This page is for everyone using my shared OpenCode Go setup, including the people I gave access to through the giveaway. The goal is simple: keep the subscription useful for all of us while also understanding how AI coding agents spend usage. Start with the high-volume models, escalate only when the task proves it needs more, and check the live usage before starting long agentic runs.

Live usage

Loading usage...

Opencode Go 1

Loading usage...

5h --
-- left reset unknown
Weekly --
-- left reset unknown
Monthly --
-- left reset unknown

Opencode Go 2

Loading usage...

5h --
-- left reset unknown
Weekly --
-- left reset unknown
Monthly --
-- left reset unknown

Last checked: --

The rule

Use Tier 1 by default. DeepSeek V4 Flash, Qwen3.5 Plus, and MiniMax M2.5 are not fallback trash; they are the models that let everyone in the shared pool code without burning through the giveaway access. They are good enough for most edits, reviews, test fixes, search, docs, and normal iteration. Tier 2 is for normal multi-file engineering. Tier 3 is for work that is stuck, architectural, or genuinely long-horizon.

Tier 1

Volume workhorse

Start here unless you already know the task is complex.

DeepSeek V4 Flash 79% SWE-Verified 31,650 req/5h
Qwen3.5 Plus 76.4% SWE-Verified 10,200 req/5h
MiniMax M2.5 80.2% SWE-Verified 6,300 req/5h
Escalate if the model is stuck, not just because a bigger model exists.
Tier 2

Standard engineering

Use for multi-file features, terminal-heavy debugging, and serious implementation work.

DeepSeek V4 Pro 55.4% SWE-Pro 3,450 req/5h
Qwen3.6 Plus 56.6% SWE-Pro 3,300 req/5h
MiniMax M2.7 56.2% SWE-Pro 3,400 req/5h
Escalate again only for planning, architecture, or failed Tier 2 attempts.
Tier 3

Use sparingly

Save these for complex refactors, hard architecture calls, or long autonomous runs.

Kimi K2.6 58.6% SWE-Pro 1,150 req/5h base
GLM-5.1 58.4% SWE-Pro 880 req/5h
MiMo-V2.5-Pro 57.2% SWE-Pro 1,290 req/5h

How the agent works

An AI coding agent is not just a chat box with a nicer terminal. It runs a loop: send the prompt and current context to the model, let the model decide whether it needs tools, execute those tools locally, feed the results back into the model, then continue until it can answer or make a change. That is why coding agents feel useful: they can read files, search the repo, edit code, run tests, inspect failures, and try again without you manually pasting every intermediate result.

The same loop is also why shared usage needs discipline. The model call is the scarce part, not the local file read or shell command by itself. Every time the agent needs to reason over new tool output, recover from a failed test, summarize a long context, or delegate to another agent routed through Go, it may create another provider request. Good prompts reduce wandering, but model choice still matters more than people expect.

Capacity

The request budget is not evenly distributed. One heavy agent session can burn 50-200 requests. That is harmless on DeepSeek V4 Flash and painful on GLM-5.1 or Kimi K2.6. If you are using the shared giveaway key, the responsible default is obvious.

Requests per 5 hours

Approximate base routing budget. Temporary promos are a bonus, not the policy.

DeepSeek V4 Flash 31,650
Qwen3.5 Plus 10,200
MiniMax M2.5 6,300
DeepSeek V4 Pro 3,450
MiniMax M2.7 3,400
Qwen3.6 Plus 3,300
MiMo-V2.5-Pro 1,290
Kimi K2.6 1,150
GLM-5.1 880

What a request means

A request is not one chat message. In OpenCode Go, the usage limits are dollar-denominated, and the published request counts are estimates based on the average token shape of a model call: input tokens, cached context, and output tokens. So the useful mental model is: one request is one model call made through the Go provider, not one thing typed into the terminal.

That matters because OpenCode is an agent loop. A single prompt can ask the model what to do, run reads or searches, feed those tool results back to the model, make edits, run tests, read the failures, and ask the model again. The local tool execution is not the expensive part by itself; the repeated model calls around those tool results are. Long context, subagents, retries, and automatic compaction can also create extra model calls when they are routed through Go.

1 short answer Usually one model call.
1 code edit Often several calls: inspect, edit, verify, respond.
1 agentic task Can be dozens of calls if it keeps reading, testing, and retrying.
1 stuck run Can reach 100+ calls if it loops through tools or subagents.

This is the real reason to start on Tier 1. High-volume models absorb normal agent behavior without punishing everyone else on the shared key.

Practical defaults

For most tasks, ask for Tier 1 first: “use DeepSeek V4 Flash unless you get stuck.” Move to Tier 2 when the task spans multiple files, needs terminal-heavy debugging, or the first pass misses the shape of the fix. Use Tier 3 only after you can explain why Tier 1 or Tier 2 is not enough.

For shared or giveaway access, do not route every agent through Kimi, GLM, or MiMo Pro. Do not upgrade search, exploration, or simple review agents. If live 5-hour usage is already high, keep new work on Tier 1 or wait for the reset.

How this page works

The widget polls /api/opencode-usage every 60 seconds. A Cloudflare Pages Function fetches both OpenCode Go dashboards and billing pages with cookies stored in Pages environment variables, extracts the rolling, weekly, and monthly usage windows plus each Go expiry date, then returns only sanitized JSON. The individual meter feeds are also available at /api/usage/opencode/1 and /api/usage/opencode/2. Successful reads are edge-cached for 45 seconds.

API

Use the JSON directly if you want to check usage outside this page:

1
2
3
curl https://nihilism.gay/api/opencode-usage
curl https://nihilism.gay/api/usage/opencode/1
curl https://nihilism.gay/api/usage/opencode/2

The combined endpoint returns accounts; each account exposes only id, label, ok, expiresAt, and windows. Each window includes usedPercent, remainingPercent, resetAt, and resetInSeconds. No cookies, workspace names, workspace IDs, or payment details are returned.