Enterprise AI Inference Costs Hit 2026 Low: The Price War and Open-Source Surge Explained

Silicon Data's LLM Token Expenditure Index — This index tracked the index-average falling from above $2 per million tokens in early June to $1.20 in the first week of August — a roughly 40% decline in two months . A snapshot on July 27 showed $1.43, confirming the downward trajectory . Silicon Data directly cited Chinese open-weight models as a primary factor dragging down the index average .

Key Competitive Actions Driving the Drop

OpenAI's GPT-5.6 Price Cuts (July 30, 2026)

OpenAI slashed API prices by up to 80%, citing infrastructure efficiency gains. The moves directly compressed the market average; CNBC and Forbes characterized them as escalating the AI price war .

• GPT-5.6 Luna (lightweight tier): dropped from $1.00 → $0.20 per million input tokens and $6.00 → $1.20 per million output tokens .
• GPT-5.6 Terra: cut 20% to $2 input / $12 output per million tokens .
• GPT-5.6 Sol (flagship): held at $5 input / $30 output per million tokens, but gained a faster mode .

OpenAI attributed the reductions to internal optimizations, including kernel rewrites and speculative decoding, which reportedly delivered 20-35% efficiency gains in inference workloads .

DeepSeek V4-Flash-0731 (July 31, 2026)

On July 31, 2026, DeepSeek released V4-Flash-0731 in public beta, holding pricing at:

• $0.14 per million uncached input tokens
• $0.28 per million output tokens
• $0.0028 per million cached input tokens (a 97-98% cache discount) .

At roughly one-tenth the cost of premium tier models, DeepSeek V4 Flash became the floor-setting price that forced broader market compression. According to benchmark firm Artificial Analysis, V4 Flash's estimated cost per task is approximately 3 cents . The model beats DeepSeek's own V4 Pro preview on every agentic benchmark DeepSeek published, at a third of the output price .

Anthropic's Claude Opus 5 (Launched July 24, 2026)

Opus 5 entered at $5 per million input tokens and $25 per million output tokens — identical to its predecessor Opus 4.8 — with a Fast mode at 2x base price ($10/$50) . While Opus 5 didn't cut prices itself, its launch at a competitive "near-frontier" tier (near Fable 5 intelligence at half the price) put pressure on rivals and expanded the low-cost option set for enterprises .

The Combined Effect

OpenAI's 80% cut (July 30), DeepSeek's sub-$0.30 Flash pricing (July 31), Anthropic's competitive Opus 5 launch (July 24), and widespread enterprise adoption of low-cost Chinese open-source models compressed the market index from $2.04 in late May to $1.16–$1.18 by early August — a 43% drop in ten weeks .

Jefferies analysts led by Thomas Chong attributed the decline to an "intensifying global price war" and a "surge in adoption of affordable Chinese open-weight models" . The market reaction was not limited to pricing: Jefferies also argued that falling token costs could boost AI demand and infrastructure investment .

A Note on the Data

Silicon Data's public index showed $1.43 on July 27 versus the reported $1.20 in early August, suggesting a steep late-July/early-August acceleration driven by the OpenAI and DeepSeek announcements . The index blends prices and usage, meaning a dip can reflect either falling list prices or a shift in demand toward cheaper models — Silicon Data itself has warned against reading it as a pure price tag .