OpenCode Makes DeepSeek V4.1 Flash Free: What the Docs Confirm and What to Watch

OpenCode now offers DeepSeek V4.1 Flash Free at $0 per million tokens on its Zen provider, and the official documentation confirms it. The catch is in the wording: the page says the model is free "for a limited time" and gives no end date, and the separate $10-a-month Go subscription treats the same model as a paid, metered one. This post separates what the pages actually say from what circulated on social media, and shows how to use the free tier without a surprise bill later.

For background on the tool itself, see our OpenCode guide. For the model, see our coverage of the DeepSeek V4.1 Flash API beta.

TL;DR: the claim, the check, the answer

What the OpenCode pages actually say

OpenCode runs two commercial surfaces, and the confusion in the headlines comes from mixing them up.

Zen (the Console) is pay-as-you-go. You add credits, get an API key, run /connect in the terminal UI, pick the OpenCode provider and then /models. Its pricing table has a block of zero-cost entries: Big Pickle, DeepSeek V4.1 Flash Free, Space Bunny Free, Cat Jam Free, LongCat 2.5 Preview Free, Step 5 Preview Free, Exo Free, MiMo-V2.6-Flash Free, MiMo-V2.5 Free, Nemotron 3 Ultra Free, Nemotron 3.5 Lightning Free and Muse Spark 1.3 Contributor Free. The paragraph below the table is specific about ours: "DeepSeek V4.1 Flash Free is free on OpenCode for a limited time. Its provider follows a zero-retention policy and does not use your data for model training."

The paid sibling, plain DeepSeek V4.1 Flash on Zen, is listed at $0.30 per million input tokens, $1.20 output and $0.006 cached read, which matches DeepSeek's own peak-hour rates.

Go is a subscription. Go costs $10 a month and Go Plus $40. Usage is defined in monthly dollar amounts, split into a 5-hour window (20 percent of the monthly limit), a weekly window (50 percent) and the monthly cap. For DeepSeek V4.1 Flash the monthly limit is listed at $80 on Go and $160 on Go Plus, up from smaller numbers earlier this month according to secondary reports. Token prices on Go are off-peak $0.15 input and $0.60 output, and double at peak. OpenCode's own estimate for typical coding-agent traffic is about 34,670 requests per five hours, 86,670 per week and 173,330 per month on Go.

So the statement "OpenCode made DeepSeek V4.1 Flash free" is true for Zen's free endpoint and misleading for Go, where the model is simply a very cheap, generously allowanced paid model. The Go page marks only three models as free (LongCat 2.5 Preview Free, Step 5 Preview Free, Cat Jam Free), and it tells subscribers that if they hit their limit "you can continue using the free models."

Why V4.1 Flash is the model to give away

DeepSeek renamed its small flagship: the pricing page now says to use deepseek-flash, the legacy deepseek-v4-flash names are retired and served by DeepSeek-V4.1-Flash, and the model supports 1M context, up to 384K output, tool calls, JSON output, the Responses and Anthropic-style APIs, and vision. It is the cheap, fast tier that coding agents hammer with thousands of small tool-calling requests. OpenCode's own estimate assumes about 71,300 cached input tokens per request, which tells you how cache-heavy agent traffic is and why the cached-read price ($0.003 per million off-peak) matters more than the headline input price.

We covered the economics in the DeepSeek V4 price increase post and the demand side in DeepSeek Flash hitting 8 trillion tokens in a day on OpenCode. For a head-to-head against other cheap models, see the small models comparison. Giving the model away is a customer-acquisition move for OpenCode and a way to gather feedback traffic for the model provider, a pattern OpenCode has used before with stealth models (see OpenCode's Omen Alpha mystery model).

What is not confirmed

Be careful with three claims that appear in reposts:

1. "Free forever" or any end date. OpenCode publishes none. A GitHub issue (anomalyco/opencode #43805) about the earlier DeepSeek V4 Flash Free model carries a commenter's report that its promotion ended and the model disappeared from the picker. That is one user's report about a previous model, not proof about this one, but it shows the pattern.
2. A 200K context cap. Secondary write-ups say the free endpoint has a 200,000-token context against 1M paid. OpenCode's Zen page does not state a cap for the free model. Test with your own long prompt before relying on it.
3. Identical model weights. The page calls the free entry a separate model ID. Whether its quantization, rate limits or queueing differ from the paid endpoint is not documented.

