OpenCode Zen: pricing, free models and the Console breach

This page is built from OpenCode's own Zen landing page, docs, status page and public model list, the makers' price pages, GitHub issues, Trustpilot, Hacker News, a Kingy AI review, a competitor's comparison and news reports on the October 2026 breach, all checked on October 21, 2026. OpenCode's claims are labelled as its own.

What OpenCode Zen sells

OpenCode Zen is the paid model service of Anomaly Innovations, the former SST team that makes the open-source OpenCode coding agent. The landing page calls it a curated set of models "tested and benchmarked for coding agents". Since OpenCode 2.0 on October 14, 2026, the docs call it the Console, alongside workspaces, budgets, provider connections and the OpenCode Go subscriptions (10 and 40 USD a month).

The price table lists about 70 paid and 14 free models: Claude, GPT, Gemini, Grok, Kimi, GLM, DeepSeek, Qwen, MiniMax and Mistral. OpenCode says it tested each model, worked with providers to serve it correctly and benchmarked each model-provider pair, but it does not publish those results or let you choose the provider. Payment is by card, Link, Alipay, UPI or WeChat Pay through Stripe (billing docs).

Prices against the official APIs

OpenCode says Zen charges per request "with zero markups" and adds a processing fee of 4.6% + 0.31 USD to each credit purchase: 0.77 USD on 10 USD and 1.23 USD on 20 USD. Most prices match the makers', with exceptions:

Gemini 3.8 Flash on Zen is already at the price Google will charge from January 2027. GPT-6.1 Sol cache reads are also 0.20 USD against OpenAI's 0.10 USD. Refunds are possible within 14 days if the credits from that charge are unused.

The October 2026 Console breach

On October 17, 2026 OpenCode said an intruder had reached "a legacy system" the night before and that third-party provider keys saved in Console workspaces may have been exposed, affecting 1.4% of Console users, all with an outside provider linked (Infosecurity Magazine). Co-founder Dax Raad said the breach involved "the OpenCode Console, not OpenCode itself", and that keys stored by the CLI and desktop apps stay on the user's machine. OpenCode linked it to a large-scale distillation attack and told affected users to revoke and replace their keys. The Hacker News reported users' own claims of unauthorised charges, such as about 281 USD on OpenAI and 168 USD on Anthropic for one user and 1,700 USD for another. By October 19 OpenCode had not said how the system was reached or whether credits, payment data or prompts were involved.

What public tests show

No independent latency, uptime or quality test of Zen has been published. OpenCode's own status page shows live time to first token and throughput per product; on October 21, 2026 pay-as-you-go read about 1.7 to 2.4 seconds to first token, while free inference was degraded by an upstream provider issue. A Kingy AI review (August 2026, updated October 2026) did not run paid inference through Zen. Kilo, a competitor, compares Zen's no-markup pricing and 1.23 USD fee with its own gateway.

What users report

GitHub issues describe paid top-ups that never reached the balance (#51562, #52186, September 2026), a "Free usage exceeded" error on a paid balance (#33318), Go checkouts charged to the Zen balance (#29182), and three Claude Fable 5 calls cut off by a content filter but billed about 20 USD in cache writes (#35475). Billing support is by email only, and many issues were closed by bots or when OpenCode 1.x issues were closed after 2.0. Trustpilot shows 2.6 out of 5 from 16 reviews, mostly about Go limits, bans and refunds. In the OpenCode 2.0 thread on Hacker News, commenters split on whether Go is the best deal or the worst open-model subscription.

MCP, Agent Skills and coding agents

• MCP: OpenCode documents a Console MCP server at https://opencode.ai/console/mcp, using OAuth with dynamic client registration, so harnesses with MCP support can manage your workspace. The Claude Code and Codex guides show one-line setup; the docs do not list its tools. The inference API has no MCP features. OpenCode itself is an MCP client.
• Agent Skills: no official skills appear in its docs or the anomalyco GitHub organisation. The OpenCode agent loads SKILL.md skills; Anthropic's Skills API is not documented for the Anthropic endpoint.
• Coding agents: Claude Code reaches Claude and most Qwen models, Codex reaches GPT, Grok and Muse Spark, and Hermes, Pi and Raycast reach every model. Kilo Gateway accepts an OpenCode Go subscription as a bring-your-own-key option.
• API formats: OpenAI Chat Completions, OpenAI Responses, Anthropic Messages and Gemini generateContent (inference docs), with a session header for consistent routing and prompt caching, and hosted web search at 0.01 USD per search. The docs map each model to an AI SDK package, and an llms.txt indexes the v2 docs.

Privacy, terms and who runs it

OpenCode says its providers keep no data and do not train on it, except listed free models; OpenAI and Anthropic keep requests 30 days under their own policies. The landing page says models are hosted in the US, the docs say the US and EU. Workspaces can connect their own Anthropic, OpenAI, Bedrock, Vertex, Azure or OpenRouter accounts through the Console (providers docs); those saved keys were what the October breach may have exposed. Budgets cap monthly spend per workspace and member (budgets docs).

How it compares with OpenRouter, Kilo Gateway and Requesty

• OpenRouter lists far more models, lets you pick hosts and adds 5.5% to purchases; its free models allow 50 requests a day.
• Kilo Gateway adds 5% to purchases and offers only a Chat Completions endpoint.
• Requesty adds 5% to top-ups and has an EU endpoint.

Who should skip it

Skip OpenCode Zen if you need responsive billing support, image or video models, or control over which provider serves a model. Until OpenCode explains the October breach, avoid saving your own provider keys in the Console.