Current public projection
AI API Pricing Index — September 2026
A reproducible snapshot of 80 public pricing records across 11 providers. Cutoff 2026-09-29; every statistic below is calculated from the same public projection that powers the AICostBudget Dataset.
First-party and cloud APIs
Conditional structured charges
At least one per-1M-token component
Comparable cached-input scalar
Verified, not inferred
The useful signal is in the pricing structure, not one cheapest-model ranking.
All medians use current public records with a compatible numeric value. Missing values are excluded, never treated as zero.
Median input price
Across 68 records with a current input-token scalar; observed range $0.10–$30 per 1M tokens.
Median output price
Across 68 records with a current output-token scalar; observed range $0.10–$180 per 1M tokens.
Median output/input ratio
Across 68 records with both prices. Output generation, not prompt ingestion, is commonly the higher list-price component.
Median cache discount
Across 57 comparable records. The observed discount range is 75%–98%.
Median Batch discount
Across 43 records with a calculation-default Batch input component; observed range 20%–50%.
Records with non-token charges
Document pages, storage, session duration, or time-based units cannot be safely flattened into a token-only price table.
Which providers contribute the most public records?
Source: AICostBudget AI API Pricing Dataset. Methodology.
How do provider median input and output prices compare?
Source: AICostBudget AI API Pricing Dataset. Methodology.
How widely are cost-saving price modes published?
A workload comparison that ignores cache and Batch pricing omits published savings modes for a large share of the tracked records.
Source: AICostBudget AI API Pricing Dataset. Methodology.
Where does token-only comparison stop working?
Source: AICostBudget AI API Pricing Dataset. Methodology.
What changed in September 2026?
14 events: 8 model additions, 3 successor comparisons, 2 lifecycle updates, and 1 time-pricing schedule update.
Claude Sonnet 5.5
Model added
Claude Opus 5.5
Model added
Claude Opus 5.5
New generation pricing
Gemini 3.8 Flash TTS
Model added
Gemini 3.8 Flash-Lite TTS
Model added
GPT-6 Luna
Model added
GPT-6 Luna
New generation pricing
GPT-6 Sol
Model added
GPT-6 Sol
New generation pricing
Grok 4.7
Model added
DeepSeek V4 Flash
Time-based pricing schedule updated
DeepSeek V4 Flash
Model lifecycle updated
DeepSeek V4 Flash Vision Exp
Model lifecycle updated
DeepSeek V4.1 Flash
Model added
Source: AICostBudget AI API Pricing Dataset. Methodology.
Key statistics
The table states the population and method so each result can be independently reproduced.
Coverage and current medians by provider
Provider medians describe the records tracked here; they are not market share or usage-weighted prices.
What this index measures—and what it does not.
Source and scope
The source is the AICostBudget public Website projection, generated from canonical provider/model records and linked official provider sources. The population is 80 public pricing records at the 2026-09-29 cutoff.
Normalization
Comparable text-token scalars are normalized to USD per 1M tokens. Provider and model aliases do not create a second research observation. Null means unavailable or not applicable, never zero.
Cache, Batch, and tiers
Cached input and Batch discounts are calculated only when both the standard scalar and a compatible published component exist. Long-context and usage tiers remain separate structured records; thresholds price the whole matching request.
Non-token pricing
Document pages, per-minute sessions, and token-storage hours keep their native billing units. They count as non-token pricing architecture and are excluded from input/output token medians.
Time and history
September events use provider-announced or official-changelog effective dates when available. The local snapshot window begins in July 2026, so this page does not claim an industry-wide annual price decline.
Verification limits
Public list prices are planning inputs, not invoices. Regional pricing, negotiated contracts, taxes, routing channels, tokenization, and workload quality can change realized cost.
Full field definitions, official source groups, exclusions, licensing, and downloads are maintained on the AI API Pricing Dataset methodology page.
Copy a citation or use the underlying data.
Link to this dated analysis for the findings; cite the Dataset when reusing the machine-readable records.
AICostBudget. "AI API Pricing Index — September 2026." AICostBudget, 2026. Based on the AICostBudget AI API Pricing Dataset. https://aicostbudget.com/en/research/ai-api-pricing-index-september-2026@article{aicostbudget_pricing_index_2026_09,
author = {{AICostBudget}},
title = {AI API Pricing Index --- September 2026},
year = {2026},
month = {September},
url = {https://aicostbudget.com/en/research/ai-api-pricing-index-september-2026},
note = {Based on the AICostBudget AI API Pricing Dataset; cutoff 2026-09-29}
}