Disclosure: some links on this page may be affiliate links; we may earn a commission at no extra cost to you. Read the disclosure.

By the toolist.ai editors. Researched from OpenCode's own pages and hands-on tests published by others, not hands-on tested by toolist.ai. Updated October 23, 2026.

OpenCode review: pricing, features and what testers found

OpenCode is a free, open-source (MIT) AI coding agent from Anomaly for the terminal, a desktop app, the web and code editors, working with 75+ model providers, a ChatGPT or Copilot plan, or local models. Anomaly sells models separately: Go costs $10 a month, Go Plus $40, and Zen is pay-as-you-go.

Our take

OpenCode suits developers who want a Claude Code-style coding agent that is free, open source (MIT) and not tied to one model company: it runs in the terminal, a desktop app, a web interface and editors such as VS Code and Zed, and works with 75+ providers, a ChatGPT Plus or Pro or GitHub Copilot sign-in, local models or a few free models. Version 2.0, which Anomaly announced on October 12, 2026 after a public beta from July, runs one background service that the terminal, desktop and web apps share, with background subagents and a Code Mode for MCP tools. Anomaly sells the models separately: Go costs $10 a month and Go Plus $40 for 29 curated models under a monthly dollar limit per model (on Go, Kimi K3's $15 works out to about 110 requests per 5 hours by OpenCode's estimate), and Zen offers 88 models (13 of them free for a limited time) pay-as-you-go at prices OpenCode says carry no markup, plus a card fee on each top-up. XDA's Mahnoor Faisal found in March 2026 that it has the read-only Plan mode she likes in Claude Code, and by September used it with a ChatGPT Plus plan, API keys and local models in place of several subscriptions; and MakeUseOf's Raghav Sethi, whose article carries a sponsored link to OpenCode, ran routine work on Go's models up to May 2026 but kept Claude for hard reasoning. The catches: Claude Pro and Max subscriptions cannot be used (Anthropic forbids it, though OpenCode's home page still lists them), Go's dollar limits go quickly on pricier models, the default Build agent edits files and runs commands without asking, version 2 dropped the language-server (LSP) support two testers credited and does not run version 1 plugins, and every cited test predates version 2.0 and Go Plus. This verdict rests on those three published tests and OpenCode's own pages.

Based on OpenCode's published plans and the hands-on tests linked below, not on a test of our own.

Good for

  • Developers who want a free, open-source alternative to Claude Code or Codex that is not tied to one model company, keeping one interface while switching models by task, as XDA's Mahnoor Faisal did
  • People who already pay for ChatGPT Plus or Pro or GitHub Copilot, or hold API keys, and want one agent across them, with local models through Ollama or LM Studio for private or free work
  • Trying agentic coding at no cost on OpenCode's limited-time free models, or for $10 a month on Go's 29 curated models, with extra usage paid from Zen credits only if you turn it on
  • Developers who want to drive the agent from scripts, other apps or their own products through its local service, API, SDK and plugins, or carry one session across the terminal, desktop and web apps (version 2)
  • Teams that want an open-source agent with free shared workspaces, budgets, SSO and SCIM and their own provider keys, or an Enterprise setup on their internal AI gateway

