OpenCode Zen and Go — what they are, what they cost, and how Agenthood connects to them

Most people know OpenCode as the open-source AI coding agent with 160K GitHub stars and 7.5M monthly developers. But the same team behind the agent also runs two model API platforms — Zen and Go — that any agent framework can use, not just OpenCode itself.

Agenthood has supported both as first-class providers since v3.0.0. Here's what they are, what they cost, what models they offer, and how to use them.

OpenCode offers three ways to use models, from free to enterprise:

OpenCode ships with built-in free models — you can install the agent and start coding immediately without any API key. These are suitable for simple tasks and evaluation, but for serious agent work you'll want Zen or Go.

Zen is a curated set of models that OpenCode has tested and benchmarked specifically for coding agents. You pay per request with transparent pricing — no markups, no hidden fees.

• Pricing: Add $20 balance (one-time), set monthly spend limits, auto-top-up when balance hits $5
• Models: Hand-picked and validated for agent coding workflows
• Privacy: All models hosted in the US, zero-retention policy
• Use anywhere: Works with any agent framework, not just OpenCode

Go is a low-cost subscription aimed at making agentic coding accessible to more developers worldwide. Instead of pay-per-token, you get generous usage limits for a flat monthly fee.

• Pricing: $5 first month, $10/month thereafter
• Models: 14 open-source models including DeepSeek V4 Flash, DeepSeek V4 Pro, GLM-5.2, Qwen3.7 Max, Kimi K2.7 Code, MiniMax M3, MiMo-V2.5-Pro, and more
• Limits: Generous per-model caps (e.g., 30,100 requests per 5 hours for MiMo-V2.5, 31,650 for DeepSeek V4 Flash)
• Top-up: Additional credit available if you hit limits

OpenCode publishes real-time usage data at opencode.ai/data. Here are the top models as of July 2026:

DeepSeek V4 Flash dominates with 75% market share on the platform — and for good reason. At $0.28 per million tokens with a 96% cache hit ratio, the effective cost is closer to fractions of a cent per session.

OpenCode reports cache ratios above 90% for most models — DeepSeek V4 Flash and V4 Pro both hit 96%. This is because coding agents send highly repetitive system prompts and file contexts. With cache, the billed tokens are a fraction of the total sent:

• Average session cost on DeepSeek V4 Flash: $0.09
• Average session cost on MiMo-V2.5: $0.04
• Average tokens per session: 6.6M

Agenthood has supported OpenCode Zen and Go as LLM providers since v3.0.0. The integration is straightforward because both providers expose an OpenAI-compatible API.

Set one environment variable:

Env
OPENCODE_API_KEY=your_key_here

Then pass the provider flag when running any agent:

Bash
# Use OpenCode Zen (pay-as-you-go)
npx agenthood run the-scribe "write a commit message" --provider opencode

# Use OpenCode Go (subscription)
npx agenthood run the-reviewer "review the last commit" --provider opencode-go

The Agenthood runtime's provider abstraction maps both OpenCode providers to the same toOpenAIMessages converter, which handles DeepSeek's specific format requirements — snake_case tool calls, correct tool_call_id ordering, and proper message role sequencing.

The actual API endpoints are:

• Zen: https://api.opencode.ai/zen/v1
• Go: https://api.opencode.ai/zen/go/v1

By default, OpenCode Go routes to deepseek-v4-flash. You can override the model with the OPENCODE_MODEL environment variable or a --model flag.

Having two provider options with different pricing models is useful:

• Zen is ideal for CI pipelines and automated workflows where you want per-run cost tracking and no monthly commitment
• Go works better for local development and interactive sessions where a flat monthly fee is simpler to budget

Both providers support the same model family, so switching between them doesn't change agent behavior — only the billing model.

Use Go for daily development. At $10/month with generous limits, it's cheaper than almost any pay-per-token alternative for regular use.

Use Zen for CI and scripts. The pay-as-you-go model means you only pay for automated runs. Set a monthly spend limit to avoid surprises.

Take advantage of the cache. Coding agent sessions are highly cacheable. The 96% cache hit rate on DeepSeek V4 Flash means you're billed for roughly 4% of the tokens you actually send.

Watch the model rankings. OpenCode publishes live data at opencode.ai/data. Models change in quality and cost — the leaderboard is a good way to spot shifts.

Full disclosure: This site uses OpenCode Go for the AI assistant on every page, and Agenthood uses OpenCode providers in its CI pipelines. The API key is set via OPENCODE_API_KEY in the deployment environment.