The DeepSeek and GLM-5.3 effect

Switching the default model to DeepSeek and adding GLM-5.3 produced the single biggest jump in usage I have ever recorded.

August 2026 — a data-driven look at my pi-go agent session history

TL;DR

Switching my coding agent’s default model to DeepSeek (via local Ollama) and adding GLM-5.3 to the rotation produced the single biggest jump in usage I’ve ever recorded. Monthly session volume went from a mid-year slump of 326 to an all-time high of 1,868 — a 5.7× increase — and the new model stack now accounts for ~54% of all sessions.

This isn’t a story about “better benchmarks.” It’s a story about adoption: when the right model becomes cheap, fast, and local, you simply use the agent more.

The Data

All numbers come from my pi-go session history (~/.pi-go/sessions), merged across two machines. Each row is one agent session.

Monthly session volume

Model mix by month (latest → oldest)

The new stack’s share of monthly usage

What Actually Happened

1. The mid-year slump (May–June)

Usage collapsed from 1,072 → 326 sessions. The old stack (GPT-5.5 + MiniMax) had plateaued. I was using the agent less, not more.

2. The DeepSeek switch (August)

I moved the default to DeepSeek-V4-Flash via local Ollama. The effect was immediate and dramatic:

• 783 DeepSeek sessions in August — the single most-used model ever in one month
• Aug 13 alone: 137 sessions — the biggest single-day spike in the dataset
• DeepSeek went from 0% → 42% of monthly usage in one month

3. GLM-5.3 joins the rotation (Aug 27)

GLM-5.3-Flash appeared Aug 27 (38 sessions) and Aug 29 (7 more), running through both Ollama and OpenRouter. It’s brand new — but it’s already the fastest-adopted model in the dataset, appearing across pi-go, ai-eng-course, and subagent task directories on day one.

4. Stealth (Aug 22+)

A third new model, Stealth/Ox-Alpha, ramped from Aug 22 (29) to Aug 25 (79). Together, DeepSeek + GLM-5.3 + Stealth = 1,040 sessions, 55.7% of August.

Why This Matters

The pattern is clear: usage growth tracked model quality-per-dollar, not marketing.

• GPT-5.5 era (Apr–Jul): capable but expensive → I rationed usage → volume fell.
• DeepSeek era (Aug): near-frontier quality, free and local via Ollama → no cost anxiety → I used the agent for everything → volume exploded.
• GLM-5.3: a fresh, fast option that slots into the rotation → more experimentation, more sessions.

The lesson for anyone building on LLMs: the friction of cost and latency is what suppresses usage. Remove it, and people don’t just do the same work cheaper — they do more work.

Projection (honest, not exponential)

I fit an exponential curve to the new-stack share and it produced nonsense (392%/month, >100% share) — because this is a step change, not organic growth. The realistic read:

• September: new stack likely 60–75% of sessions (DeepSeek stays dominant, GLM-5.3 and Stealth keep growing)
• October: 70–85%, approaching a ceiling
• Realistic ceiling: ~85–90% — the remaining other bucket is subagent/ACP sessions (Codex, Claude, Gemini) that will always exist

The growth is real, but it’s adoption saturation, not a hockey stick.

Data Notes

• Source: ~/.pi-go/sessions/*/meta.json, merged from local + dr-mac-studio via rsync (4,805 sessions total)
• “GLM-5.3” = glm-5.3-flash:cloud (Ollama) + z-ai/glm-5.3-flash (OpenRouter)
• “Other” includes subagent/ACP sessions (Codex, Claude, Gemini) that don’t record a model name
• Token figures excluded here; prompt-token counts are inflated by context re-sending