OpenCode Go Usage Limits: Which Models Last, Which Burn Out

In our run, eight models outlasted the five-hour window and eight ran dry within the hour

Runway is how long an agent can call a model nonstop before its five-hour share is gone: the requests the share buys in a 100k-token session, times the seconds each call took in our run. Real sessions also wait on tools and on people, so in practice a window lasts longer than this.

• Long-lasting and quick in our run: Muse Spark 1.3 Contributor, Muse Spark 1.2 Contributor and DeepSeek V4.1 Flash outlasted the window and answered in about 3 seconds a call. MiMo-V2.6-Flash lasted longest of all, but at 11.8 seconds a call it spent 56 seconds of model time per task against 16 for Muse Spark 1.3 Contributor.
• Quickest to run out: GLM-5.2, Gemini 4 Argon, Qwen3.8 Max, Kimi K3.1 Pro, Step 5 Preview, GLM-5.3, Grok 4.7 and Kimi K3. Grok 4.7 cost $0.0913 per task in our run, 41× DeepSeek V4.1 Flash. Go Plus caps for these models are 3× to 8× Go's, at 4× the price.
• Session length moved the projected counts most: Kimi K3.1 Pro gets 199 requests per five hours at 50k and 85 at 150k. Claude Haiku 5.5 prices rise 5× above 100k, so its runway falls from 1.7 h at 100k to 15 min at 150k.
• DeepSeek ran off-peak: our run and OpenCode's published counts both use DeepSeek's off-peak price. On weekdays, 01:00–04:00 and 06:00–10:00 UTC, DeepSeek charges 2×. If Go's meter bills at the peak rate, DeepSeek counts would halve: DeepSeek V4.1 Flash would still outlast the window (12 h), while DeepSeek V4 Pro would drop from 1.2 h to 37 min.

These are observations from one run of our own test: a small Node project, four tasks, one session per model. Every model passed every task, so the groups reflect cost and speed in that run, not capability, and not how a model will behave on anyone else's code. Other projects, prompts and session lengths will give different numbers. The models table carries the broader Artificial Analysis scores.

One session, four tasks, every call logged and priced

We gave each of the 28 paid Go models a fresh copy of a small Node project with an AGENTS.md, a test suite and one real bug. OpenCode v2.1.4 ran its default build agent with every tool call approved, and in one session each model had to:

1. find why npm test fails, fix it and rerun the suite;
2. bill a new cache_write token count, with tests;
3. answer a question about the allowance windows without editing a file;
4. rename a function across src/ and test/ and rerun the tests.

Hidden checks confirmed the fix, the billing change, the rename and a passing suite. All 26 models that ran passed every check. MiMo-V2.5 and MiMo-V2.5-Pro, still on the pricing page, were refused by the server as deprecated. The run spent $1.70 of Go allowance.

OpenCode's session store logged each model call, which is what the Go console counts as a request, with its tokens, cost and time. Our sessions were short: median prompts of 11k to 21k tokens, 34k at most. Every agent step re-sends the conversation, so we projected each model to longer sessions. Its measured fresh input, cache writes and output per call stay fixed; the rest of the prompt is re-read from cache and priced at the Go tier for that size. The count is the five-hour share, a fifth of the monthly cap, divided by that cost. Runway multiplies the count by the model's mean seconds per call.

The groups use 100k, just above the largest prompt OpenCode's own profiles assume.

DeepSeek's calls were billed at the off-peak rate, $0.15 per million fresh input tokens, which is also the price behind OpenCode's published counts. Gemini 4 Argon is priced at its discounted launch rate, as OpenCode prices it, and has no published profile, so it is compared at 55k.

Each model ran once, 11 to 28 calls, so a repeat could shift individual figures by several points. Runway counts model time only, and longer prompts answer more slowly, so a real window lasts longer than the figure. The weekly cap is half the monthly one, two and a half five-hour shares: a model that empties every window hits it in its third. Cache writes are an open question. OpenCode's client prices them at the listed rate, and seven models wrote them (Claude Haiku 5.5, GPT 5.6 Luna, GPT 6 Luna, Kimi K3.1 Pro, Qwen3.7 Plus, Qwen3.8 Flash and Qwen3.8 Max), but the Go page doesn't say whether its meter counts them. If it doesn't, those models gain.

At OpenCode's own session size, most published counts hold

OpenCode Go costs $10 a month. The usage-limits page turns each model's cap into a request count using a token profile: a little fresh input, a large prompt re-read from cache, a short reply. For Kimi K3.1 Pro that is 1,050 fresh tokens, 76,500 cached and 300 output, priced at $0.0204, which gives 147 requests per five hours.

Priced at the prompt size its profile assumes, each model's measured request lands within 3% of the published count for seven models, among them Gemini 4 Argon, GPT 5.6 Luna and Kimi K3.1 Pro.

Eight fall short by a fifth or more, and the cause is output. OpenCode assumes Grok 4.7 writes 120 tokens per request; it averaged 623, most of it reasoning. GLM-5.3 wrote 638 against an assumed 150, and Claude Haiku 5.5 616 against 240. Those three get 69% to 75% of what the page says.

Four published counts don't follow from OpenCode's own profile at all. Kimi K2.7 Code's 1,350 is 36% above what its profile and prices give, and Kimi K3 (12%), GLM-5.3 (12%) and GLM-5.2 (11%) are over too.

In our run the method held up. Several of its inputs were generous.

The session you carry decides the count

OpenCode's profiles assume prompts of 33k to 90k tokens. Fresh input and output barely change as a session grows; the cache re-read grows with every step, and even at cache prices it ends up as most of the bill. Kimi K3.1 Pro gets 199 requests per five hours at 50k, 119 at 100k and 85 at 150k. DeepSeek V4.1 Flash, published at 43,330, gets 28,590 at 100k.

Claude Haiku 5.5 has a cliff. Above 100k tokens every price on it rises 5×, and its count falls from 1,922 at 100k to 291 at 150k.

Go Plus raises the expensive caps most

Go Plus costs $40, four times Go. Its caps are 2× Go on DeepSeek V4.1 Flash, MiMo-V2.6-Flash, Muse Spark 1.2 Contributor and Muse Spark 1.3 Contributor, 3× on six models, 4× on most, and 8× on GLM-5.3. The 2× models are the cheap ones, where Go rarely runs out: even at 150k, DeepSeek V4.1 Flash gets 23,542 requests per five hours. In our run, the caps that ran out first were in the minutes group, and Plus takes Kimi K3.1 Pro at 100k from 119 to 479.

Of the four models with a 2× Plus cap, MiMo-V2.6-Flash, Muse Spark 1.3 Contributor, Muse Spark 1.2 Contributor and DeepSeek V4.1 Flash already outlasted the five-hour window on Go in our run.