Muse Wonder Reaches OpenCode: A Coding Agent Meets a 30-Second First Token

Meta Muse Wonder became selectable inside OpenCode on 21 October 2026, nine days after Meta Superintelligence Labs released the model on 12 October. That is the fact worth reporting, and it is a smaller fact than the headline suggests: the model did not launch here, it was added to a coding harness that already had a long list of models. What makes the addition interesting is the collision it sets up. A terminal agent is the least forgiving place to run a model that Artificial Analysis measures at 30.35 seconds to first token, and Muse Wonder is exactly that model — 52 on the independent Intelligence Index, $4.00 per million input tokens and $20.00 per million output, and a first-token wait roughly eight times the board median. Whether a coding agent can absorb that wait is the question this piece is actually about. When a first-token figure like this one changes after a provider ships a fix, the change is logged with its date and sources in the AI Incident Archive, next to the rest of the model and vendor failures we track.

What actually changed on 21 October

OpenCode publishes its model list in its documentation, and Muse Wonder now appears there under the identifier muse-wonder-1.0, priced at $4.00 input, $20.00 output and $0.20 cached read per million tokens — the same rate card Meta publishes, with no discount attached. The entry is served through an OpenAI-compatible Responses endpoint, which means an existing OpenCode configuration does not need a bespoke client to reach it; it needs a model string and a key.

The honest qualification: OpenCode's own changelog does not name the model. Its newest entries for that week, v2.1.0 on 20 October and v2.1.2 on 21 October, describe free-limit recovery, model visibility and SDK changes. So the availability is documented in the model list rather than announced as a headline feature, which is the ordinary way harness support lands and also the reason a reader may see the model in a picker before they see a blog post about it. We could not read the vendor's social post directly — it returned a payment wall — so the date here rests on the listing plus the post's own timestamp, not on a press release.

Nothing about the model itself changed on 21 October. Its weights did not open, its price did not move, and no new benchmark appeared. The only new thing is reach: a model that previously required wiring up Meta's API by hand is now one line of configuration inside a tool a lot of people already have open.

The number that decides whether this is pleasant

Artificial Analysis runs every model through the same harness, and for an agentic coding tool its latency figures matter more than its composite score. For Muse Wonder at its max reasoning setting:

• Time to first token — 30.35 seconds, against a board median of 3.96 seconds

• Output speed once it starts — 145.1 tokens per second, 26th of 185 measured models and well above the 79.2 median

• Cost per Intelligence Index task — about $4.52, against a blended list rate of $2.94 per million tokens

• Output tokens consumed on the index run — about 130 million, against an 85 million median

• Context window — 1,048,576 tokens, with text, image, speech and video accepted as input and text returned

Those two latency numbers describe a specific shape of model: it thinks for a long time, then writes quickly. In a chat window that is tolerable. In an agent loop that makes eight tool calls, the 30-second wait is paid per turn, and an eight-turn task can spend four minutes simply waiting for the first token of each step. The 145 tokens per second is real and it is why the model does not feel slow once it is talking — but it cannot buy back the silence before the first word.

That is not an argument against using it. It is an argument for choosing the right turns. Muse Wonder's strongest vendor-reported claim is long-horizon professional work — Meta's own table puts Muse Wonder Full at 75.3 on JobBench, a long-horizon agentic benchmark and the one place its numbers diverge most sharply from its sibling. If your OpenCode session is a single hard refactor with a large plan, the wait is amortised over a long generation. If it is forty small edits, it is paid forty times.

The repository that is not a release

Readers who go looking for the weights will find meta-01/muse-wonder on Hugging Face and may reasonably conclude something open happened. It did not. The repository contains a .gitattributes file and nothing else: no model card, no config.json, no safetensors shards, no licence file. Downloads are not tracked, and the page states that the model is not deployed by any inference provider on that platform. Read plainly, that repository is a name reservation — a placeholder holding a string, not a checkpoint.

So the correct statement of Muse Wonder's openness is: proprietary, API-only, with a bare repository squatting on the name. Anyone describing it as open-weight, downloadable, or fine-tunable is describing a different model. The same caution applies to the two-checkpoint structure — muse-wonder-1.0 is the callable model, and muse-wonder-full-1.0 is the higher-scoring configuration Meta sells through its Muse Code subscription rather than per token. OpenCode lists the first one.

Where OrcaRouter sits, and where it does not

We do not serve Muse Wonder. There is no meta/muse-wonder in the OrcaRouter model catalogue, and no amount of adjacent Meta availability changes that — the Meta checkpoints we do carry are meta/muse-spark-1.2 at $1.25 input and $4.25 output per million tokens, and its 1.1 predecessor, both on a 1,048,576-token window. Saying so plainly matters more than it usually would, because a model that is only reachable through two providers is exactly the case where a router's honest answer is "not this one yet" rather than a suggestive silence.

What the routing layer is genuinely good for here is the experiment. A ten-day-old model with a 30-second first token and only one independent evaluation behind it is not something to put a production path behind on faith. Point a slice of traffic at it from the same key that reaches the models you already trust, keep the rest of the traffic where it is, and let automatic failover catch the timeouts while you find out whether the long-horizon behaviour is worth the wait. That is a cheaper way to answer the question than migrating a workflow and hoping.

The other half of the routing argument is arithmetic. Because we pass provider list prices through with no markup, a vendor price cut shows up on our side the same day, and a model we do not host simply does not appear — the catalogue is a statement about what is reachable, not a wish list. When Muse Wonder lands here, the price you see will be Meta's price.

What would make this a bigger story

Three things, each checkable in a single page load. A Muse Wonder entry in OpenCode's changelog rather than only its model list would mean the integration is being treated as a feature rather than a catalogue row. A second independent evaluator publishing a long-horizon number would firm up the one claim that currently rests entirely on Meta's own table. And any movement in the Hugging Face repository — a model card, a licence, a single safetensors file — would change the openness question, which today has a one-word answer.

Until then the right reading of 21 October is narrow and useful: a frontier model got easier to try in the tool where most people would try it, and the thing that will decide whether it stays configured is not its index score but whether 30 seconds of silence per turn is a price the work can pay.