Setting up a backup AI provider in case Opencode Go is down

TL;DR: Use NanoGPT (or something else) as a pay-as-you-go backup for OpenCode Go outages, with fallbacks from different model families than your primary.

Disclaimer: I AI-slopped this article but added edits manually.

OpenCode Go charges $10/month for $60 of usage.
I like it a lot.
BUT the service is being more and more used, it goes down sometime, forcing a massive timeout on you. Can't blame them, it's hard to find GPUs these days.

But I want to keep working.

The Alternative

NanoGPT runs on different providers. I load it with $15 minimum balance and only spend during OCG outages. Cost per outage day runs about $1.30 for my typical usage.

The Backup Config

This is my LLM Proxy config', the details are not important, but I decided that for NanoGPT my models would look like this.
This is a simplified configuration but you get the gist.

NANOGPT_MODELS='{
  "trash": { "id": "Meta-Llama-3-1-8B-Instruct-FP8" },
  "trash-fallback": { "id": "Meta-Llama-3-1-8B-Instruct-FP8" },
  "quick": { "id": "nex-agi/nex-n2-mini" },
  "quick-fallback": { "id": "gemma-4-e2b-it" },
  "low": { "id": "nex-agi/nex-n2-mini" },
  "low-fallback": { "id": "tencent/hy3" },
  "high": { "id": "upstage/solar-pro4:thinking" },
  "high-fallback": { "id": "tencent/hy3" },
  "deep": { "id": "zai-org/glm-5.2:thinking", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "high" } } },
  "deep-fallback": { "id": "qwen3.8-27b:thinking", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "high" } } },
  "max": { "id": "google/gemini-3.7-flash", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "high" } } },
  "max-fallback": { "id": "zai-org/glm-5.2:thinking", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "high" } } },
}'

For comparison this is my setup with Opencode GO

OPENCODE_GO_MODELS='{
  "trash": { "id": "hy3", "options": { "extra_body": { "thinking": { "type": "disabled" } } } },
  "trash-fallback": { "id": "mimo-v2.5", "options": { "extra_body": { "thinking": { "type": "disabled" } } } },
  "quick": { "id": "mimo-v2.5", "options": { "extra_body": { "thinking": { "type": "disabled" } } } },
  "quick-fallback": { "id": "hy3", "options": { "extra_body": { "thinking": { "type": "disabled" } } } },
  "low": { "id": "mimo-v2.5", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "low" } } },
  "low-fallback": { "id": "hy3", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "low" } } },
  "high": { "id": "minimax-m3", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "high" } } },
  "high-fallback": { "id": "qwen3.7-plus", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "high" } } },
  "deep": { "id": "minimax-m3", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "medium" } } },
  "deep-fallback": { "id": "qwen3.7-plus", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "medium" } } },
  "max": { "id": "glm-5.2", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "high" } } },
  "max-fallback": { "id": "qwen3.7-max", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "high" } } },
  "artistry": { "id": "mimo-v2.5", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "high" } } },
  "artistry-fallback": { "id": "mimo-v2.5", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "high" } } },
  "writing": { "id": "mimo-v2.5", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "high" } } },
  "writing-fallback": { "id": "minimax-m3", "options": { "extra_body": { "thinking": { "type": "enabled" }, "reasoning_effort": "high" } } }
}'

Why These Fallbacks

MiMo V2.5 Thinking (Xiaomi family) sits at 56.8 LMArena and costs $0.14/$0.28 per million tokens. That matches the capability of Qwen3.7 Plus which I use on OCG, but cheaper and from a different family.

Kimi K2.6 Thinking (Moonshot family) hits 61.8 LMArena at $0.50/$2.60. That's close to Qwen3.7 Max which I use as my OCG max fallback, with similar pricing tier and a different provider.

If MiniMax or Z.AI has a regional outage, the Xiaomi and Moonshot models still work.

Costs

It's not worth it to subscribe to NanoGPT's subscription for this use case.
It costs $12 per month, but with this setup here is what I expect to pay on pay-as-you-go.

Outage days/year	Backup cost
5	~$7
10	~$14
30	~$41

Here are other models I considered on NanoGPT for my setup,

A few alternatives worth keeping in mind for later:

• Ling 3.0 Flash Thinking ($0.08/$0.22, 50.6 LMArena, inclusionAI) - cheapest viable option. Useful if I ever need to cut backup costs further, though it drops 6 LMArena points below MiMo.
• Qwen3.8 27B Thinking ($0.20/$1.40, 68.1 LMArena, Alibaba) - higher LMArena than MiniMax M3, but Qwen family. Same as my OCG fallback, so it defeats the provider diversity point.
• MiMo V2.5 Pro Thinking ($0.44/$0.87, 60.2 LMArena, Xiaomi) - solid alternative for the max slot if Kimi K2.6 ever has issues.
• GLM 4.7 Flash Thinking ($0.07/$0.40, 45.3 LMArena) - very cheap but the capability drop is noticeable.
• DeepSeek V4 Pro Off-Peak ($0.66/$1.98) - decent price tier for max, but peak hours jump to $1.32/$3.96 and privacy guarantees are limited.

Cohere North Mini Code (36.5 LMArena) and Ring 2.6 1T (42.8) both scored too low to make the cut.

---

---