Models
Choose the Right Model
Chinese open-weight models deliver proprietary-tier coding at a fraction of the cost. Hereβs how to pick and access them.
Open Weights > Closed Gardens
The AI landscape has shifted. Chinese open-weight models now deliver 80β95% of proprietary model quality (Claude Opus, GPT-5) at 10β50x lower cost. For web development with AI agents, this changes the economics entirely.
| Open-Weight Chinese | Proprietary (US) | |
|---|---|---|
| Cost per 1M tokens | $0.14 β $0.87 | $3 β $15 |
| Context window | Up to 1M | 200K β 1M |
| Code quality (SWE-bench) | 70β76% | 72β78% |
| Weights available | Yes β open | No β closed |
| Self-host option | Yes | No |
| Censorship risk | Provider-level only | Provider + model-level |
The Models
| Model | Provider | Context | Max Output | Coding Rank | Best For |
|---|---|---|---|---|---|
| GLM-5.2 | Zhipu AI (z.ai) | 1M | 128K | #1 Code Arena | Long tasks, full-project context, heavy coding |
| DeepSeek V4 Pro | DeepSeek | 1M | 384K | Top-tier | Planning, architecture, complex reasoning |
| DeepSeek V4 Flash | DeepSeek | 1M | 384K | Strong | Fast iteration, editing, high-volume tasks |
| Kimi K2.7 Code | Moonshot AI | 128K | 32K | Strong | Precise code generation, refactoring |
| MiniMax M3 | MiniMax | 128K | 32K | Solid | Balanced coding, general tasks, cost-efficient |
| MiMo-V2.5-Pro | MiniMax | 128K | 32K | Strong | Frontend, design-heavy tasks, creative coding |
Sources: GLM-5.2 specs from docs.bigmodel.cn. DeepSeek V4 pricing from api-docs.deepseek.com. Code Arena rankings from bigmodel.cn (June 2026). OpenCode Go model list from opencode.ai/go.
GLM-5.2 Deep Dive
GLM-5.2 is the current standout for AI-powered web development. Key facts:
- 1M context window β fits entire project codebases in a single prompt. Tested at 850K+ tokens in real coding sessions.
- #1 on Code Arena β the global blind-test leaderboard for coding ability, ahead of all proprietary models as of mid-2026.
- Between Claude Opus 4.7 and 4.8 on SWE-bench verified, FrontierSWE, and PostTrainBench.
- Thinking mode with reasoning tokens for complex debugging and architecture decisions.
- Function calling + MCP support β works with OpenCodeβs tool system and PageWeave MCP tools.
- Open weight β can be self-hosted or run through API. No vendor lock-in.
DeepSeek V4
DeepSeek V4 comes in two tiers, both with 1M context:
| V4 Flash | V4 Pro | |
|---|---|---|
| Input (cache miss) | $0.14 / 1M tokens | $0.435 / 1M tokens |
| Output | $0.28 / 1M tokens | $0.87 / 1M tokens |
| Max output | 384K | 384K |
| Thinking mode | Yes (default) | Yes (default) |
| Tool calls | Yes | Yes |
| Concurrency | 2,500 | 500 |
How to Access These Models
Path 1: OpenCode Go (Recommended)
$10/month ($5 first month). One subscription gives you all six models above with generous rate limits. Open OpenCode Desktop, go to Settings β Providers, select OpenCode Go, sign in, and youβre done. Works with any agent that accepts OpenAI-compatible endpoints.
| Model | Requests per 5 hours |
|---|---|
| DeepSeek V4 Flash | 31,650 |
| Qwen 3.7 Plus | 4,300 |
| DeepSeek V4 Pro | 3,450 |
| MiMo-V2.5-Pro | 3,250 |
| MiniMax M3 | 3,200 |
| Kimi K2.7 Code | 1,150 |
| Qwen 3.7 Max | 950 |
| GLM-5.2 | 880 |
Get all models with one subscription
Subscribe to OpenCode Go β $10/mo$5 for your first month. Cancel anytime.
Path 2: Direct API (Pay-Per-Token)
More control, pay only for what you use. Each provider has its own API and pricing:
| Provider | Models | API Endpoint | OpenCode Setup |
|---|---|---|---|
| Zhipu AI (z.ai) | GLM-5.2, GLM-5.1 | open.bigmodel.cn |
Settings β Providers β Z.AI |
| DeepSeek | V4 Pro, V4 Flash | api.deepseek.com |
Settings β Providers β DeepSeek |
| Moonshot AI | Kimi K2.7 Code | platform.moonshot.ai |
Settings β Providers β Moonshot AI |
| MiniMax | M3, MiMo-V2.5-Pro | platform.minimax.io |
Settings β Providers β MiniMax |
All four providers are natively supported in OpenCode. Open the Desktop app, go to Settings β Providers, select the provider, sign in or paste your API key, and youβre done.
Path 3: Free Tier (Testing Only)
OpenCode includes free models for testing. Run /models in the TUI to see whatβs available. Free tier models have lower rate limits (~200 requests per 5 hours) and are suitable for exploration, not production work.
Recommendations by Task
| Task | Recommended Model | Why |
|---|---|---|
| Planning & architecture | DeepSeek V4 Pro | Best reasoning, 384K output, excellent at multi-step plans |
| Heavy coding sessions | GLM-5.2 | 1M context fits entire projects, #1 coding rank |
| Full-project context | GLM-5.2 or DeepSeek V4 Pro | Both have 1M context, handle 850K+ token sessions |
| Fast editing / iteration | DeepSeek V4 Flash | Fastest, cheapest, 31K+ requests on Go plan |
| Frontend & design work | MiMo-V2.5-Pro | Strong at creative coding, visual output |
| Budget-conscious heavy use | MiniMax M3 | 3,200 req/5hr on Go, solid across all tasks |
Model freshness note: The Chinese open-weight model landscape evolves weekly. GLM-5.2 launched June 2026. DeepSeek V4 launched mid-2026. Check opencode.ai/go for the current model lineup, which is updated as new models ship.
Our recommendation: Start with OpenCode Go ($10/mo). You get all six models with generous limits. Use GLM-5.2 for heavy sessions, DeepSeek V4 Flash for quick edits, and DeepSeek V4 Pro for planning. One subscription, no per-token anxiety.
Continue