A checkbox next to an exit door illustrating the opt-out and data-retention choices for a free hosted model

Privacy: free does not mean unconditional

The Zen privacy section says OpenCode's providers generally follow a zero-retention policy and do not train on your data, then lists exceptions. Several of the free models, including Big Pickle, Exo Free, both MiMo free models and Muse Spark 1.3 Contributor Free, say collected data may be used to improve the model during the free period. Nemotron 3 Ultra Free is described as trial use only. DeepSeek V4.1 Flash Free is not on the exception list, and the page says its provider follows a zero-retention policy.

One more wrinkle: the Go privacy table marks DeepSeek V4.1 Flash with "0 days*" data retention, with an asterisk. We could not read the footnote in the fetched copy, so check it if retention matters to you. If you are handling proprietary code, read the table yourself and compare with the options in our private DeepSeek access via Venice post. Remember that DeepSeek is a China-based provider and that your own company policy may restrict sending code there regardless of what the hosting intermediary promises.

A gauge with a rising needle representing the monthly dollar usage limit on OpenCode Go

How to try it in five minutes

1. Install OpenCode and sign in to the Console to get an API key (Zen needs billing details on file even for free models, per the docs flow).
2. In the terminal UI run /connect, choose the OpenCode pay-as-you-go provider and paste the key.
3. Run /models and select DeepSeek V4.1 Flash Free, or set it in opencode.jsonc:

{
  "$schema": "https://opencode.ai/config.json",
  "model": "opencode/deepseek-v4.1-flash-free"
}

4. For direct API use, call https://opencode.ai/inference/v1/chat/completions with model deepseek-v4.1-flash-free and the @ai-sdk/openai-compatible package.
5. Keep a paid fallback: either the same model without the "Free" suffix on Zen, or a Go subscription, so a sudden end of the promo does not stall your agent mid-task.

Note that OpenCode asks clients on Go to send a stable x-opencode-session header so it can route and cache prompts; the Go page lists validated third-party clients such as Claude Code, Codex and Hermes.

What this means for what you pay

If you run an agent all day on a small model, the free tier is a real saving, but it is a trial, not a plan. The more durable number is the Go allowance: at the documented estimate of roughly 173,000 requests a month for $10, DeepSeek V4.1 Flash is among the cheapest ways to keep an agent loop running, and its off-peak pricing means scheduling heavy batch work on weekends or outside 01:00-04:00 and 06:00-10:00 UTC halves the cost.

Reliability is the other variable. DeepSeek's API went down for about five hours today, as we reported in the outage and 7-day credit refund post. A free endpoint carries no uptime commitment, so build model fallbacks into your harness: route to another cheap model when the first one errors. StepFun's free Step 5 preview, covered in our Step 5 Preview launch post, is a sensible second choice on the same provider.

If your coding agent runs unattended with shell access, remember that a free model is still a model that can take dangerous actions. AgentBeam, the agent security platform from the explainx.ai team, stops AI agents before they take dangerous actions, which matters more when cost stops being the brake on how much an agent does.

How to verify this yourself

• Open the Zen docs and search for "DeepSeek V4.1 Flash Free" in the pricing table and the Free models section.
• Compare with the Go docs usage limit table.
• Fetch https://opencode.ai/inference/v1/models to see which model IDs are live, as the docs suggest.
• Check DeepSeek's pricing page for the direct rates.

Related reading

• OpenCode: open-source AI coding agent guide
• DeepSeek V4.1 Flash API beta
• DeepSeek 5-hour outage and credit refund
• Small models compared: Haiku 5.5, GPT-6 Luna, DeepSeek V4.1 Flash and more
• DeepSeek Flash at 8 trillion tokens a day on OpenCode
• DeepSeek V4 price increase

Prices, limits and model lists are accurate as of October 15, 2026 and OpenCode changes them often. Re-check the official pages before you depend on them.