Watch out for

  • You cannot use a Claude Pro or Max subscription: OpenCode's provider docs say Anthropic explicitly prohibits the plugins that allowed it, version 1.3.0 (March 22, 2026) removed the one OpenCode bundled, and The New Stack dates Anthropic's block on third-party agents to January 2026. OpenCode's home page FAQ still lists Claude Pro and Max among the subscriptions you can use. Claude in OpenCode now means an Anthropic API key or Zen credits (Claude Opus 5.5 at $4 per million input tokens and $20 per million output tokens on Zen), and Go includes only Claude Haiku 5.5.
  • Go and Go Plus are not unlimited: each model has a monthly dollar limit counted at its token price, of which 20% can be used per 5 hours and 50% per week. OpenCode's own estimates for Go run from about 110 Kimi K3 requests per 5 hours (490 a month) to 45,300 for the Muse Spark Contributor models, based on token counts per request that it says vary in practice. Several Trustpilot reviewers say their limits ran out fast; one says a single unfinished GLM 5.3 task used 101.3% of the 5-hour, 40.5% of the weekly and 20.3% of the monthly allowance. OpenCode says the limits may change, the model list changes as it tests new ones, and some terms are temporary (Gemini 4 Argon Contributor's doubled limit and two free models are 'limited time').
  • Go and Go Plus bill monthly only (no yearly price), renew automatically until you cancel, and a refund is possible only within 14 days of the charge and only if you have not used that period's allowance (Zen credits: within 14 days, if unused). Zen top-ups add a processing fee of 4.6% plus $0.31 ($1.23 on $20), and the terms make you responsible for taxes and send disputes to binding arbitration in Delaware unless you opt out by mail within 30 days. The '$5 first month' MakeUseOf's tester describes was on the Go page until at least August 23, 2026 and gone by August 24, and the $5 referral credit ended when referrals stopped being recorded on September 22, 2026.
  • OpenCode says it stores no code or context except when you use its free models, and several of those (Big Pickle, Flint 40B Free, Exo Free and the free MiMo models) may use what you send to improve the model, while NVIDIA's free Nemotron endpoints log use and warn against personal or confidential data; setting OPENCODE_DISABLE_FREE_MODELS=1 hides them. On Go, Gemini 4 Argon Contributor and the Muse Spark Contributor models may use your prompts and outputs to train Google's or Meta's models, OpenAI, Anthropic and xAI models keep requests for 30 days, and DeepSeek's zero-retention agreement is renewed monthly (current through October 31, 2026). The Zen page says its models are hosted in the US, the Console docs in the US and EU.
  • By default the Build agent edits files and runs shell commands without asking; in version 2 it also reads and edits outside the project without a prompt (version 1 asked for outside directories), and only reading .env files asks first. Add ask or deny rules if you want approvals, as OpenCode's permission docs show for shell commands and git push.
  • Version 2 is new and still moving: Anomaly announced it on October 12, 2026, version 1 plugins do not run in it, its server API is different, and 2.1 (October 20) changed SDK and plugin contracts again. OpenCode's V2 migration guide says version 2 does not run language servers or produce LSP diagnostics, which the home page still advertises and which MakeUseOf's tester and XDA's September tester credited before 2.0. Session sharing (/share links), which XDA's March tester liked and the home page still lists, was discontinued in both versions on October 15, 2026 (1.19.1 and 2.0.30). Version 1 still gets releases (1.19.4 on October 22, 2026), the two are no longer installed side by side by default, and version 1 has stopped installing updates automatically since 1.19.0 (October 11).
  • All three cited tests predate version 2.0 and Go Plus, and none covers Zen, the desktop app or the Console. The two XDA pieces are one writer's personal accounts, neither naming a specific model; the only test of Go, MakeUseOf's (published May 17, 2026, with a sponsored link to OpenCode), used MiniMax M2.7 and GLM-5, neither of which is in Go's line-up today. We found no independent hands-on test of version 2.0 or Go Plus; The New Stack's October 12, 2026 article on 2.0 reports Anomaly's announcement rather than a test.

What OpenCode does

Price

Free plan: OpenCode itself is free and open source under the MIT License: the terminal app, the desktop app for macOS, Windows and Linux, the web interface and the editor extensions (VS Code, Cursor, Zed, Windsurf, VSCodium) cost nothing. It comes with a rotating set of limited-time free models (13 in OpenCode's docs on October 23, 2026) that need no account, some of which may use what you send to improve the model. Otherwise the AI is paid for separately: your own API key from 75+ providers, a ChatGPT Plus or Pro or GitHub Copilot sign-in, local models through Ollama or LM Studio at no charge, or OpenCode's Zen credits or Go subscription. A Claude Pro or Max subscription cannot be used (Anthropic prohibits it), so Claude needs an API key or Zen credits.

PlanBilled monthlyBilled yearly, per monthIncludes
Open source (free) free not listed The MIT-licensed agent in the terminal, the desktop app, the web interface and editor extensions; limited-time free models with no account needed (some may use your data to improve the model); your own API keys from 75+ providers, ChatGPT Plus or Pro and GitHub Copilot sign-ins, or local models, each billed by its provider; no Claude Pro or Max sign-in
Zen (pay-as-you-go) not listed not listed Prepaid credits in the OpenCode Console for 88 models in OpenCode's docs (13 of them free for a limited time), among them Claude, GPT, Gemini, Grok, Muse, Qwen, Kimi, GLM, DeepSeek, MiniMax and Mistral models, charged per request at per-token prices OpenCode says carry no markup, from $0.05 input and $0.40 output per million tokens (GPT 5 Nano; Jev 1.13, a decision model that does not write text, charges only $0.042 per million input tokens) to $30 and $180 (GPT 5.5 Pro and GPT 5.4 Pro); the Zen page offers a $20 top-up plus a $1.23 card processing fee, while the billing docs offer $10, $50, $100 or any amount plus 4.6% and $0.31; optional auto top-up and monthly spend limits; requests stop when the balance reaches zero; usable from other agents through an API key
Go $10 not listed Monthly subscription to 29 curated models that OpenCode's docs call open coding models (Kimi K3, GLM-5.3, DeepSeek V4 Pro, Qwen3.8 Max, MiniMax M3, MiMo and others), plus OpenAI's GPT 6 Luna and GPT 5.6 Luna, Anthropic's Claude Haiku 5.5, xAI's Grok 4.7, Google's Gemini 4 Argon Contributor and Meta's Muse Spark Contributor models; each model has a monthly dollar limit counted at its token price ($15 for Kimi K3 up to $80 for DeepSeek V4.1 Flash), 20% of it per 5 hours and 50% per week; optional extra usage paid from Zen credits; works in OpenCode or other agents through an API key; cancel any time; refund within 14 days only if the allowance is unused
Go Plus $40 not listed The same models and token prices as Go with higher monthly limits: for the 12 paid models the Go page charts, 2 to 4 times Go's (Kimi K3 $60 instead of $15, MiniMax M3 $180 instead of $60, DeepSeek V4.1 Flash $160 instead of $80), with the same 5-hour and weekly shares; monthly, cancel any time; on the Go page since September 28, 2026
Enterprise not listed not listed Contact sales, per-seat pricing: a central config tied to your SSO and internal AI gateway that can disable all other providers, no token charges when you use your own gateway, OpenCode's statement that it stores no code or context data, and a SOC 2 Type 2 report in its Trust Center. The Console's team features (workspaces, budgets, policies, SSO and SCIM) need no Enterprise contract and have no seat fees

