Why GitHub Copilot’s $180/mo Cost Crisis Drove Our Team to Self-Hosted DeepSeek — And How We Cut 70% Off the Bill

TL;DR: GitHub Copilot Cowork bills hit ~$180/mo per heavy user because all traffic routes through premium Anthropic/OpenAI models with no routing controls. We moved planning to self-hosted DeepSeek-V4 via Ollama + Continue.dev, kept execution on Claude Sonnet via OpenRouter, and dropped effective cost to ~$45/mo. Full comparison →

The Context

Three-dev team, 120+ hours/month AI-assisted coding across TypeScript/Go/PostgreSQL. Started on GitHub Copilot Business ($19/mo) in 2024. May 2026: Copilot Cowork launched with “unlimited” premium models. Reality: heavy users (8+ hrs/day) saw bills of $140–180/mo when Microsoft passed through Anthropic/OpenAI costs 1:1. No per-seat caps, no routing to cheaper models, no visibility into token breakdown. Microsoft’s own leak (early June 2026) confirmed they’re exploring self-hosted DeepSeek-V4 and MAI family models for a cheaper tier — admitting the current model is unsustainable.

What We Tested

Configuration	Monthly Cost (Heavy User)	Planning Quality	Execution Quality	Control
Copilot Cowork (default)	$180	Excellent (Opus/GPT-4o)	Excellent	None — all premium
Copilot + VS Code (no Cowork)	$19	Good (GPT-4o mini)	Degraded on complex refactors	None
Continue.dev + Ollama (DeepSeek-V4) + OpenRouter (Claude Sonnet)	~$45	Excellent (DeepSeek-V4 32B)	Excellent (Claude 3.5 Sonnet)	Full — per-task routing, budgets, local planning
VS Code + Cline + OpenRouter (mixed)	~$55	Excellent (Haiku/4o-mini)	Excellent (Sonnet)	Full — BYOK, per-key budgets

The Pivot Point

June 8, 2026: One dev’s Copilot bill hit $176 for the month. Breakdown showed 68% of tokens were “planning” — thinking through approach, writing pseudocode, exploring files. Those tokens billed at Opus/GPT-4o rates ($15–30/M). Switched that dev to Continue.dev with Ollama-hosted DeepSeek-V4 for planning (qwen2.5-coder:32b works too), routed execution to Claude 3.5 Sonnet via OpenRouter. Next month: $42. Planning quality indistinguishable; execution quality identical. Microsoft’s DeepSeek pivot announcement two weeks later confirmed we weren’t alone.

What We Use Now

Token-routed stack:

• Planning/Exploration: ollama run deepseek-v4:32b (or qwen2.5-coder:32b) via Continue.dev — local, free, unlimited, 128k context
• Execution/Code Gen: anthropic/claude-3.5-sonnet via OpenRouter — pay-per-token, ~$3/M in, $15/M out
• Budget enforcement: Continue.dev maxTokens per request + OpenRouter per-key daily caps ($5/day default)
• Fallback: If local model fails, auto-route to openrouter/anthropic/claude-3-haiku (cheap cloud planning)

Team convention documented in .continue/cost-policy.md:

“Default: local planning, cloud execution. Override only with // @premium comment. Budget alert at 80% daily cap.”

When You’d Choose Differently

• Copilot Business ($19/mo) still wins for teams <20 hrs/mo AI usage, no DevOps capacity for local models, or strict compliance requiring Microsoft data residency.
• Cursor Pro ($20/mo) if you need Composer’s multi-file editing UX and accept the context-loss tradeoffs on long sessions.
• Windsurf ($15/mo) if you want Cascade’s persistent context in a VS Code fork without managing local infra.

Tool Crucible Rating

Dimension	Rating (1–5)	Notes
Overall (Cowork)	2	Capable but economically broken for heavy users; no routing controls
Overall (Routed Stack)	5	70% savings, better control, same quality — but requires setup
Cost Predictability	1 (Cowork) / 5 (Routed)	Cowork: surprise bills. Routed: hard caps, visible per-task costs
Vendor Lock-in	1 (Cowork) / 5 (Routed)	Cowork: trapped in Microsoft pricing. Routed: swap any model/provider
Setup Effort	5 (Cowork) / 2 (Routed)	Cowork works instantly. Routed needs Ollama, Continue.dev, OpenRouter config

---

This is part of our AI Coding Cost Optimization series. See full comparison: GitHub Copilot Pricing 2026: Cowork Crisis, DeepSeek Pivot, and the BYOK Alternative