DeepSeek V4.1 Flash has a temporary free route in OpenCode. Select opencode/deepseek-v4.1-flash-free. The official listing, checked October 17, 2026, describes the offer as limited-time. OpenCode Go remains a paid subscription.
For a first trial, the useful questions are straightforward: which entry should you select, what can become billable, and where will your code go? This guide checks the offer and setup against documentation and source. We have not run a task through this exact free route.
October 17 model update: DeepSeek separately released V4.1 Flash 1017 on its own website, app and API. OpenCode’s Free listing does not identify the dated checkpoint it serves, so this offer check does not establish access to that new revision.
Console security notice — sources captured October 19: OpenCode says a breach affected 1.4% of Console users, a subset with external providers linked, and says affected users were emailed. Read the credential-route distinction and verification limits before linking a provider. Direct local BYOK and Console-stored linked credentials are different paths; neither is declared proven safe here.
Choose the Free entry
Three similar selections have different billing relationships:
| OpenCode selection | Access route |
|---|---|
opencode/deepseek-v4.1-flash-free | Temporary free model |
opencode/deepseek-v4.1-flash | Metered Console model |
opencode-go/deepseek-v4.1-flash | Included usage within a paid Go plan |
Sources: live model catalog, Console pricing, Go configuration.
Go costs $10/month; Go Plus costs $40/month. Existing subscribers should select the separate Free entry to try this offer. The announcement does not turn a Go subscription into a free plan. Go’s own allowance and balance-fallback settings belong to that paid route.
Our DeepSeek V4.1 Flash guide covers the model and direct API. Use it for that background; direct-provider prices and earlier tests do not establish results for this OpenCode promotion.
Start without going through a paid signup
The released v1.19.3 client source includes access to free OpenCode models without a personal API key or connected account. When credentials are absent, it loads zero-input-cost models through its built-in public access path. This is a source check, not a completed fresh-install test.
For the official OpenCode client:
- Install or update it using the official instructions. The version inspected here was v1.19.3.
- Open a small project containing public or synthetic code.
- Enter
/modelsand select DeepSeek V4.1 Flash Free. - Confirm the full selection contains both
opencode/and the-freesuffix before submitting the task.
You can also choose the model when starting the client:
opencode --model opencode/deepseek-v4.1-flash-freeOpenCode documents both interactive selection and the model flag. If the entry is missing or a request is rejected, inspect the error and current listing. A payment prompt alone does not prove you must buy Go to use the free model.
If you already have paid providers configured, check the rest of the session too. OpenCode documents auxiliary title-model calls and configurable credit reload. Choosing a free primary model does not establish that every tool, subagent or background request is free. See the OpenCode permissions guide.
Scoped offer recheck, October 20: the Free entry remains listed as limited-time; its catalog still specifies 200,000 context / 32,000 output tokens. Go remains paid at $10/$40 monthly, and its DeepSeek privacy agreement still names October 31. This does not renew the October 17 released-client source inspection or establish fresh-install access, account quota, incident resolution or a Free-offer end date.
Limits: what is actually established?
The OpenCode-maintained model configuration lists 200,000 tokens of context and 32,000 output tokens, with input, output and cached-read prices set to zero. These are catalog settings; we did not test the server’s boundary or a maximum-length request.
We found no exact offer end date or fixed daily-request/concurrency quota in the reviewed free-model documentation. Rate limits, reliability and account-specific access remain unmeasured here. Use a bounded trial before depending on the route for time-sensitive work.
Go’s documentation separately dates its monthly DeepSeek zero-retention agreement through October 31, 2026. That date describes the Go privacy agreement; it does not establish when this free promotion ends.
Privacy: check this specific route
OpenCode explicitly says the Free model’s provider follows zero retention and does not use submitted data for training. Attribute that to OpenCode’s provider statement, rather than treating it as an independently audited result.
Its broader privacy policy distinguishes prompts passed upstream without storage by OpenCode from account, payment and device/IP information it collects. Request-content retention and account-data retention are different questions.
OpenCode’s hosted-service terms also allow content from unpaid accounts to help improve its services. The reviewed pages do not explain how that provision applies to this free route. The provider’s no-training statement is therefore not a blanket promise that nobody uses the content for another purpose.
A local coding client can still send selected code, prompts and tool output to a hosted service. Before connecting a private repository, identify which files and tools the client can access, which services receive the context, and whether session sharing is enabled. Our OpenCode privacy and sharing section explains those separate paths; its older pricing roster retains its own review date.
The data-retention explainer is useful for the distinction between training, storage and deletion. Its Claude/OpenAI policies are not the authority for this DeepSeek endpoint.
Where DeepSeek’s security history fits
AIHackers already covers DeepSeek’s 2025 database exposure and what Wiz confirmed. That historical infrastructure incident does not establish an exposure in today’s OpenCode service, or demonstrate that the current route is safe.
Read model capability, provider security, data terms and client permissions as separate evaluations. For a practical repository-access checklist, see the six questions from our ZCode investigation. Those questions transfer; the incident findings do not automatically transfer to OpenCode.
Make the first trial useful
Choose one small bug with a reproducible failure and a clear expected result. A suitable first prompt is:
Fix this one failing case in the supplied public or synthetic code. Explain the cause, make the smallest necessary change, and show how to verify the result. Ask before running commands or changing dependencies. Stop after one proposed fix and report anything you could not verify.
Review the proposed commands before executing them. A prompt expresses your request; configure actual permissions in the client as well.
Keep four observations: whether the exact route responded, whether the fix passed your check, how much review it needed, and what usage or cost information the service returned. Record unavailable measurements as unknown.
Our earlier $1 AI experiment used a different access route. Its outputs are not a test of this free model, and this offer check adds no new performance score. The cost-saving playbook explains how to compare useful work once you have results.
Before turning a trial into an automated public benchmark or output service, check the current service terms. They contain internal-use and output-extraction restrictions alongside an output-ownership clause; free pricing alone does not settle that use case.
For other available routes, use Free Stack; compare recurring paid options in the October buying guide. For this offer, start with the exact Free entry, a small inspectable task and the current provider terms.