Independent/Reader-funded/Infrastructure, not tokens
DeAINEWS

AI you control — open models, private inference, and the networks that run them.

Industry & Business

Price Index Delta: DeepSeek Doubles, Kimi K3 Splits, gpt-oss Reverts

OpenRouter list-price delta, 2026-10-05: DeepSeek V4.1 Flash doubled to $0.30/$1.20 per million tokens; Kimi K3 split its spread; gpt-oss-120b reverted.

DeAI is powered by Morpheus (mor.org). We cover competing providers on the same terms — see our methodology.

Two rows of price tags on a hardware-store pegboard, one tag flipped over showing a higher number on its reverse — the week's split between falling input and rising output list prices. Illustration: DeAI
Two rows of price tags on a hardware-store pegboard, one tag flipped over showing a higher number on its reverse — the week's split between falling input and rising output list prices. Illustration: DeAI

The fourth DeAI Price Index observation is dated 2026-10-05 UTC, two weeks after issue three (no observation ran on 2026-09-28 — see Background). Comparing OpenRouter's public catalog against the 2026-09-21 snapshot, the headline move is deepseek-v4.1-flash, up 100% to $0.30 input / $1.20 output per million tokens — the aggregator row now lists DeepSeek's official peak rate — while Moonshot AI's Kimi K3 split its spread: input down 61% to $0.67, output up 65% to $14.00.

Key facts

  • DeepSeek V4.1 Flash on OpenRouter: $0.15/$0.60 → $0.30/$1.20 (+100%) — the row moved from the off-peak rate to the official peak rate, which is unchanged. Fireworks separately raised its own DeepSeek V4.1 Flash serverless pricing Oct 1, nearly doubling output cost.
  • Kimi K3 on OpenRouter: $1.70/$8.50 → $0.67/$14.00 (input -61%, output +65%) — a 21× input/output spread, observed 2026-10-05 on the OpenRouter models API. This is an aggregator row, not an official Moonshot price.
  • DeepSeek V4 Pro (0423) on OpenRouter: $0.9553/$1.9105 → $0.2088/$0.4176 (-78%) — while the sibling deepseek-v4-pro-0813 row lists $4.50/$5.00 and the ~deepseek/deepseek-pro-latest alias $1.901/$4.20.
  • gpt-oss-120b: $0.15/$0.60 → $0.037/$0.17 — reverted to the value observed on 2026-09-07 and 2026-09-14, confirming the 2026-09-21 reading flagged in issue three as a catalog data revision, not a price move.
  • GLM-5.3 Flash: $0.09/$0.30 → $0.15/$0.50 (+67%) — back to its 2026-09-14 level; full-size GLM-5.3 blew out to $0.05 in / $7.00 out (input -95%, output +145%).
  • Hermes 4 70B is still absent from the OpenRouter catalog — the third consecutive observation without it. nousresearch/hermes-4-405b remains listed.

What happened

Ten of the seventeen tracked rows moved week-over-week against the 2026-09-21 baseline, and for the first time in four observations the biggest mover is an increase. The deepseek-v4.1-flash row doubled to $0.30/$1.20. That figure is not new: it is exactly DeepSeek's official peak cache-miss rate, re-verified unchanged on DeepSeek's own pricing page today (off-peak $0.15/$0.60, peak $0.30/$1.20, cache-hit input $0.003 off-peak / $0.006 peak, peak hours 01:00–04:00 and 06:00–10:00 UTC Monday–Friday). Issue three caught the row listing off-peak pricing on a Monday; issue four catches it listing peak pricing. The practical read for builders: this row tracks the peak/off-peak schedule, and which rate you see depends on when you look.

The timing is notable. On October 1, Fireworks raised its own DeepSeek V4.1 Flash serverless pricing — Standard tier from $0.22 input / $0.66 output to $0.30 / $1.20 per million tokens, nearly doubling output cost, "to bring our pricing in line with current market rates." OpenRouter's row now lists the same $0.30/$1.20 figures. We cannot attribute intent from a catalog row, but two independent surfaces moving to identical numbers within four days is the closest thing to a market repricing this basket has produced.

The Kimi K3 row tells the week's second story: a scissors move. Input fell 61% to $0.67 while output rose 65% to $14.00 — a 21× input/output spread, the widest in the basket. Issue three recorded Kimi K3's 36% cut; this week's shape suggests the aggregator's backend mix changed rather than Moonshot rewriting a price card. At $0.67 input it is now cheaper than Claude Sonnet 4.6 ($3.00/$15.00) on input but more expensive on output; the cheapest open-weight input row is now deepseek/deepseek-v4-flash-0731 at $0.0152.

The gpt-oss-120b flag, resolved

The data-quality flag from issue three is now closed. openai/gpt-oss-120b listed $0.037/$0.17 on 2026-09-07 and 2026-09-14, $0.15/$0.60 on 2026-09-21, and back to $0.037/$0.17 today — exactly the reversion the flag anticipated. The 2026-09-21 reading was a catalog artifact, and the index's decision to record and flag rather than average it away is what makes the series interpretable. Downstream derivatives should treat the 2026-09-21 gpt-oss row as noise.

