Large Language Model Cache Economics: Why Frontier API Prices Fell 80 Percent
A series of tit-for-tat price reductions between OpenAI and Anthropic, alongside aggressive open-weight competition, dramatically lowered model costs. Because agentic coding workflows spend 96% to 99% of total tokens on cached context, cuts to cache read pricing directly crater real-world inference bills.
Why it matters
In agentic software engineering pipelines, conversations append past execution history turn-by-turn. When cache hit rates reach 90% to 100%, cache read fees constitute almost the entire bill, making cache-ratio reductions far more impactful than baseline input or output rate cuts.
TL;DR
- 01In iterative agent loops, cached context accounts for 96% to 99% of processed tokens, meaning cache read pricing dominates overall bills.
- 02Anthropic cut cache read costs by 75% with Fable 5.1 ($0.25/MTok vs $1/MTok for Fable 5), yielding a 70% to 75% net cost drop for agent workflows.
- 03OpenAI dropped GPT-5.6 Luna pricing by 80% to compete with low-cost open-weight models like DeepSeek V4 Flash.
- 04Open-weight competition on gateways like OpenRouter continues to push down cache fees, with DeepSeek Flash V4.1 available across 29 providers at $0.003/MTok cache reads.
Key facts
- GPT-5.6 Luna price cut
- 80% reduction to $0.20 / $1.20 / $0.02 per MTok
- Fable 5.1 cache read discount
- 75% reduction to $0.25 per MTok compared to Fable 5
- Agent cache token share
- 96% to 99% of total tokens per turn
- Vercel token cost index
- Fell to 43.3% of April 2026 cost
The Shift to Cache-Centric Pricing
Headline LLM pricing previously prioritized standard input and output tokens. However, the aggressive release cycle between Anthropic and OpenAI shifted competitive pressure to include aggressive cache read pricing:
- GPT-5.6 Luna: Dropped 80% to
$0.20input,$1.20output, and$0.02cache read per million tokens. - Claude Fable 5.1: Retained
$10input and$50output, but cut cache reads by 75% to$0.25/ MTok compared to Fable 5 ($1 / MTok). - Claude Opus 5.5: Reduced compute serving requirements to offer
$4input,$20output, and$0.20cache read. - GPT-6.1 Sol: Halved its cache reads to
$0.10/ MTok (pricing at$2/$10/$0.10), matching Anthropic's 5%-of-input ratio.
Why Cache Read Ratios Dictate Coding Agent Bills
During iterative coding sessions in Claude Code or Codex, the context window retains complete file trees, documentation, and earlier execution turns. Cache hit rates typically hover between 90% and 100%. Because cached tokens represent between 96% and 99% of total processed volume in long sessions, cutting cache read rates by 75% directly reduces overall execution costs by 70% to 75%.
Open-Weight Commodity Pressure
Gateway routing data shows that proprietary frontier labs are responding directly to open-weight model economics. DeepSeek Flash V4.1 maintains 29 independent hosting providers on OpenRouter with cache read rates around $0.003 / MTok. Simultaneously, Moonshot AI's Kimi K3.1 Pro scores 55 on the Artificial Analysis index (approaching Opus 5.5 at 58) while charging $2 / $10 / $0.20 per million tokens.
✓ When to use
- Agentic software engineering systems like Claude Code, Codex, or OpenCode with long session memory.
- Multi-turn interactive conversational interfaces processing large shared documentation context.
✕ When NOT to use
- Single-turn API calls without conversational history where context caching cannot be leveraged.
- Workloads where queries vary completely across calls with near 0% cache hit rates.
What to do today
- Audit existing LLM API pipelines to verify whether prompt caching headers and mechanisms are properly enabled.
- Benchmark models using effective cost-per-turn calculations that incorporate 90%+ cache hit rates rather than raw input token rates.
- Evaluate multi-provider routing on OpenRouter to capture the lowest available cache read tiers for secondary agent tasks.