What Is OpenCode Go?

OpenCode Go is a $10/month subscription to a curated set of coding models, sold by Anomaly, the company behind OpenCode. You subscribe, get an API key, and plug it into OpenCode or another coding agent as a model provider. It isn't an agent. It doesn't edit files or run commands. It's the thing the agent calls.

What $10 buys: included usage on 31 paid models plus 2 limited-time free models (33 total, as of September 26, 2026). OpenCode sets each model's allowance as a monthly dollar amount of $15, $30, or $60 at Go's token prices. Each model also has a 5-hour cap (20% of its monthly amount) and a weekly cap (50%).

At the limit: requests stop until the window resets, unless you turn on Use balance, which bills the overflow to a pay-as-you-go Zen/Console balance. OpenCode says free models keep working.

Privacy isn't uniform. Most Go models are listed as 0-day retention and no training. Grok and GPT Luna models are retained for 30 days. The two Muse Spark Contributor models are used for training. DeepSeek's zero-retention agreement runs only through September 30, 2026 unless OpenCode renews it.

Outside OpenCode: OpenCode says Go works "with any agent" and lists Claude Code, Codex, and five other clients as validated, with conditions on user agent and session headers.

OpenCode Go vs OpenCode vs Zen/Console vs free models

Search results, YouTube titles, and even some reviews blur these together. They're separate things you pay for (or don't) separately.

OpenCode's docs say Go is "completely optional and you don't need to use it to use OpenCode." Go and Zen/Console share an account and console, which is why they get confused. They're still different products with different model lists and different billing.

OpenCode, the agent itself: what it does, the current v2 install, providers and local models, and what's free

entity↗

On the naming: the current docs at /docs/zen still say "Zen." The v2 docs call the same service "OpenCode Console" and "the OpenCode pay-as-you-go provider." The API endpoints still sit under opencode.ai/zen/. It's one service under two names.

What $10 a month gets you

About "$5 for the first month": Google's AI Overview for "opencode go pricing" and at least one third-party model directory still said on September 26 that Go costs $5 for the first month. OpenCode's current product page, FAQ, and docs say $10/month and don't mention an introductory price. Treat $5 as out of date unless you see it at checkout.

Some older reviews also describe Go as a handful of Chinese open models. That's no longer accurate. The roster now includes Grok, GPT Luna, and Meta's Muse Spark Contributor models alongside GLM, Kimi, Qwen, MiniMax, MiMo, DeepSeek, LongCat, and Hy.

Current OpenCode Go models (33, as of September 26, 2026)

OpenCode's product FAQ and both versions of its Go docs list the same 33 models. The docs warn: "The list of models may change as we test and add new ones." It changes often, so check the date on any list you read, including this one.

The monthly limit is the dollar value of included usage for that model at Go's token prices. The request estimates are OpenCode's own, based on assumed tokens per request that are mostly cached input (for example, about 55,000 cached, 1,000 fresh input, and 200 output tokens for GLM-5.3-Flash). Your requests will differ, so use these to compare models, not to plan a workload.

By tier, that's 11 models at $15, 4 at $30, and 16 at $60, plus the two free ones. The strongest-sounding names (Grok 4.7, Kimi K3, Qwen3.8 Max, GLM-5.3, DeepSeek V4 Pro) mostly sit in the $15 tier with the lowest request estimates.

OpenCode explains the spread: it gets bulk discounts or reserved GPU capacity on most models and passes that on as higher allowances. For new models, or models whose public price is already low, it hasn't, and you get "a little more than if you paid the model providers directly."

Because limits are dollar amounts, a request billed at the peak rate uses twice as much of a DeepSeek model's allowance as the same request off-peak. That's our arithmetic from the published prices, not something OpenCode spells out.

How OpenCode Go usage limits work

This is the most confusing part of Go, and OpenCode's own pages don't fully explain it.

What OpenCode publishes

The Go docs say:

Each model has the following usage limits: 5-hour — 20% of the monthly limit; weekly — 50%; and monthly — 100%.

They illustrate it with a $60 model: $12 per 5 hours, $30 per week, $60 per month.

The product page FAQ says Go's quotas are "roughly equivalent to base allowances of $12 per 5 hours, $30 per week, and $60 per month; model-specific allowances may differ."

