State of AI 2026
We analyzed 18 trillion tokens of anonymized API traffic across 104 models and 18 providers to understand how developers actually use AI — and where the market is heading.
18T
Tokens analyzed
104
Models tracked
18
Providers
250k+
Developer accounts
Key findings
01
The open-source surge
Open-weight models grew from 12% to 32% of token share in 12 months. DeepSeek R1 and Qwen 3 235B alone account for 18% of all traffic — driven entirely by price performance, not marketing.
02
Programming dominates
34% of all tokens are consumed by programming tasks — a figure that rose 9 percentage points year over year as agentic coding tools like Cursor and Kilo Code scaled to hundreds of billions of tokens.
03
The agentic shift
Average prompt length grew 3.8× since early 2024. Tool-calling requests now account for 41% of traffic. Reasoning models handle 52% of all tokens — up from 8% eighteen months ago.
04
Asia Pacific catching up fast
APAC's token share grew from 13% to 24% in one year, driven by developer adoption of Qwen, DeepSeek, and locally deployed Llama variants. Chinese-developed models now power 31% of all APAC requests.
05
The Glass Slipper effect
Models that precisely match a user's workload generate 4× better 90-day retention. Frontier models launched before a task category existed rarely recover — model-task fit at launch determines long-term adoption.
06
Price drives switching
83% of model switches occur within 48 hours of a competitor price drop. DeepSeek V3's January launch triggered the largest single-day model migration in Nova's history — 2.1B tokens shifted in 6 hours.
Token share by provider
Share of total tokens processed · trailing 12 months
OPEN VS CLOSED SOURCE
BY PROVIDER
What developers actually use AI for
Task category distribution by token volume · text models only
Geographic distribution
Token share by region · April 2026 vs April 2025
Methodology
This report analyzes 18 trillion tokens of API traffic from January 2025 through April 2026. All data is aggregated and anonymized — no prompt content, user identifiers, or request-level metadata is retained or used. Task classification uses a lightweight routing classifier applied to anonymized request metadata only. Geographic attribution is based on account registration region, not IP geolocation. Provider market share reflects tokens processed through Nova's API, not global internet traffic.
Access the underlying data
Researchers and journalists can request anonymized dataset exports and custom cuts.
Visit the data page →