Price Index Delta: DeepSeek V4 Pro Snaps Back, GLM-5.3 Flash Scissors

The fifth DeAI Price Index observation is dated 2026-10-12 UTC, one week after issue four. Comparing OpenRouter's public catalog against the 2026-10-05 snapshot, seven of seventeen tracked rows moved. The headline move is deepseek/deepseek-v4-pro, up 358% to $0.95526 input / $1.91052 output per million tokens — a full reversion to its September level that confirms last week's $0.2088/$0.4176 row as the anomaly, not a price cut. The week's scissor award goes to GLM-5.3 Flash: input down 85% to $0.0232, output up 140% to $1.20.

Key facts

• DeepSeek V4 Pro (0423) on OpenRouter: $0.2088/$0.4176 → $0.95526/$1.91052 (+358%) — exactly the 2026-09-21 level, observed 2026-10-12 on the OpenRouter models API. The official DeepSeek page lists only V4 Pro 0813 ($1.32/$3.96 peak, re-verified 2026-10-05) — no 0423 price card exists to cut.
• GLM-5.3 Flash on OpenRouter: $0.15/$0.50 → $0.0232/$1.20 (input -85%, output +140%) — the second consecutive scissor move on this row, which has now repriced in both directions in as many weeks.
• Kimi K3 on OpenRouter: $0.67/$14.00 → $0.73/$10.00 (input +9%, output -29%) — the 21× input/output spread from issue four narrowed to about 14×.
• GLM-5.3 (full-size): $0.05/$7.00 → $0.079/$4.80 (input +58%, output -31%) — partial mean reversion toward the $0.91/$2.86 row of two weeks ago.
• DeepSeek V4 Flash 0731: output $1.28 → $0.64 (-50%); input $0.0152 → $0.016. Sibling 0423 row flat at $0.03/$1.28, and kimi-k2.5 rose 11% on both axes ($0.45/$2.25 → $0.50/$2.50) — the basket's only uniform increase.
• deepseek-v4.1-flash held at $0.30/$1.20 for a second straight Monday — the peak/off-peak schedule test proposed last week did not flip. Hermes 4 70B is absent a fourth consecutive observation.

What happened

The deepseek-v4-pro move is the story, because it is a correction of the index's own prior observation. Two weeks ago the row listed $0.95526/$1.91052; last Monday it listed $0.2088/$0.4176, a 78% collapse that issue four reported as the basket's steepest cut; today it is back to the September figure to four significant digits. A value that returns exactly to its prior level is the signature of a catalog data artifact, the same pattern the index documented and flagged on gpt-oss-120b two issues running. The practical reading: no DeepSeek V4 Pro price change ever took effect. The official DeepSeek pricing page never listed a 0423 checkpoint price at either level — its only V4 Pro card is 0813 at $1.32/$3.96 peak, unchanged since re-verification on 2026-10-05.

GLM-5.3 Flash's scissor is the second consecutive weekly repricing of that row, in opposite directions: last week input rose 67% while output rose 67% in step; this week input collapsed 85% to $0.0232 — the cheapest input in the basket — while output jumped 140% to $1.20, a 52× input/output spread. Kimi K3 moved the opposite way, narrowing its spread as output fell 29% to $10.00. Full-size GLM-5.3 partially mean-reverted from last week's blowout. None of these rows has a published first-party explanation; the index records catalog rows, not intents.

Why it matters

• Issue 4's biggest cut did not survive contact with a second observation. Builders who re-priced workloads onto deepseek/deepseek-v4-pro at $0.2088/$0.4176 last Monday were reading a catalog artifact. The two-observation confirmation rule from the gpt-oss-120b flag keeps proving out: in five issues, every single-reading outlier — the 2026-09-21 gpt-oss spike, last week's V4 Pro collapse — has reverted.
• Aggregator scissor rows are drifting wider, not converging. GLM-5.3 Flash now spreads 52× between input and output; Kimi K3 sits at 14× after narrowing. A workload that is input-heavy (RAG, short answers) and one that is output-heavy (long generation, agents) now face radically different relative economics on the same model row from week to week.
• The peak/off-peak schedule test failed. Issue four predicted the deepseek-v4.1-flash row might flip back to the $0.15/$0.60 off-peak rate later in the week; two Mondays running it has listed the $0.30/$1.20 peak rate. Either the row is now pinned to peak, or the schedule is no longer reflected in this row at all. Either way, a Monday-only index cannot distinguish the two — a sampling-frequency limit worth recording.

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; issue four caught the first increase and resolved the gpt-oss-120b flag. This issue's DeepSeek reversion strengthens the series' core discipline: flag outliers when observed, confirm on a second observation, never average them away.

A process note, per the standing gap report: the automated DIPT pricing fetcher again produced no d_pricing candidates in this week's sweep (0 in data/tracker/{candidates,resolved}_2026-10-12.json), so the snapshot was fetched directly for this job using the documented Issue-1 methodology — the fourth week running without a wired OpenRouter adapter. The raw pull is dated and archived (data/tracker/raw_openrouter_models_2026-10-12.json, 458 models fetched live 2026-10-12 ~13:00 EEST; basket extraction in pricing_fetch_2026-10-12.json); nothing here derives from provider marketing claims.

What's next

The next observation is scheduled for Monday, 2026-10-19. Watch items: whether the deepseek-v4-pro row holds at $0.9553/$1.9105 (the two-observation confirmation it needs), whether glm-5.3-flash's scissor stabilizes or scissors again, whether kimi-k3 output keeps normalizing toward a conventional spread, whether deepseek-v4.1-flash ever shows an off-peak reading again, and whether Hermes 4 70B returns to the catalog. The machine-readable snapshot updates at /data/prices.json alongside this story, with the 2026-10-05 snapshot preserved in the file's archive chain.