Monthly release · October 2026
What the routing layer measured in October 2026.
Monthly open data from Requesty's production gateway: which coding agents and tools teams use, how often providers fail and why, latency, caching and reasoning effort. Every number has a trend, a method and a download.
Refreshed monthly · CC BY 4.0 · RSS · index.json · Weekly charts
Finding of the month · Coding agent adoption
OpenCode was used by 61.2% of organizations running a coding agent through Requesty in October 2026, ahead of Claude Code at 21.8%.
- OpenCode61.2%
- Claude Code21.8%
- Kilo Code6.2%
- GitHub Copilot5.2%
- Pi3.9%
- Qwen Code3.1%
See the trend and data →
What changed in October 2026
Time to first token by provider
Vertex (Gemini): median time to first token 2.06 s → 3.62 s (+1.56 s)
Prompt cache hit rate by provider
Vertex (Gemini): cache hit rate 11.7% → 51.1% (+39.3 pp)
Provider error rates
DeepInfra: provider error rate 22.9% → 7.7% (-15.2 pp)
Reasoning effort mix
none: share of requests that set effort 17.5% → 26.9% (+9.4 pp)
Find a dataset
- Coding agents and toolsMonthly · October 2026
Coding agent adoption
Which coding agents do teams run through an LLM gateway?
OpenCode was used by 61.2% of organizations running a coding agent through Requesty in October 2026, ahead of Claude Code at 21.8%.
- Coding agents and toolsMonthly · October 2026
Tools models call
Which tools do LLMs call most often in production?
3.3% of chat requests in October 2026 returned at least one tool call. The most widely called tool name was "read", seen at 37.0% of organizations using tools.
- ReliabilityMonthly · October 2026
Provider error rates
How often do LLM providers return errors, and what kind?
At the median major provider in October 2026, 0.07% of requests failed with a provider 5xx and 0.39% were rate limited with a 429.
- LatencyMonthly · October 2026
Time to first token by provider
Which LLM provider returns the first token fastest?
Melious had the fastest median time to first token among major providers in October 2026 at 603 ms. The median major provider took 1.71 s.
- CachingMonthly · October 2026
Prompt cache hit rate by provider
What share of input tokens do LLM providers serve from prompt cache?
Fireworks served 94.0% of input tokens from cache in October 2026, the highest among major providers. The median major provider served 80.4%.
- ReasoningMonthly · October 2026
Reasoning effort mix
Which reasoning effort level do developers ask for?
7.0% of chat requests set a reasoning effort in October 2026. Of those, "low" was the most requested level at 38.7%.
- ReliabilityUpdated · Oct 2026
Operational metrics per provider, October 2026
Which AI provider has the highest success rate?
In October 2026 Parasail had the highest success rate at 100% and DeepInfra the lowest at 91% among the 15 busiest providers.
- LatencyUpdated · Oct 2026
Latency leaderboard, October 2026 (top 10 providers by volume)
Which AI provider has the lowest latency?
Among the ten busiest providers in October 2026, Typesafe had the lowest median total latency at 225ms and DeepInfra the highest at 8.0s.
- LatencyUpdated · Oct 2026
Provider throughput density, October 2026
Which AI provider has the highest tokens per second?
In October 2026 Typesafe led at a median 889 output tokens per second of total latency. Parasail was slowest at 28 tokens per second.
- Coding agents and toolsUpdated · Oct 2026
Reasoning-token share of provider output, October 2026
Which providers serve the most reasoning tokens?
In October 2026 reasoning tokens made up 78% of output on Fireworks, the highest share, and 6% on Mistral, the lowest among providers with meaningful reasoning volume.
- Coding agents and toolsUpdated · Oct 2026
Tool-call share of output tokens and requests, October 2026
Which providers handle the most tool calls?
In October 2026 Melious had the highest tool-call share at 66.4% of output tokens and 66.2% of requests.
- Coding agents and toolsUpdated · Oct 2026
finish_reason mix per provider, October 2026
Which provider serves the most agentic traffic?
In October 2026 Vertex (Claude) had the highest tool_calls share at 34%, followed by Sference at 31% and DeepSeek at 31%.
- Coding agents and toolsUpdated · Oct 2026
finish_reason mix per model, October 2026
Which LLM is used most for tool calling?
In October 2026 deepseek-v4.1-flash-1017 ended 73% of successful completions with tool_calls, the highest among the busiest models. gpt-4.1-mini was at 0%, used almost entirely for plain completions.
- ReliabilityUpdated · Oct 2026
Provider error code distribution, October 2026
What is the most common LLM API error?
In October 2026, 429 rate limits made up 66.4% of provider-returned errors on the Requesty gateway. 400 Bad Request was next at 18.5%, and 503 Service Unavailable at 6.3%.
- LatencyUpdated · Oct 2025 to Oct 2026
p50 latency year over year: October 2025 to October 2026
Have LLM providers gotten faster in the last year?
