dottxt tool-call index: DeepSeek tops 100M daily calls, failure rates vary 0.04%–17% by provider
remilouf · x · 2026-11-21
dottxt's weekly "This week in tool calls" (Nov 9–15, 2026) tracks real-traffic function-calling failure rates:
- New models: Inkling 2 Small (5 providers, 1.4% breakage), Nex-N2.6 (0.62%), ERNIE-5.5-Flash-Preview (1.2%).
- Busiest models: DeepSeek V4.1 Flash 1017 at 101.7M calls/day (0.038% on Morph but 15% on AtlasCloud); GLM 5.3 Flash at 22.8M calls/day (0% best, 2.6% worst endpoint).
- Hour-by-hour regressions: gpt-oss-120b on CoreWeave hit 50% breakage, fixed to 6.6% Monday 16:00; DeepSeek V4.1 Flash 1017 on AtlasCloud worsened 1.7%→17% since Nov 3; DeepSeek V4 Pro 0423 on Azure rose 3.3%→11%.
- GLM 5.3 Flash on Novita climbed 0.9%→2.3%.
Takeaway: the same model's tool-call reliability varies wildly across inference providers — endpoint choice matters as much as model choice.
More from coding & agent
- Founder's rebuttal to McKinsey: enterprises need a first brain, not an AI second brain — vasuman · 2026-11-21
- Devs: LLMs absorb layered complexity until context explodes; tools and MCP are too heavy — dscape · 2026-11-21
- Inkling 2 Mini on Ollama: MLX build, one-command hooks into Claude Code and OpenCode — ollama · 2026-11-21
- Unguided coding agents pass only 52% of web checks; Modern Web Guidance lifts it to 91% — rseroter · 2026-11-21
- n8n launches Agents: your existing workflows become an AI agent's toolkit — SucceededMind · 2026-11-21
- Nous Research picks winners of Hermes Agent plugin hackathon with ~1,000 entries — alexcovo_eth · 2026-11-21