Prices last checked: 2026-10-30 on OpenCode's pricing page. Full breakdown: OpenCode pricing.

Price history

No price change since we first recorded these prices on October 23, 2026. We re-check them every week and list each change here.

What hands-on testers found

We have not tested OpenCode ourselves. These are the published tests we found that used it hands-on, newest first; follow the links for the full reviews.

XDA · Mahnoor Faisal · September 27, 2026

What they tested: Her working setup as of September 2026, before Anomaly announced version 2.0: a ChatGPT Plus subscription signed in inside OpenCode, Claude and other occasional models through API keys, and local models for lighter tasks, in place of separate subscriptions for several coding tools. She names no version or task and does not mention Go or Zen; its OpenCode screenshots repeat those of her May 2, 2026 XDA piece, and the article does not link to OpenCode.

  • Signing in with her ChatGPT Plus plan let her use the OpenAI models she already paid for inside OpenCode, but because Anthropic does not allow Claude Pro or Max there, she reaches Claude through Anthropic's API instead.
  • Keeping one interface and swapping the model underneath meant she no longer paid a monthly plan for each coding tool, paying API usage only for the models she uses now and then and nothing for local ones.
  • Beyond model choice she found it a full coding agent: it works across a project, edits files, runs terminal commands, uses LSP integration to understand the code, and offers Plan and Build modes, per-action permissions for edits and shell commands, and MCP servers.

MakeUseOf · Raghav Sethi · May 17, 2026

What they tested: OpenCode on the Go plan up to May 2026 (version 1), which he says then cost $5 for the first month and $10 a month after: MiniMax M2.7 for most work and GLM-5 for harder problems, with Ollama for trivial tasks and an API key for serious ones, on a handful of personal projects including a personal finance dashboard and a script that sorts files in a watched folder. Neither model is in Go's line-up today, and the test predates version 2.0 and Go Plus. Tied: the article carries a sponsored link to OpenCode.

  • MiniMax M2.7 through Go handled his TypeScript and Python work well for what it cost, but on multi-file refactoring, architectural choices and other work that needs deep reasoning, Go's open models fell noticeably short of Claude Sonnet or Opus in his use, so he kept Claude for those.
  • Putting routine work (functions, features from a spec, tests, simple fixes) on OpenCode with Go and only the hard problems on Claude cost him less than Claude Pro alone had, and he built his projects for well under the price of one month of Claude Pro.
  • He credits its LSP integration (the agent sees the same diagnostics as your editor) and separate sessions per task, and repeats the OpenCode team's warning that some MCP servers, GitHub's in particular, fill the context quickly.

XDA · Mahnoor Faisal · March 19, 2026

What they tested: OpenCode in March 2026, on version 1 before 1.3.0 (March 22, 2026) removed the bundled Claude Pro or Max sign-in, as a non-developer who also uses Claude Code; her screenshots show it planning, then building, a Pac-Man game. She does not name the model she used or mention Go or Zen, and the test predates version 2.0.

  • She found it handles the same jobs as Claude Code from the terminal (explaining code, adding features, fixing bugs, running commands and changing several files at once) for free, with a choice of over 75 providers rather than Anthropic's models alone.
  • Plan mode, one of her favorite Claude Code features as a non-developer, has a counterpart: Tab switches OpenCode into a read-only Plan mode to work out the approach, and back to Build mode to carry it out.
  • She liked referencing project files in a prompt with @, dropping images such as mockups into the terminal, /undo and /redo, /share links to a whole conversation, and themes and custom commands; at the time it also offered a Claude Pro or Max sign-in, which she noted Anthropic did not officially support.

What users say

Test result

toolist.ai has not run its own test of OpenCode; the findings above come from the published tests linked. See how we research and test.

Best coding tools · OpenCode pricing

Do you make OpenCode? Claim this listing to correct it; it's free. To see whether ChatGPT, Claude, Gemini and Perplexity recommend it, and how many readers we send you: toolist.ai for AI companies.

Sources