opencode v2.x Free Tier Dries Up After One Conversation: A Metering Change Breaks the Daily Quota

Follow my public account: [编程朝花夕拾] to get first-hand content.

01 Introduction

Since opencode upgraded to 2.x, free quota gets exhausted after a few conversations, sometimes even in one conversation, which is really frustrating. My first reaction was that it's impossible to use up the quota. Two days ago I used the free quota for a whole day's tasks, and even shared an article about using the OC tool to import opencode's free quota into Workbuddy in one click.

Lo and behold, these two days it can't be used. I went to anomalyco/opencode and searched, and found that the comment section is exploding, with all kinds of complaints about the free quota issue. Let's sort this out together.

02 Origin

After downloading version 1.19.5 from GitHub, when working on a project, the model simply stopped working and told me to download from https://opencode.ai/download. I already have the latest version, why do I need to download again?

(Image)

With no other choice, I had to download from the official website:

(Image)

After downloading, I found it's already version 2.2.5.

(Image)

The new version indeed changed many things. Most directly, it puts free models separately for easy selection, which is a pleasant surprise. No need to pick from the full list.

(Image)

Just when I was overjoyed, I wanted to use deepseek v4.1 flash to optimize a small plugin I wrote earlier, and after two rounds the quota was exhausted, needing to wait until the next day for reset.

(Image)

What's going on here?

03 Tracking

opencode has 224k stars on GitHub, which shows a huge number of people paying attention. Others must be facing similar problems.

(Image)

So I looked through issues, issues about the "free" keyword:

(Image)

The dense issues are in English, so I used AI to analyze the timeline of free quota exhaustion, which is much more convenient than looking one by one.

(Image)

It seems v2.x introduced many usage thresholds and also brought some problems. As of now, they haven't been fixed.

04 v2.x Experience Gets Worse

4.1 Change in Metering Unit

In issue #58751, a reporter directly pulled the data from local sqlite. He used opencode/deepseek-v4.1-flash-free, same account, same machine:

(Image)

In the first half, two lines are 6 seconds apart with almost the same volume but opposite outcomes — the reporter judges this is a rolling bucket at the boundary, not a hard per-request limit. His old session with 985 user messages and about 447k token assembled context returns provider.free-limit on every build; a new session with 23k tokens created in the same minute returns normally 18 seconds later.

In the second half, heavy usage on the 20th–21st had zero errors, but from the 24th the first request was blocked. The reporter's own judgment is "Something changed on the service side around 21–23 Oct". This timing coincides exactly with the batch of rate-limiting infrastructure in v2.1.2 (10/21) and the pinned post (10/26) opening.

So a more accurate description of "quota got smaller" is: the metering model changed from a lenient daily count to a rolling budget deducted by token/request characteristics. Your quota number may not have changed, but each round of dialogue has become more expensive.

4.2 Sessions Get Fatter

One request can consume what used to take ten times as much. v2 agents read files more eagerly, LSP context is more complete, and long sessions rely on auto-compaction; a single build request's assembled context can easily reach hundreds of thousands of tokens. Add the detail from #58205's log: within the same second, the main model reports "Free usage limit reached" while the small title-generation model gpt-5.4-nano reports "Insufficient account funds". Even invisible requests like automatic title generation are knocking on the free pool's door. #58176 and #58521 specifically record that title generation under the free tier is rejected with 403 FreeTierError.

05 Summary

opencode's free model used to be quite robust, but now suddenly being heavily rate-limited may make many developers stay away. If you really need a model for your work, you'll need a stable BYOK. Free models are either rate-limited or reduced in intelligence, which may bring some trouble.