Price Index Delta: Kimi K3 Cuts 36%, DeepSeek Reverts to Off-Peak

The third DeAI Price Index observation is dated 2026-09-21 UTC, one week after issue two. Comparing OpenRouter's public catalog against the 2026-09-14 snapshot, the biggest move is Moonshot AI's Kimi K3, down 36% to $1.70 input / $8.50 output per million tokens from $2.65/$13.28 — while DeepSeek's deepseek-v4.1-flash row halved to $0.15/$0.60, which is exactly DeepSeek's official off-peak rate.

Key facts

• Kimi K3 on OpenRouter: $2.6481/$13.2827 → $1.70/$8.50 per million tokens (-36% input, -36% output), observed 2026-09-21 on the OpenRouter models API. This is an aggregator row, not an official Moonshot price.
• DeepSeek V4.1 Flash on OpenRouter: $0.30/$1.20 → $0.15/$0.60 (-50%) — now matching the official off-peak rate. The official page is unchanged since 2026-09-14: $0.15/$0.60 off-peak, $0.30/$1.20 peak.
• GLM-5.3 Flash: $0.15/$0.50 → $0.09/$0.30 (-40%) — giving back last week's doubling (it was $0.075/$0.25 on 2026-09-07).
• DeepSeek V4 Pro (0423) on OpenRouter: $1.60/$3.20 → $0.955/$1.91 (-40%) — back to the official $1.32/$3.96-adjacent level; the official anchor is unchanged.
• gpt-oss-120b: $0.037/$0.17 → $0.15/$0.60 — see the data-quality note below; we flag rather than celebrate this one.
• Hermes 4 70B is still absent from the OpenRouter catalog (absent since 2026-09-07); nousresearch/hermes-4-405b remains listed.

What happened

Six of the sixteen tracked rows moved week-over-week, and for the first time since the index started, the direction is almost uniformly down. The Kimi K3 cut is the headline: at $1.70/$8.50 the Moonshot flagship drops out of the premium band (where it sat beside Claude Sonnet 4.6 at $3/$15) and into mid-pack territory, roughly 12× the price of DeepSeek V4.1 Flash at off-peak rates.

The DeepSeek rows tell a more interesting story than the percentages suggest. Last week the deepseek-v4.1-flash row listed at $0.30/$1.20 — DeepSeek's official peak rate — while the official page also published a $0.15/$0.60 off-peak schedule. Today the aggregator row lists $0.15/$0.60. That is consistent with the row repricing to the off-peak schedule, and worth reading alongside DeepSeek's own page, which is unchanged: off-peak $0.15/$0.60, peak $0.30/$1.20, cache-hit input $0.003–$0.006, with peak hours 01:00–04:00 and 06:00–10:00 UTC Monday–Friday. The deepseek-v4-flash-0731 legacy row also split its spread — input fell to $0.04 while output rose to $0.16 — and the deepseek-v4-pro row gave back all of last week's 67% jump, returning to $0.955/$1.91.

Z.ai's GLM-5.3 Flash retraced to $0.09/$0.30, below its 2026-09-07 level of $0.075/$0.25 but well under last week's $0.15/$0.50. The full-size GLM-5.3 fell 35% to $0.91/$2.86.

The gpt-oss-120b anomaly, flagged

One row moved the "wrong" way and we are not smoothing it over. openai/gpt-oss-120b listed at $0.037/$0.17 in both the 2026-09-07 and 2026-09-14 snapshots; today it lists $0.15/$0.60 — a 4× jump that would make it one of the week's biggest movers. Two things give us pause. First, the new rate exactly matches both the deepseek-v4.1-flash off-peak row and OpenRouter's platform-standard $0.15/$0.60 tier, which is a suspicious coincidence for a distinct model. Second, the raw API value (0.00000015 $/token) is exactly the same figure DeepSeek's V4.1 Flash carries. We cannot distinguish a genuine repricing of the OpenRouter backend mix from a catalog data revision with this week's fetch alone, so we record the observed value and flag it. If next week's snapshot reverts, this was noise; if it holds, the cheapest open-weight row in the basket is no longer the cheapest.

