OpenCode launched on June 19, 2025 as an MIT-licensed terminal coding agent: a Go CLI with a Bubble Tea TUI, SQLite-backed session storage, and a Bun/JavaScript server underneath. A year on, the project reports 900+ contributors, 13,000+ commits, and v1.16.0 landed on June 5, 2026. Those numbers put it in the same conversation as Claude Code and Codex CLI for daily terminal use - but the design goal was different from the start.
The Pitch: One CLI, Any Model
Where Claude Code is built around Anthropic's models and Codex CLI around OpenAI's, OpenCode is explicitly provider-agnostic. It supports 75+ providers - Anthropic, OpenAI, Google, Mistral, AWS Bedrock, Groq, OpenRouter, and local models via Ollama - and lets you switch providers mid-session without reconfiguring your project.
In practice that means a workflow like: explore a codebase and draft a plan with a cheap, fast model, then switch to a stronger model for the actual implementation, then switch again to a local model for anything you don't want leaving your machine. The agent, the session history, and your terminal habits stay the same - only the backend model changes.
Pricing: Bring Your Own Key, or Don't
OpenCode itself is free and open source. Three ways to pay for the models it calls:
- Free (BYOK): $0 for OpenCode - you pay your model provider directly at their normal API rates.
- Go plan: $5 for the first month, $10/month after, bundling provider access.
- Zen plan: $20/month plus pay-per-use for heavier usage.
The BYOK option is the one worth noting: it decouples your tool from any single vendor's subscription, which matters if you're already tracking token costs across providers (see our guide on AI cost modeling).
Where It Fits Next to Claude Code and Codex CLI
On the official Terminal-Bench 2.0 leaderboard, Codex CLI with GPT-5.5 is listed at 82.2%, Gemini CLI with Gemini 3.1 Pro at 61.4% and 59.4% across two entries, Claude Code with Claude Opus 4.6 at 58.0%, and an older OpenCode run with Claude Opus 4.5 at 51.7%. Treat those numbers as a snapshot, not a universal ranking: agent versions, model choices, and evaluation dates differ, and OpenCode's result changes with the model you point it at.
The official Terminal-Bench site also lists Terminal-Bench 2.1 as active, but its benchmark page currently says the tasks have not been uploaded yet. Until that leaderboard is public, the safer comparison is not "who won 2.1?" but the design trade-off: vendor-specific CLIs optimize a tight model-tool pairing, while OpenCode optimizes for a workflow that can survive a provider outage, a pricing change, or a better model from a different vendor next quarter.
What This Means for Your Setup
- Not a verdict on quality: OpenCode's growth is a signal that "model-agnostic" has real demand, not proof that it outperforms vendor-specific agents on a given task. Output quality still depends mostly on the model you select.
- Useful for cost control: BYOK plus mid-session model switching gives you a lever - cheap models for exploration, stronger ones for the final diff - that's harder to get inside a single-vendor subscription.
- Local models are a first-class option: Ollama support means sensitive codebases can stay on a local model without leaving your terminal workflow, relevant if you've read our guide on sanitizing code before sending it to AI or on local and private AI models.
- Try it alongside, not instead of: Most teams reporting good results run a model-agnostic CLI for exploration and cost-sensitive work, and keep a vendor-specific agent for the highest-stakes diffs - consistent with the "run more than one tool" pattern covered in our multi-agent coding stack roundup.
If you're picking your first CLI tool, our CLI-first AI development guide and tool comparison for vibe coders cover the fundamentals that apply regardless of which agent you choose.