Applied to each tier, the documented math is:

The 5-hour window is rolling. Because the weekly cap is half the monthly one, you can't spend a model's whole month in one week. Because the 5-hour cap is 20% of the monthly one, you can't spend a week's worth in one sitting either.

What OpenCode doesn't say

The docs describe caps per model. The FAQ describes base allowances for the plan. Neither says how the two relate. Two readings fit the official wording:

1. Separate budgets. Each model has its own $15, $30, or $60 per month, independent of the others. Using up Kimi K3 wouldn't touch your GLM allowance.
2. One shared pool with per-model ceilings. The plan has one $12 / $30 / $60 pool, and each model's limit caps how much of that pool it can use. Spending on one model leaves less for all the others.

OpenCode's documentation doesn't settle this. We haven't tested it, so we can't either.

OpenCode also says: "Usage limits may change as we learn from early usage and feedback."

What happens when you hit the limit

Use balance is something you enable, and it needs prepaid credit. With it on, requests past your Go limits are charged per token from that balance at pay-as-you-go prices. OpenCode doesn't say whether the fallback is charged at Go's rates or Zen/Console's. For some models they differ: DeepSeek V4 Pro lists at $1.74 input / $3.48 output per million tokens on Zen/Console, against $0.66 / $1.98 off-peak on Go.

Zen/Console balances can auto-reload. By default, OpenCode's docs say, a balance that drops below $5 reloads $20. If you turn on Use balance, check your auto-reload and monthly spend limit settings.

Privacy and training by model

The roster table above has each model's status. Here's what's different and why it matters for code.

• Most Go models: OpenCode lists them as "Not used" for training with 0 days retention.
• Grok 4.7 and 4.6: not used for training, 30 days retention. OpenCode notes that xAI's zero-data-retention mode disables API features that depend on stored data, including the stateful Responses API, Files and Collections, and the Batch API.
• GPT 6 Luna and GPT 5.6 Luna: not used for training, 30 days retention. OpenCode says OpenAI generates abuse-monitoring logs for all API usage and keeps them up to 30 days.
• Muse Spark 1.3 Contributor and 1.2 Contributor: used for training and not zero-retention. OpenCode repeats Meta's terms: "Heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models." In a coding agent, prompts and completions include your source code. These are also the models with the highest request estimates on Go.
• DeepSeek: listed at 0 days, with a condition. The Go docs say: "ZDR agreement is renewed monthly. The current agreement is valid through September 30, 2026."

For more on the Contributor tradeoff, including Meta's own Standard and Contributor pricing, see our Muse Spark page.

Muse Spark: Meta's Standard vs Contributor tiers, what each costs, and what Contributor lets Meta train on

entity↗

Go vs Zen/Console: which billing fits

Go makes sense when your usage is steady and fits inside the caps of the models you like, especially the $60-tier ones. Zen/Console makes sense when you need models Go doesn't carry, when your usage is bursty enough to blow through 5-hour caps, or when you want one bill per token with no windows. Lots of people will want both: Go for everyday work and a Zen/Console balance behind Use balance for overflow.

Using OpenCode Go outside OpenCode

OpenCode says you can use Go "with any agent." The docs add conditions, because Go is "designed for OpenCode and other coding agents that produce similar types of requests," and traffic is monitored for abuse.

OpenCode lists these clients as validated, adding that it doesn't guarantee they'll keep working:

OpenCode also lists known problematic clients with missing or partial session support: DeepSeek Harness, GitHub Copilot Chat, Kimi Code, and MiMo Code.

What "validated" covers for Claude Code: the entry is about session-header handling. It doesn't say every Go model works through Claude Code, or which endpoint format Claude Code should use. Go's endpoint table serves only MiniMax and Qwen models on the Anthropic-style /messages route. The rest are listed on /chat/completions or /responses. OpenCode doesn't document whether other models can be reached through /messages. We haven't run Go through Claude Code, so we're not publishing setup steps or a compatibility list.

Endpoints and model IDs

The models endpoint is thinner than it sounds. When we requested it without a key on September 26, it returned 43 model IDs with only an ID, a timestamp, and an owner, with no prices, limits, or context sizes. Ten of those IDs aren't in the documented roster, including models OpenCode lists as deprecated on Zen (such as minimax-m2.5 and kimi-k2.5) and an unlisted omen-alpha. Don't read the endpoint as the list of models your subscription includes. The documented roster is the better guide.

