricing candidates in this week's sweep (1 in data/tracker/resolved_2026-10-26.json, a Phala row with no machine-readable pricing source configured), so the snapshot was fetched directly for this job using the documented Issue-1 methodology — the fifth week running without a wired OpenRouter adapter. The raw pull is dated and archived (data/tracker/raw_openrouter_models_2026-10-26.json, 467 models fetched live 2026-10-26 ~13:00 EET; basket extraction in pricing_fetch_2026-10-26.json); nothing here derives from provider marketing claims.

What's next

The next observation is scheduled for Monday, 2026-11-02. Watch items: whether the deepseek-v4.1-flash row holds at $0.02/$1.00 (the two-observation confirmation it needs — the first test of a new shape rather than a reversion), whether the new deepseek-v4.1-flash-1017 checkpoint settles at its $0.15/$0.60 entry row, whether gpt-oss-120b and glm-5.3-flash stay put now that they have reverted, whether kimi-k3's 131× spread narrows or is the new normal, and whether Hermes 4 70B returns to the catalog. The machine-readable snapshot updates at /data/prices.json alongside this story, with the 2026-10-12 snapshot preserved in the file's archive chain.