The Coding Agent Economy.
• $92 avg cost per active user / month
• Claude powers 92% of all coding agent spend (up from 68%)
• Cache hit rates jumped 52% → 86%
https://t.co/CG9rIsKXkq
The throughput density data suggests something counterintuitive:
the highest throughput providers are not necessarily serving the largest requests.
They are serving a massive number of relatively small generations extremely efficiently.
A lot of AI infrastructure performance right now looks less like “big intelligence” and more like high frequency inference systems.
Congrats @GroqInc
https://t.co/8ED0RDo3QT
The surprising thing in the latency data is how compressed the top providers have become.
For a lot of workloads, the gap between “fast” and “slow” providers is now smaller than the variance introduced by tool calls, long context, and agentic execution itself.
Model latency is starting to matter less than workflow latency. Congrats @xai https://t.co/XXys3M1CTJ
Most AI teams have zero control over which models employees and agents can actually use.
Today we’re launching Approved Models + Access Lists in Requesty.
You can now:
• approve models org-wide
• restrict models by API key or group
• enforce regional/compliance policies
• standardize model usage across teams
AI governance is becoming critical infrastructure.
https://t.co/1DWRNdZOut
The open source model market is consolidating much faster than expected.
A handful of OSS families now dominate traffic share while most new releases barely register.
The gap between “models people talk about” and “models people actually use in production” is getting very large.
@deepseek_ai is still dominating!
Jan → Apr 2026 data from Requesty ↓
https://t.co/outpgd6i2w
The interesting metric is not tool call request share.
It’s tool call token share.
Once workflows become agentic, token consumption shifts dramatically toward tool execution:
retrieval
code output
tool responses
intermediate reasoning
The number of requests can look normal while the token profile completely changes.https://t.co/b95ln2sqOL
One of the clearest signals of how people actually use AI might be finish reasons.
Anthropic direct traffic is now 52% tool calls.
OpenAI direct is just 3%.
You can literally see the difference between conversational usage and agentic workflows in the data.
April 2026 data from Requesty ↓
https://t.co/uFQMpkLRZr