Muse Spark in OpenCode Go: Eligibility Before Price
MODELS & COSTS
Muse Spark in OpenCode Go: Eligibility Before Price
A low reseller rate is useful only if the provider permits your account to use the model. Muse Spark 1.2's contributor listing also carries data-use conditions. Regional access, the actual billed route and the intended repository's privacy needs belong in the same decision.
In this article
The country restrictions documented at the time excluded several markets, including the European regions, Canada, Vietnam and Argentina discussed below. A reseller listing does not establish eligibility for a particular account. Check the provider's country list and the actual account response before designing a workflow around it.
The relevant comparisons are capability, OpenCode integration and access terms. They answer different questions and should be checked separately.
Muse Spark's hosted model
Muse Spark is the first model from Meta Superintelligence Labs (MSL) — the new Meta division run by Scale AI founder Alexandr Wang. It's the opening act of Meta's Muse family, and it's a different animal from the open-weights Llama line: Muse Spark is closed-weights and hosted. You can't download it, you can't run it on your own GPU, you can only reach it through Meta's own API (or partners like OpenRouter and now OpenCode).
Capabilities and API access
- 1M-token context window — the "throw the whole repo at it" flagship feature.
- Truly multimodal input — text, image, video, audio, PDF. It reasons over screenshots and documents, not just code.
- Agentic-first design — parallel tool calls, sub-agent delegation, computer and browser use, built to run long multi-file coding tasks.
It's the same engine that already powers Meta AI — which launched in Korea back in May. The consumer app and the developer API are two very different gates, and that distinction is the whole plot of this post.
Benchmark sources and comparison limits
Short version: yes, it's competitive — and it climbed hard between 1.1 and 1.2. Here's the story in Meta's own official numbers.
Muse Spark 1.1 → 1.2 (Meta's official eval report)
| Benchmark | 1.1 | 1.2 | What it measures |
|---|---|---|---|
| DeepSWE | 53.3% | 59.3% | Real-world software engineering, long agentic tasks |
| Terminal-Bench 2.1 | 80.0% | +6.7 pts | Terminal / CLI agent work |
| Meta Internal Coding Bench | — | 70.6% | Meta's own coding eval |
| AC Bench | 48.4% | 54.8% | Agentic coding (3rd-party) |
On DeepSWE 1.1 — the benchmark devs now watch most closely — Muse Spark 1.2's 59.3% lands it third, right behind Opus 5 (65.0) and GPT-5.6 Terra (64.8).
And on Meta's own internal coding bench, 1.2's 70.6% actually beats GPT-5.6 Terra (65.4) and Gemini 3.6 Flash (63.9) — while sitting about nine points behind Claude Opus 5 (79.4). That's a remarkably candid admission from Meta: even on a test it designed itself, Anthropic still wins.
How to read this section: I'm using Meta's vendor-reported figures for the headline tables and flagging where third parties (Artificial Analysis, Vals) verify them. Treat any single-source number as "says Meta."
Contributor pricing
Meta split Muse Spark 1.2 into two tiers by model ID:
| Tier | Model ID | Input | Output | The catch |
|---|---|---|---|---|
| Contributor | muse-spark-1.2-contributor | $0.10/M | $0.20/M | Meta keeps your data for training · select countries only |
| Standard | muse-spark-1.2 | $1.25/M | $4.25/M | Data not retained for training |
The Contributor tier at ~$0.30/M total undercuts almost everything on the board — cheaper than DeepSeek V4 Flash and MiMo-V2.5, and orders of magnitude below Opus or Sol. That price is real, but it's priced that way because you're the product: your code and prompts become Meta's training data. If you're stringing together anything proprietary, that trade is a real cost, not a freebie.
OpenCode Go integration
Here's the part that got me excited before the wall hit back: OpenCode Go — the cheap $5/month-first subscription for open coding models — has added Muse Spark 1.2 Contributor to its lineup. Go's model list now reads:
Go is designed specifically "for international users," hosted across the US, EU, and Singapore for stable global access — which is why Korea and most of the world can use it for everything else. It's the confirmation that Meta sees agent-CLI distribution as the beachhead for Muse: an OpenAI-SDK-compatible, "bring your own harness" model baked straight into the tools people already run.
But here's the thing that trips everyone up — and it's the whole point of this post:
Go is a proxy — it resells the route. The actual inference still runs on Meta's own infrastructure behind Meta's own regional gate. If Meta's whitelist excludes your country, OpenCode Go can't do anything about it, because the model never reaches Go's servers — your request dies at Meta's door.
That's why you get "not available in your region" even though you're paying for a plan that lists the model. It's not a billing problem. It's not an OpenCode problem. It's a Meta jurisdiction problem.
Regional eligibility
Meta restricts the Muse Spark developer Model API with a three-layer gate, which is why workarounds mostly fail:
| Layer | What it checks | Can you beat it? |
|---|---|---|
| 1 · Geo IP | Your IP must resolve to an allowed country | VPN gets past this |
| 2 · Account region | Your Meta dev account's profile region must be allowed (effectively US for much of the rollout) | ← stops most people |
| 3 · Phone / SMS | Verify with a phone number from the allowed region | ← final wall |
A VPN gets you past layer 1, then layers 2 and 3 stop you cold. That's why the community reports are so consistent: EU users get the wall, Canadian users get the wall, and "my country is definitely not sanctioned" users in Vietnam and Argentina get the wall too. It's a region whitelist, not a sanctions filter.
The frustrating irony for Korean readers: Meta AI (the consumer chatbot) has been live in Korea since May 2026 — but that's the hosted app in your browser. The developer Model API is a separate availability decision, and Korea simply isn't in the whitelist yet. Same company, same model, two entirely different doors — one open to you, one locked.
Meta says it's "expanding global access" and that the Contributor tier is "available in select countries." Translated for normal people: they know it's restricted, they're working through a country list, and it's not you yet.
Deployment limits
- Closed weights. No local install, no self-hosting, no fine-tuning. This is the anti-Llama. Your GPU can't touch it.
- Region wall is the biggest blocker today, and it's a business/legal decision, not a technical one. Expect slow, uneven rollout.
- Contributor tier = data mining. Cheap by design. Read the retention terms before feeding it your codebase.
- Harness-inflated gains. Some of the 1.1→1.2 benchmark jump is the Muse Code agent, not pure model. Meta's own chart labels admit the 1.1 baseline ran in a different harness.
- Missing numbers. Meta hasn't published full 1.2 eval tables (no SWE-bench Verified at launch); the coding strength is real but partially self-reported.
The coding results and Contributor price make this a candidate for a controlled trial where access is supported. They do not resolve the tier's data-use conditions or guarantee the same price through every partner.
Before adopting it, confirm regional eligibility, data handling, rate limits and the partner's billed price. Contributor access should be evaluated on its own terms, especially for private repositories. Retain a fallback model for an account that loses access or reaches its limits.
Related on this blog: Meta just open-sourced its agent model: Muse Glimmer, a 30B that runs on your machine · Xiaomi AI Cube: 120B Local Models at 4.5× DGX Spark's Bandwidth
Comments
Post a Comment