OpenCode Go: $10 a month for eighteen models

Most subscription pitches in this space are a flat fee for one vendor's models. OpenCode Go is a different shape: $5 for the first month, then $10/month, for access to eighteen open-weight coding models behind one API key — and it works with any agent that speaks the OpenAI or Anthropic API, not just OpenCode's own client.

That last detail is the one that makes it interesting, and it is the one the marketing underplays.

What you get

The stated design goal is roughly six times the subscription price in usage value, funded by bulk discounts and reserved capacity. At $10 for up to $60 of metered usage, that is the arithmetic.

Read the caps properly

The three limits are the whole product, and they are dollar-denominated, not request-denominated. That has consequences people discover the annoying way.

Because the caps are in dollars, your effective request count depends entirely on which model you pick and how much context you push. Running a large-context model hard will exhaust $12 in an afternoon; running a small one on short prompts will not come close. Two people on identical plans can have completely different experiences and both be describing it accurately.

The 5-hour cap is the one that bites. $12 per rolling five hours is generous for normal work and restrictive for a long agentic session that re-reads a large codebase on every turn. If you routinely run agents that chew context, budget around that window rather than the monthly figure.

The part that makes it worth more than it looks

You subscribe through OpenCode Zen, get an API key, and connect it with the TUI's /connect command. From there it behaves like any other provider.

But because the endpoint is OpenAI/Anthropic-compatible, that key is not locked to OpenCode's client. Point your own agent loop at it. Point a framework at it. Use it as the cheap tier in a routing setup while keeping a frontier key for the hard cases. The subscription becomes a general-purpose cheap-inference budget rather than a single tool's licence.

That also limits the lock-in. If you stop paying, you change a base URL and an API key — you are not unwinding an integration.

Who this is actually for

• Anyone currently paying per-token for open-weight models. If your metered spend on this class of model is above about $10/month, the subscription is strictly better until you hit the caps.
• People who want to compare models without eight accounts. Eighteen models behind one key is a genuinely good evaluation harness. Running the same regression set across several of them is a couple of hours' work.
• Cost-sensitive agent builders. As the cheap tier in a routed system, with something stronger held in reserve for escalation.

Who it is not for

• Anyone who needs a specific frontier model. The catalogue is open-weight. If your work depends on a particular closed model's behaviour, this does not replace it — though it may still be worth holding alongside.
• Heavy continuous agentic use. You will meet the 5-hour cap regularly. Metered API access may genuinely suit you better, which is an unusual thing for a subscription to admit.
• Anyone who cannot accept 0–30 day retention. Check the per-model table against your own data policy before you connect anything real.

The honest summary

The catalogue is explicitly stated to change as models are tested and added, so treat any specific model list — including this one — as a snapshot. What is stable is the shape: a small monthly fee, dollar-denominated caps, and a compatible endpoint you can point anything at.

For $5 in the first month, the evaluation costs less than deliberating over it. Run your own regression set across three or four of the models and you will know within an afternoon whether the caps fit your usage.

You can subscribe at opencode.ai/go.