Tracking your usage

OpenCode's only documented method is the web console: "You can track your current usage in the console." The docs don't describe what the console shows. Screenshots in GitHub issues show rolling, weekly, and monthly usage percentages next to a Usage History of per-model dollar amounts.

Several open issues (#41206, #41391, and others linked from them) report that the dashboard's percentages don't match the dollar amounts in Usage History. Issues #41125, #42935, and #50258 report prompt caching dropping out on DeepSeek models, which makes requests bill at full input rates and burn allowance faster. We can't confirm any of these without a subscription.

Known issues with OpenCode Go right now

Open reports in OpenCode's GitHub repository as of September 26, 2026. They're unverified by us, and we're not listing every one.

• Model picker gaps. Issue #51212 reports that the OpenCode Go section of the /models picker in OpenCode 1.18.32 left out 10 models the Go API returned, including every DeepSeek model, and that those models still worked when selected directly. That's one report, not proof the roster is wrong. If a Go model you expect is missing, try setting opencode-go/<model-id> directly.
• Quota accounting. Multiple reports that usage bars fill faster than Usage History suggests (see above).
• Caching on DeepSeek. Multiple reports of lost prompt-cache hits.
• Endpoint errors. Reports of intermittent 400, 403, 500, and 503 responses on specific models or routes.

Who OpenCode Go may fit

Picking the model inside the subscription is still a separate decision from picking the subscription.

How to choose a model for a task by reasoning depth, latency, and cost

concept↗

The production coding-agent workflow, for checking whatever model you run before trusting its diffs

workflow↗

What we haven't verified

• How far any model's allowance goes in real work, or how many tasks fit in a 5-hour window.
• Whether Go's limits are separate budgets per model or one shared pool.
• Whether Use balance charges Go or Zen/Console token prices.
• Whether free models stay usable after the Go limit is hit, as OpenCode says, or get blocked, as one user reports.
• What DeepSeek retention applies after September 30, 2026.
• Which Go models work through Claude Code, Codex, or other validated clients, and on which endpoint.
• Whether the /models picker gap in #51212 affects current OpenCode builds generally.

How we separate vendor documentation, community reports, and our own inference

concept↗

Change log

Sources & Documentation

1. OpenCode: Go product page and FAQ — $10/month, 33-model list, request estimates, base allowances ($12/5h, $30/week, $60/month), free-model quota of 200 requests/day, privacy table, refund terms.
2. OpenCode docs: Go — Per-model limits (20% / 50% / 100%), pricing and monthly limits, request estimates and assumptions, Use balance, validated clients, endpoints, models endpoint, privacy and the DeepSeek ZDR date. Last updated September 26, 2026.
3. OpenCode v2 docs: Console / Go — Same roster and limits under the Console naming. Use balance falls back to the pay-as-you-go balance.
4. OpenCode docs: Zen — Pay-as-you-go catalog and pricing, free models, auto-reload, monthly limits, deprecated models (including MiniMax M2.5), Zen privacy and US hosting.
5. OpenCode v2 docs: Console / Models — Zen renamed OpenCode Console, the 'OpenCode pay-as-you-go provider.'
6. OpenCode: Go models endpoint — Returned 43 model IDs without prices or limits when requested on September 26, 2026.
7. OpenCode: home page — OpenCode as the open-source coding agent, separate from Go.
8. GitHub: anomalyco/opencode issue #41391 — Community report: quota accounting vs Usage History, and a non-maintainer's proposed multiplier.
9. GitHub: anomalyco/opencode issue #49014 — Community report: one model's 5-hour cap blocked all Go models.
10. GitHub: anomalyco/opencode issue #49020 — Community report: undocumented /zen/go/v1/usage returns 403 for an active subscription.
11. GitHub: anomalyco/opencode issue #51212 — Community report: /models picker omits 10 Go models.
12. GitHub: anomalyco/opencode issue #42035 — Community report: identically named Zen and Go models cause 'subscribe to Go' errors.
13. GitHub: anomalyco/opencode issue #47703 — Community report: free models blocked along with Go models at the limit.