CHEAPEST INPUT TODAY Nova Micro $0.035/MBLENDED PRICE SPREAD 1102× (Nova Micro → GPT-5.4 Pro)Δ 2 FLAGGED sources disagree — see Data & Alerts131 MODELS · 16 PROVIDERS re-checked every morning by a GitHub ActionFOLLOW PRICE CHANGES commit feed → Data & Alerts

September 2026 Price Movement Report

Published 2026-10-01

This is the second entry in the monthly series that started with the August 2026 Price Movement Report: what actually moved in LLM pricing over the previous month, sourced entirely from docs/pricing-changelog.md, the append-only log our daily sync writes every time it touches a price, a model, or a provenance field. Same three things, same order: the largest confirmed move, the full tally by kind, and the live disputed count.

September was quiet, and we'll say so plainly. Seven price rows changed, on two models, and only one of those two was a vendor charging a different amount. Before any of that, though, we owe a correction to last month.

A correction to the August report

The August report described itself as covering every change the sync recorded in August. Its tally stopped at 27 August and missed the month's last four days. The log itself was complete; our count of it was not.

The four missing days included one set of price changes we should have reported. On 2026-08-30, three retired xAI models moved to a single new rate:

  • xai-grok-3 and xai-grok-4: $3 / $15 → $1.25 / $2.50 per million tokens (output −83%)
  • xai-grok-4-1-fast: $0.20 / $0.50 → $1.25 / $2.50 per million tokens (input +525%, output +400%)

This wasn't xAI repricing anything in August. xAI's own migration notice says retired Grok models are billed at grok-4.3's rate of $1.25 in / $2.50 out; our catalog caught up with that on the 30th, and on 2 September we confirmed it against xAI's retirement notice itself. But by the August report's own terms these belonged in it, and they overturn two claims it made:

- Grok 3's and Grok 4's $12.50 drop on output is a bigger dollar swing than GPT-5.6's $10 cut, which we called the largest of the month. - Grok 4.1 Fast's +525% on input is a bigger percentage than DeepSeek V4 Flash's +371%, which we called the largest single move.

The corrected August tally, all four missing days included:

  • Review: 255 (we reported 229)
  • Price: 77 across 20 distinct models (we reported 69 across 17)
  • Added: 76 (we reported 73) — 70 at the 2026-08-01 cold start, 6 genuine additions after it, not 3
  • Coverage: 31 (we reported 28)
  • Metadata: 26 (we reported 25)
  • Provider: 5 and Corrected: 1, unchanged

We've added a note to the top of the August report pointing here, and left its text as published. From this report on, the tally is taken on the first of the month, from the whole month, not while the draft is being written.

The largest move: DeepSeek cut V4 Flash

On 2026-09-16, DeepSeek cut the rate on its cheap model, confirmed against the vendor's own pricing page:

  • deepseek-deepseek-v4-flash input: $0.44 → $0.30 per million tokens (−32%)
  • deepseek-deepseek-v4-flash output: $1.32 → $1.20 per million tokens (−9%)
  • deepseek-deepseek-v4-flash cached input: $0.014 → $0.006 per million tokens (−57%)

DeepSeek now calls the model deepseek-flash (DeepSeek-V4.1-Flash). Its page says the old deepseek-v4-flash name is still accepted and billed at the new price, so the catalog row keeps its id.

Read it against last month and this is a small give-back, not a reversal. The 2026-08-16 repricing took V4 Flash's output from $0.28 to $1.32; at $1.20 it still costs about 4.3 times what it did at the start of August. V4 Pro didn't move.

A price change that changed nobody's bill: Claude Sonnet 5

On 2026-09-01, the log shows Sonnet 5's price falling from $3 / $15 to $2 / $10 per million tokens. Nobody was charged less. The August report told you the $2 / $10 introductory rate would expire on 31 August and step up to $3 / $15. Anthropic cancelled that step-up: its pricing page now says the introductory rate is the standard price and the increase won't happen. The catalog folded the intro rate into the base price and dropped the expired intro block. You paid $2 / $10 on 31 August and you pay $2 / $10 now.

That's the whole Price count for September: four rows for Sonnet 5, three for DeepSeek.

The rest of the month, by kind

Every changelog entry in September, tallied by the label the sync gave it:

  • Review: 299 — a disagreement between sources noted, updated, or cleared; no price changed
  • Metadata: 30 — eighteen status changes (nine of them the archiving described below), five context or output limits (Claude Sonnet 4.5's context window went from 200K to 1M tokens), five display names, two capability scores
  • Coverage: 16 — a field the catalog wasn't tracking before now is, or one a vendor stopped publishing
  • Price: 7 — across 2 models, described above
  • Added: 6 — Claude Fable 5.1, GLM-5.2, Gemini 3.8 Flash, Grok 4.7, Claude Opus 5.5, Claude Sonnet 5.5

No Provider or Corrected entries this month.

Month over month

The comparison the series exists for, August (corrected) against September:

  • Models added: 6 after launch, plus 70 at launch → 6
  • Models moved to legacy: 1 → 7
  • Models archived as retired: 0 → 9
  • Models re-priced: 20 → 2, only 1 of them by the vendor
  • Changelog entries: 471 → 358

"Legacy" means a model is no longer offered as current upstream; the catalog keeps its page rather than deleting it. September's seven were Grok 3, Grok 4 and Grok 4.1 Fast (the retired models above), Mistral Medium 3, Kimi K2.5, Gemini 3 Pro Preview and Kimi K2 Thinking. An eighth, Gemini 3.1 Flash-Lite Preview, was marked legacy on 2026-09-23 and restored the next day.

"Archived" is new this month. On 2026-09-29 we began marking models that a vendor has shut down entirely as deprecated: the page and its price history stay, but the row is frozen and no longer offered in any model picker. Nine were archived that day: Claude Opus 4.1, DeepSeek R1, DeepSeek V3, DeepSeek V3.2, Gemini 3 Pro Preview, Mistral Medium 3, Kimi K2 Thinking, Kimi K2.5 and Grok 2.

August's re-pricing number is inflated by the pipeline's first days: 43 of its 77 Price rows landed on 2026-08-02. September is the first month that reads as the market rather than the machinery, and the market was calm.

What's still disputed

When the August report went out, 17 models carried an open needsReview flag. As of this report it's 1, from our own health endpoint:

curl -s https://promptspend.dev/v1/health

That drop needs reading carefully, because most of it isn't disputes being settled. The count had reached 23 by 2 September, when 16 flags were cleared in one pass by reading each vendor's own pricing page. Eight were still open on 28 September. On the 29th, the Grok 4.7 flag cleared, and the other seven all belonged to models that were then archived as retired (Claude Opus 4.1, DeepSeek R1, Gemini 3 Pro Preview, Mistral Medium 3, Kimi K2 Thinking, Kimi K2.5 and Grok 2). A model its vendor no longer sells can't be re-checked against anything, so archiving closes its flag rather than resolving it.

The one flag open today is new: on 1 October, Kimi K3's recorded price of $3 / $15 per million tokens disagreed with OpenRouter's $0.723 / $10. A flag means two sources disagree right now, not that either is necessarily wrong.

Reading this report

Next month's report goes out on 1 November, with the same three things in the same order. September is what a quiet month looks like, and we'd rather publish it than skip it.

Standard-tier, global-endpoint list prices in USD. This build read the catalog generated 2026-11-17.