Why it matters

Three weeks of observations now show the same pattern from different directions: aggregator rows swing 30–100% week to week while first-party pages barely move. For a builder, the practical consequences are concrete:

• Kimi K3's window. A 36% cut on a frontier open-weights row is a real signal that Moonshot wants volume — but it is an aggregator rate, so verify against Moonshot's own platform before committing a production workload.
• Off-peak is now visible in aggregator rows. DeepSeek's peak/off-peak split appearing in the v4.1-flash row means time-of-day routing decisions are starting to show up in the catalogs builders actually browse, not just in billing docs.
• Cache rates matter more than headline rates. DeepSeek's cache-hit input at $0.003–$0.006 is 25–50× cheaper than its peak cache-miss input. At these spreads, a workload's cache-hit ratio changes the effective price more than any of this week's list moves.

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 where an official page publishes one. Issue two recorded the DeepSeek V4.1 Flash launch week. One process note for this issue: the automated DIPT pricing fetcher produced no d_pricing candidates in this week's sweep, so — as with issue two — the snapshot was fetched directly for this job using the documented Issue-1 methodology. The raw pull is dated and archived; nothing in this report is derived from provider marketing claims.

The checkpoint-portability rule from issue one still applies and got more tangled this week: OpenRouter now carries both deepseek/deepseek-v4-pro (0423) at $0.955/$1.91 and deepseek/deepseek-v4-pro-0813 at $0.66/$1.98, alongside the ~deepseek/deepseek-pro-latest convenience alias at $0.66/$1.98. Three DeepSeek "Pro" routes, three prices, one billing page.

What's next

The next observation is scheduled for Monday, 2026-09-28. Watch items for that run: whether the gpt-oss-120b row holds at $0.15/$0.60 or reverts (our data-quality flag), whether Kimi K3's cut shows up on Moonshot's own platform page or stays an aggregator phenomenon, whether Hermes 4 70B returns to the catalog, and whether the DIPT pricing fetcher gets its OpenRouter adapter wired so this snapshot stops depending on a manual fetch. The machine-readable snapshot updates at /data/prices.json alongside this story.

FAQ

What changed in LLM API prices this week?

On OpenRouter's public catalog, comparing 2026-09-21 to 2026-09-14: Kimi K3 fell 36% to $1.70/$8.50 per million tokens, deepseek-v4.1-flash halved to $0.15/$0.60, GLM-5.3 Flash fell 40% to $0.09/$0.30, DeepSeek V4 Pro fell 40% back to $0.955/$1.91, and gpt-oss-120b rose to $0.15/$0.60.

How much does Kimi K3 cost now?

Moonshot AI's Kimi K3 listed at $1.70 input / $8.50 output per million tokens on OpenRouter on 2026-09-21, down from $2.65/$13.28 a week earlier — a 36% input / 36% output cut. That is the aggregator row; Moonshot's own platform rate is not published on this page.

Why did gpt-oss-120b jump from $0.037 to $0.15 input?

OpenRouter is an aggregator: each listed rate reflects the backend mix serving that slug on the snapshot date. The $0.037/$0.17 figure was observed on 2026-09-07 and 2026-09-14; on 2026-09-21 the row listed $0.15/$0.60. This may be a data revision rather than a real price move, and we flag it rather than averaging it away.

Is DeepSeek V4.1 Flash cheaper at off-peak hours?

Yes. DeepSeek's official page bills V4.1 Flash (model name deepseek-flash) at $0.15/$0.60 off-peak and $0.30/$1.20 peak per million tokens, with off-peak covering all weekend hours and weekday non-peak windows. On 2026-09-21 OpenRouter's deepseek-v4.1-flash row listed $0.15/$0.60 — the off-peak rate.