Price Index Delta: DeepSeek V4.1 Flash Lands, GLM Flash Doubles

The second DeAI Price Index observation is dated 2026-09-14 UTC, one week after issue one. Comparing OpenRouter's public catalog against the 2026-09-07 snapshot, the biggest move is on DeepSeek rows: the deepseek-v4-flash-0731 listing fell 57% to $0.06 input / $0.12 output per million tokens, while a new deepseek-v4.1-flash row appeared at $0.30 / $1.20 — matching DeepSeek's official peak rate.

Key facts

• DeepSeek V4 Flash 0731 on OpenRouter: $0.14/$0.28 → $0.06/$0.12 per million tokens (-57%), observed 2026-09-14 on the OpenRouter models API. This is an aggregator row, not an official DeepSeek price.
• New row: deepseek/deepseek-v4.1-flash at $0.30/$1.20, matching the official peak cache-miss rate on DeepSeek's pricing page, which lists V4.1 Flash at $0.30/$1.20 peak and $0.15/$0.60 off-peak.
• DeepSeek V4 Pro on OpenRouter: $0.955/$1.911 → $1.60/$3.20 (+67.5%). DeepSeek's own page now lists V4 Pro at $1.32/$3.96 peak and confirms the model stays online past September 14 — the phase-out was withdrawn.
• Z.ai GLM-5.3 Flash doubled: $0.075/$0.25 → $0.15/$0.50 (+100%) on the same aggregator comparison.
• Moonshot AI's Kimi K3 fell ~11%: $3.00/$15.00 → $2.65/$13.28 on OpenRouter.
• Nous Research's Hermes 4 70B no longer appears in the OpenRouter catalog as of 2026-09-14; the larger Hermes 4 405B remains listed at $1.00/$3.00.

What happened

Issue one of the index, dated 2026-09-07, recorded a 17-model basket of OpenRouter list prices plus the official DeepSeek V4 Flash anchor. Today's fetch of the same surface shows five changed rows, one new row, and one removal.

The DeepSeek changes are the story. On 2026-09-10 DeepSeek shipped V4.1 Flash, and the official pricing page now bills the deepseek-flash model name as DeepSeek-V4.1-Flash: $0.30 input / $1.20 output per million tokens at peak, $0.15/$0.60 off-peak, $0.003–$0.006 for cache-hit input, with peak hours defined as 01:00–04:00 and 06:00–10:00 UTC on weekdays. The legacy deepseek-v4-flash name still resolves but is served by V4.1-Flash and billed at Flash rates. OpenRouter's new deepseek/deepseek-v4.1-flash row at $0.30/$1.20 mirrors the official peak rate exactly — the first time this index can anchor an aggregator row to a same-day first-party list price.

The same catalog's deepseek-v4-flash-0731 row, which matched the official $0.14/$0.28 anchor last week, now lists at $0.06/$0.12 — below any rate on DeepSeek's own page. That is a cheaper-backend routing outcome on the aggregator, not a first-party cut, and builders reading it as "DeepSeek costs six cents" would be wrong twice: the official cache-miss floor for the current Flash model is $0.15 off-peak, and the served checkpoint behind an aggregator slug is whatever the router chooses.

V4 Pro is the reversal. DeepSeek had signaled V4 Pro traffic would route to Flash from today; the pricing page now says the company "decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged" — listed at $1.32/$3.96 peak, $0.66/$1.98 off-peak. OpenRouter's V4 Pro row moved the other direction, up 67.5% to $1.60/$3.20. Neither number cancels the other: they are different products (first-party vs aggregated) observed on the same day.

Away from DeepSeek, Z.ai's GLM-5.3 Flash row doubled to $0.15/$0.50, and Moonshot AI's Kimi K3 row fell to $2.65/$13.28 — an ~11% cut that leaves it tied with Anthropic's Claude Sonnet 4.6 listing at the top of the basket's open-weight rows. Nous Research's Hermes 4 70B, present last week at $0.13/$0.40, is absent from today's catalog response; Hermes 4 405B remains at $1.00/$3.00. We have not verified whether the 70B removal is a delisting or a transient catalog gap.

Why it matters

List-price deltas decide routing budgets. A builder who sized a workload on last week's deepseek-v4-flash-0731 row is looking at a 57% cheaper aggregator quote today — but the durable, contract-grade number is DeepSeek's official V4.1 Flash rate, and the two should never be conflated. The V4 Pro reprieve removes this week's forced-migration risk for existing Pro users, at unchanged first-party billing. The GLM-5.3 Flash doubling is the largest single increase in the basket and lands on the row that was the second-cheapest open-weight option last week; at $0.15/$0.50 it now sits above Llama 4 Scout ($0.10/$0.30, unchanged) on input price.

The standing caveats from issue one apply in full: these are list rates, not invoices; aggregator rows reflect backend mix; and a cheaper row that fails your evals is not cheaper. For the eval side see GPT-5.5 vs open models and the cheapest LLM API roundup.

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. Today's delta also closes the loop on two stories from this week: the V4.1 Flash launch and silent-swap risk we covered on September 12, and the September 11 PULSE on the same launch — both flagged that official confirmation had to come from DeepSeek's pricing page, which is exactly what today's fetch provides. Builders running DeepSeek endpoints should also note the checkpoint-portability rule from issue one: deepseek-v4-flash, deepseek-v4-flash-0731, and deepseek-v4.1-flash are three different slugs at three different prices today.

What's next

The next observation is scheduled for Monday, 2026-09-21. Watch items for that run: whether Hermes 4 70B returns to the OpenRouter catalog, whether the GLM-5.3 Flash increase holds or reverts, and whether DeepSeek's off-peak/peak split starts appearing in aggregator rows rather than only on the first-party page. 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 2026-09-14 to 2026-09-07: DeepSeek V4 Flash 0731 fell 57% to $0.06/$0.12 per million tokens, GLM-5.3 Flash doubled to $0.15/$0.50, DeepSeek V4 Pro rose 67% to $1.60/$3.20, and Kimi K3 fell about 11% to $2.65/$13.28.

How much does DeepSeek V4.1 Flash cost on the official API?

DeepSeek's pricing page lists V4.1 Flash (model name deepseek-flash) at $0.30 input / $1.20 output per million tokens at peak cache-miss rates, $0.15/$0.60 off-peak, and $0.003–$0.006 for cache-hit input, as of 2026-09-14.

Is DeepSeek V4 Pro being discontinued?

No. DeepSeek's pricing page states it decided to continue providing V4 Pro API services after September 14, 2026, with billing unchanged — reversing an earlier phase-out plan covered in our September 12 story.

Why did OpenRouter's DeepSeek rows change if official billing is unchanged?

OpenRouter is an aggregator: each listed rate reflects the backend mix serving that slug on the snapshot date, not a first-party price list. A 57% drop on the 0731 row means cheaper backends now serve it — it is not an official DeepSeek price cut.