Why it matters

  • Peak/off-peak now visibly oscillates in aggregator rows. A builder pricing a workload off the OpenRouter catalog on a Monday morning will read a different rate than one checking Friday night. For DeepSeek V4.1 Flash the swing is 2×, and cache-hit input ($0.003–$0.006) is 25–50× cheaper than cache-miss — time-of-day and cache-hit ratio now swing effective cost more than any list move this week.
  • Fireworks' hike may be the first of a wave. The week-ahead brief flags Fireworks' Oct 1 move as the pricing baseline other providers may follow. If OpenRouter's deepseek rows track official peak rates, the cheapest frontier-quality tier just doubled for peak-hour workloads.
  • Aggregator rows remain unusable as official prices. Three DeepSeek "Pro" routes now list $0.2088/$0.4176 (0423), $4.50/$5.00 (0813) and $1.901/$4.20 (alias) simultaneously — a 21× spread under one brand. The checkpoint-portability rule from issue one applies: verify against the first-party page before committing spend.

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 two covered the DeepSeek V4.1 Flash launch week; issue three caught the catalog moving down.

One process note: there was no 2026-09-28 observation — the scheduled Monday run did not produce a snapshot that week, so this issue compares against 2026-09-21 across a two-week gap. That gap makes the DeepSeek doubling harder to date precisely: it could have landed any time between September 22 and October 5. This is recorded rather than smoothed over. Also 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}_2026-10-05.json), so the snapshot was fetched directly for this job using the documented Issue-1 methodology — the third week running without a wired OpenRouter adapter. The raw pull is dated and archived (data/tracker/pricing_fetch_2026-10-05.json); nothing here derives from provider marketing claims.

What's next

The next observation is scheduled for Monday, 2026-10-12. Watch items: whether the deepseek-v4.1-flash row flips back to off-peak later in the week (the schedule test), whether Kimi K3's scissors shape holds or normalizes, whether other providers follow Fireworks' hike, whether Hermes 4 70B returns, and whether the DIPT OpenRouter adapter gets wired so the snapshot stops depending on manual fetches. The machine-readable snapshot updates at /data/prices.json alongside this story, with the 2026-09-21 snapshot preserved in the file's archive chain.

Questions

What changed in LLM API prices this week?
On OpenRouter's public catalog, comparing 2026-10-05 to 2026-09-21: deepseek-v4.1-flash doubled to $0.30/$1.20 per million tokens, kimi-k3's input fell 61% to $0.67 while output rose 65% to $14.00, deepseek-v4-pro fell 78% to $0.2088/$0.4176, and gpt-oss-120b reverted to $0.037/$0.17.
Why did DeepSeek V4.1 Flash double in price on OpenRouter?
The OpenRouter deepseek-v4.1-flash row moved from the $0.15/$0.60 off-peak rate it listed on 2026-09-21 to $0.30/$1.20 — DeepSeek's official peak cache-miss rate, unchanged on the official page. Separately, Fireworks raised its own DeepSeek V4.1 Flash serverless pricing Oct 1, nearly doubling output cost. The aggregator row now lists peak pricing.
Was the September 21 gpt-oss-120b price jump real?
No. The $0.15/$0.60 figure observed on 2026-09-21 was flagged at the time as a possible catalog data revision. On 2026-10-05 the row reverted to $0.037/$0.17 — the same value seen on 2026-09-07 and 2026-09-14 — confirming the one-week reading as a data artifact, not a price move.
How much does Kimi K3 cost now?
Moonshot AI's Kimi K3 listed at $0.67 input / $14.00 output per million tokens on OpenRouter on 2026-10-05, versus $1.70/$8.50 a week earlier. Input fell 61%, output rose 65% — a 21x input/output spread. This is the aggregator row, not an official Moonshot price.

Sources

  1. OpenRouter API — list models — OpenRouter
  2. DeepSeek API Docs — Models & Pricing — DeepSeek
  3. Fireworks AI changelog — DeepSeek V4.1 Flash serverless price update (Oct 1, 2026) — Fireworks AI
  4. Open-Model Inference Price Deltas (Issue 3) — DeAI News

About DeAI

DeAI is an independent publication covering open-weight AI models, private inference, and decentralized infrastructure — the tools for running AI you actually control. We test providers on price, privacy, and refusal behavior and publish the numbers, not the vibes. DeAI is powered by Morpheus (mor.org), a decentralized inference marketplace, and covers it on the same terms as every other provider.

Powered by Morpheus and StrandCMS

Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider — we rank it wherever the data lands. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more about the Morpheus Inference API →

Sponsor disclosure — not editorial

Powered by Morpheus and StrandCMS. Morpheus is a decentralized inference marketplace, covered on the same terms as every other provider. StrandCMS is the open-source, agent-first framework this site is built on.

Learn more →