Weekly Build Brief — 21 September 2026
Dated changes in the LLM API catalogue and GPU rental floor between 2026-09-20 and 2026-09-21: 5 price move(s), 5 GPU rate move(s), 6 added / 6 removed model(s). Read from source APIs, with the monthly bill for a stated workload.
Window: 2026-09-20 → 2026-09-21 (two consecutive dated snapshots of the catalogue, read from the source APIs — never written by a model).
What moved in the API catalogue
| Model | Per 1M tokens | Before | After | Change |
|---|---|---|---|---|
deepseek/deepseek-v4-flash |
input | $0.0372 | $0.0899 | +141.7% up |
deepseek/deepseek-v4-flash |
output | $0.0745 | $0.1797 | +141.2% up |
deepseek/deepseek-v4-pro-0813 |
output | $1.7345 | $3.9600 | +128.3% up |
deepseek/deepseek-v4-pro-0813 |
input | $0.5782 | $1.3200 | +128.3% up |
deepseek/deepseek-v4-pro |
input | $0.4223 | $0.9483 | +124.6% up |
Full dated log: /llm-costs/changes/.
What it costs on a fixed workload
Reference workload: 20M input tokens + 5M output tokens per month, priced with the observed rates above — recompute it yourself, the rates are the only input.
| Model | Before | After | Monthly difference |
|---|---|---|---|
deepseek/deepseek-v4-flash |
$1.64 | $2.70 | +1.05 $/month |
deepseek/deepseek-v4-flash |
$2.17 | $2.70 | +0.53 $/month |
deepseek/deepseek-v4-pro-0813 |
$35.07 | $46.20 | +11.13 $/month |
deepseek/deepseek-v4-pro-0813 |
$31.36 | $46.20 | +14.84 $/month |
deepseek/deepseek-v4-pro |
$17.93 | $28.45 | +10.52 $/month |
Live rates: /api-pricing/.
GPU rental floor
| GPU | Before | After | Change |
|---|---|---|---|
RTX 5090 |
$0.0135 | $0.2129 | +1477.0% |
RTX PRO 6000 S |
$1.0689 | $1.4689 | +37.4% |
RTX 3090 |
$0.1213 | $0.0874 | -27.9% |
RTX PRO 6000 Max-Q |
$1.0681 | $0.8022 | -24.9% |
H100 NVL |
$2.1356 | $1.9763 | -7.5% |
Lowest observed hourly price per card: /gpu-pricing/.
Catalogue entries and exits
| Movement | Model |
|---|---|
| added | SamLowe/roberta-base-go_emotions |
| added | joeddav/xlm-roberta-large-xnli |
| added | lpiccinelli/unidepth-v2-vitl14 |
| added | peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF-MTP |
| added | stabilityai/sdxl-turbo |
| added | unsloth/gemma-4-12b-it-GGUF |
| removed | CompVis/stable-diffusion-safety-checker |
| removed | Qwen/Qwen2.5-1.5B |
| removed | bartowski/XYZAILab_XYZ-Aquila-mini-GGUF |
| removed | jinaai/jina-embeddings-v2-small-en |
| removed | meta-llama/Llama-3.2-1B |
| removed | mudler/ced-gguf |
6 added, 6 removed between the two snapshots.
Published this week
| Date | Pipeline | Identifier |
|---|---|---|
| 2026-09-20 | east_west_room | pdf-to-markdown-saas-vs-mineru |
| 2026-09-14 | east_west_room | chinese-tts-stack-licensed-cosyvoice-vs-chattts |
| 2026-09-14 | east_west_room | curl-cffi-beats-cloudflare-without-browser |
| 2026-09-14 | east_west_room | ms-swift-chinese-finetuning-standard |
| 2026-09-21 | openrouter_catalogue | deepseek/deepseek-v4-flash-0731 |
| 2026-09-21 | openrouter_catalogue | ~deepseek/deepseek-v4-flash-latest |
| 2026-09-21 | openrouter_catalogue | deepseek/deepseek-v4-flash |
| 2026-09-21 | openrouter_catalogue | deepseek/deepseek-v4-flash |
| 2026-09-21 | openrouter_catalogue | ibm-granite/granite-4.2-8b |
| 2026-09-21 | openrouter_catalogue | ibm-granite/granite-4.2-8b |
Every entry is a page on this site; the registry that produces it is data/content_registry.json.
How this page is built
- Prices, GPU rates and catalogue entries are read from the source APIs and stored as dated snapshots (
data/history/pricing/,data/history/gpu/,data/history/models/). - This page compares 2026-09-20 and 2026-09-21 and shows only what crossed the materiality thresholds (5% and $0.02 per 1M tokens for prices, 5% for GPU rates).
- No language model writes this page. The only arithmetic is the monthly bill above, from the rates shown.
- Nothing to report means no page: a week without a material change is not published.
- Observed 2026-09-21 05:55 (local time).
Was this article helpful?
Let us know to improve our AI generation.