# LLM Prices Plummet: Breakthroughs and Savings Ahead! Newsy: LLM Cost Tracker Published: 2026-09-16T15:00:02.388Z Source: https://newsyai.com/edition/0235b589-800f-4c6a-9b4e-cb26510089d4 Archive: https://newsyai.com/newsy/7e294c66-a84b-45ba-b656-6f55351953db.md --- **# LLM Costs Keep Falling: Fresh Price Cuts, KV Breakthroughs, and Smarter Power Use** **Weekly briefing on advances that shrink the cost of large language models (September 9–16, 2026)** ## Pricing announcements from AI labs this week - No broad list-price drops on the biggest flagship models from OpenAI, Anthropic, or Google. Core rates stay steady.[[1]](https://www.aipricing.guru/pricing/) - **DeepSeek** cut prices on its V4.1 Flash model around September 11: input dropped from $0.22 to $0.15 per million tokens and output from $0.66 to $0.60 (about 15% lower overall). Off-peak and cached rates remain ultra-low, keeping pressure on the market.[[2]](https://www.aipricing.guru/) - **OpenAI** reduced prices on the GPT-5.6 series (reported September 15): frontier Sol tier blended price cut ~29%, mid-tier Terra ~20%, and budget Luna tier ~80%. These changes make high-volume and lighter workloads noticeably cheaper.[[3]](https://benchlm.ai/token-price-index) **Bottom line**: Flagship stability continues, but targeted cuts from DeepSeek and OpenAI on mid- and value-tier models drive effective costs lower, especially for cached or high-volume use. ## 1) New arXiv papers advance KV-cache efficiency for long contexts and multimodal models **What’s new** Recent papers (submitted September 8–10) include **Grouped Value Attention (GVA)**, which stores grouped values and reconstructs keys on demand to cut persistent cache size by 45–47% with near-identical accuracy. **OmniKVQuant** brings effective 2-bit quantization to multimodal (audio/video/text) models by handling modality-specific geometry and drift, preserving performance on benchmarks. **py-kvcache** shows external SSD-based KV offloading can be 2× faster than prior systems with smart preloading, and **KVShareArena** benchmarks reuse across different contexts and model checkpoints for retrieval and agent workloads.[[4]](https://arxiv.org/abs/2609.13285) **Why it matters for LLM costs** These techniques shrink memory needs and speed up loading or reuse without retraining or big quality losses. Less memory per session or faster cache hits means the same hardware handles more users or longer conversations, directly lowering energy and infrastructure costs per token. **Tie to the larger LLM economy** Layered on earlier KV optimizations, these multiply throughput across inference engines. Cheaper long-context, multimodal, and agentic applications encourage higher usage volumes, more training data, and faster iteration on the next wave of efficient models and hardware. ## 2) NVIDIA DSX platform boosts tokens per watt through power and grid co-design **What’s new** September 15 updates on NVIDIA’s DSX reference design highlight system-level optimizations: dynamic power management lets clusters deliver 23–24% more token throughput within the same power budget, with projections of up to 40% more GPU capacity in suitable environments. The platform integrates compute, networking, cooling, and grid signals for better overall efficiency.[[5]](https://blogs.nvidia.com/blog/from-megawatts-to-tokens-how-nvidia-maximizes-ai-factory-production/) **Why it matters for LLM costs** Power and cooling have long been major expenses. Squeezing more intelligence from every watt reduces the electricity bill per token and allows denser deployments without new grid connections, lowering both capital and operating costs for providers. **Tie to the larger LLM economy** These hardware and facility gains compound with software efficiencies (KV techniques, routing) and competitive pricing. Denser, cheaper-to-run infrastructure accelerates scaling, supports broader adoption, and reinforces the cycle of higher capability at lower prices. ## This week’s impact in plain English Targeted price reductions from DeepSeek and OpenAI, plus new research on smarter KV caching and NVIDIA’s latest power-optimization wins, show steady progress on “intelligence per dollar and per watt.” No single headline crash, but the combination of lower effective rates and technical gains keeps the long-term downward cost curve intact amid rising competition. ## Sources and further reading - AI Pricing Guru and LLMRates trackers (September 16, 2026 snapshots).[[1]](https://www.aipricing.guru/pricing/) - BenchLM Token Price Index update (September 15, 2026).[[3]](https://benchlm.ai/token-price-index) - arXiv papers: GVA (2609.13285), OmniKVQuant (2609.11582), py-kvcache (2609.11744), KVShareArena (2609.10266) (September 8–10, 2026).[[4]](https://arxiv.org/abs/2609.13285) - NVIDIA DSX platform update (September 15, 2026).[[5]](https://blogs.nvidia.com/blog/from-megawatts-to-tokens-how-nvidia-maximizes-ai-factory-production/) - TokenCost and related pricing blogs for model and rate adjustments (mid-September 2026).[[6]](https://tokencost.app/blog) Stable prices on flagships plus ongoing technical and operational gains keep the long-term cost trajectory firmly downward.