Finding 01
The low cost cluster is already deep.
165 priced models sit at or below $0.50 blended per one million tokens using a 3 input to 1 output workload mix. That makes the bottom of the market wide enough for real substitution pressure.
Research note · Live pricing dataset
API Pricing Compression Monitor
A live BenchGecko note tracking where AI API prices are compressing, which providers anchor the low cost floor, and which premium models still sit far above the market median.
Dataset date
Aug 12, 2026
BenchGecko generated data
438
Priced models
61 providers with input and output prices
$0.4
Median input
per one million input tokens
$1.44
Median output
per one million output tokens
165
Low cost band
models at or below $0.50 blended
Finding 02
Premium pricing still has a long tail.
45 priced models sit above $5 blended. The monitor treats these as premium outliers until benchmark strength, latency, context, or specialist capability explains the gap.
Finding 03
Output tokens remain the expensive side.
The median output price is $1.44 versus $0.4 for input. The 90th to 10th percentile spread is 83.3x for output and 42.9x for input.
Low cost floor
Cheapest blended token prices in the current dataset.
LFM2-2.6B
liquid · input $0.01 · output $0.02
$0.013
blended
LFM2-8B-A1B
liquid · input $0.01 · output $0.02
$0.013
blended
Ling-2.6-flash
inclusionai · input $0.01 · output $0.03
$0.015
blended
Llama 3.1 8B Instruct
Meta · input $0.02 · output $0.03
$0.023
blended
Mistral Nemo
Mistral AI · input $0.02 · output $0.03
$0.023
blended
Granite 4.0 Micro
ibm-granite · input $0.017 · output $0.112
$0.041
blended
Llama 3 8B Lunaris
sao10k · input $0.04 · output $0.05
$0.042
blended
Nex-N2-Mini
nex-agi · input $0.025 · output $0.1
$0.044
blended
Gemma 2 9B
Google DeepMind · input $0.03 · output $0.09
$0.045
blended
Qwen2.5 Coder 7B Instruct
Alibaba Qwen · input $0.03 · output $0.09
$0.045
blended
Provider compression table
Providers with at least three priced models, ranked by median blended price.
Premium watchlist
High priced models that need capability context before price alone is interpreted as market power.
o1-pro
OpenAI · input $150 · output $600
$262.5
GPT-5.4 Pro
OpenAI · input $30 · output $180
$67.5
GPT-5.5 Pro
OpenAI · input $30 · output $180
$67.5
Claude Opus 4.6 (Fast)
Anthropic · input $30 · output $150
$60
Claude Opus 4.7 (Fast)
Anthropic · input $30 · output $150
$60
GPT-5.2 Pro
OpenAI · input $21 · output $168
$57.75
GPT-5 Pro
OpenAI · input $15 · output $120
$41.25
GPT-4
OpenAI · input $30 · output $60
$37.5
Methodology and caveats
Prices are taken from BenchGecko model records and expressed per one million tokens where available.
The blended price uses a 3 input to 1 output workload mix. It is a comparison lens, not a universal workload model.
Compression does not mean quality convergence. Benchmark coverage, latency, context, tools, and reliability still matter.