Comparing October 2025 with October 2026, Nebius median latency fell 90%, from 2.8s to 288ms.
- Coding agents and toolsUpdated · Oct 2026
Cache hit rate by coding agent, October 2026
Which coding agent has the best cache hit rate?
In October 2026 Cline led at 96% of input tokens from cache. Qwen Code was lowest at 82%. Most agents sat above 90%.
- ReliabilityUpdated · Oct 2026
Error rate by coding agent, October 2026
Which coding agent has the lowest error rate?
In October 2026 Cursor had the lowest error rate at 1.0% of calls, and Qwen Code the highest at 10.9%.
- Coding agents and toolsUpdated · Oct 2026
Tool-call finish rate by coding agent, October 2026
Which coding agent makes the most tool calls?
In October 2026 Roo Code ended 95% of calls with a tool call, the highest share. Cline was lowest at 45%.
- Coding agents and toolsUpdated · Oct 2026
Claude model share by coding agent, October 2026
Which models do coding agents use most?
In October 2026 Claude models took 92% of Qwen Code spend and 89% of Claude Code spend. Hermes Agent spent the least on Claude at 1%.
- Coding agents and toolsUpdated · Oct 2026
Streaming adoption by coding agent, October 2026
Do coding agents use streaming?
In October 2026 8 of 12 agents streamed nearly every call. Claude Code streamed the least at 87%.
- Coding agents and toolsUpdated · Oct 2026
Session depth by coding agent, October 2026
How many API calls does a typical coding agent session make?
In October 2026 Roo Code had the deepest sessions at a median 41 calls per trace, with a p95 of 169. Claude Code sessions had a median of 5 calls.
- LatencyUpdated · Oct 2026
Claude Code median latency by provider and model, October 2026
Is Claude faster on Anthropic, Bedrock or Vertex?
In October 2026 Vertex (Sonnet) was fastest at 2.6s median provider latency. Bedrock (Opus) was slowest at 6.1s.
- Coding agents and toolsUpdated · May 2025 to Oct 2026
Average cost per user per month by coding agent, May 2025 to October 2026
Which coding agent has the highest average cost per user?
In October 2026 Claude Code users averaged $131 per month, up from $78 in April. OpenCode users averaged $44 and Kilo Code users $25.
- Coding agents and toolsUpdated · Jan 2026 to Oct 2026
Family share within open-weight traffic, January to October 2026
Which open-source LLM is most used?
In October 2026 DeepSeek held 60% of open-weight traffic on the Requesty gateway.
- ReliabilityUpdated · Jan 2026 to Oct 2026
Policy vs direct eventual success rate, January to October 2026
Do LLM fallbacks improve reliability?
In October 2026, 97.8% of requests sent through a Requesty routing policy eventually succeeded, against 95.5% of direct requests.
Earlier editions (21 notes kept for citations)
- finish_reason mix per provider, April 2026 · newer: Oct 2026
- finish_reason mix per model, April 2026 · newer: Oct 2026
- Token-weighted tool_calls share per provider, April 2026 · newer: Oct 2026
- Family share within OSS-routed traffic, Nov 2025 - Apr 2026 · newer: Jan 2026 to Oct 2026
- Latency leaderboard per provider, April 2026 · newer: Oct 2026
- Provider throughput density, April 2026 · newer: Oct 2026
- Streaming TTFT vs total latency, April 2026 · newer: Monthly
- p50 latency YoY: April 2025 vs April 2026 · newer: Oct 2025 to Oct 2026
- Operational metrics per provider, April 2026 · newer: Oct 2026
- Prompt-cache hit rate per provider, April 2026 · newer: Monthly
- Reasoning-token share of provider output, April 2026 · newer: Oct 2026
- Provider error code distribution, April 2026 · newer: Oct 2026
- Policy vs direct eventual success rate, Jan-Apr 2026 · newer: Jan 2026 to Oct 2026
- Average cost per user per month by coding agent, May 2025 to April 2026 · newer: May 2025 to Oct 2026
- Prompt-cache hit rate by coding agent, April 2026 · newer: Oct 2026
- Claude model share by coding agent, April 2026 · newer: Oct 2026
- Error rate by coding agent, April 2026 · newer: Oct 2026
- Tool-call finish rate by coding agent, April 2026 · newer: Oct 2026
- Claude Code median latency by provider and model, April 2026 · newer: Oct 2026
- Session depth by coding agent, April 2026 · newer: Oct 2026
- Streaming adoption by coding agent, April 2026 · newer: Oct 2026
Fetch /data/index.json for every series, month and download URL, or /data/llms.txt for a plain-text catalog. Every page has JSON, CSV and Markdown exports.
index.jsonFree to reuse under CC BY 4.0 with credit to Requesty Data. Each month has a permanent URL with APA and BibTeX, so quoted numbers never change under you.
See an exampleNew months land early in the following month. Follow the RSS feed, or read the weekly Gateway Charts for what changed this week.
RSS feed