@meta/muse-spark-1.3 is live on Brainbase

Muse Spark 1.3 is now available on Brainbase as @meta/muse-spark-1.3. Use it with any supported harness and sandbox through the same Universal Managed Agents API you already use for every other model.

Meta released Muse Spark 1.3 on September 2, three weeks after 1.2, with a narrower brief than its predecessor: sustain longer-horizon work, juggle several workflows in a single thread, and get there in fewer turns. It keeps the one-million-token context window, native multimodal input, and the same per-token prices as 1.2.

#Why Muse Spark 1.3 matters for agents

The headline change is not a benchmark. Meta reports that 1.3 uses roughly 20% fewer tool calls and 25% fewer tokens than 1.2 to complete the same agentic tasks, and that it now asks clarifying questions instead of guessing when a task is underspecified. For an agent that runs for an hour, those two properties compound: fewer tool calls means fewer sandbox round-trips, and fewer tokens means the context stays useful for longer before compaction.

The second change is long context that actually holds up. Muse Spark 1.2 already exposed a million tokens, but its recall fell off sharply past 256K. On MRCR, 1.3 scores 98.5 in the 256K–512K band and 98.1 in the 512K–1M band, against 66.3 and 55.5 for 1.2 — and ahead of GPT-5.6 Sol on both. That is the difference between a context window you can advertise and one you can plan an agent around.

#Muse Spark 1.3 benchmarks

On Meta's published evaluation set, Muse Spark 1.3 ties GPT-5.6 Sol at 88.8 on Terminal-Bench 2.1 and edges Opus 5 on DeepSWE v1.1 and SWEAtlas CodeBase QnA. It sits between the two on GDPVal-AA v2 — 1754 Elo against Sol's 1710 and Opus 5's 1824 — and trails Sol on DeepSearchQA and the Agentic Instruction-Following Index. Select a metric below to compare the reported results.

Independent testing from Artificial Analysis scores Muse Spark 1.3 at 62 on its Intelligence Index, up eight points from 1.2 and level with the Opus 5 / Fable 5 tier. One caveat travels with that number: the model was among the more verbose in the evaluation, emitting about 120M output tokens against a median of 72M. Meta's 25%-fewer-tokens claim is relative to 1.2, not to the field.

#Muse Spark 1.3 pricing and context

Per-token pricing is unchanged from 1.2. Meta's framing is “frontier performance that's almost too cheap to meter,” and against the models it benchmarks itself on, that holds: input is a quarter of Opus 5's rate and output is a sixth.

Meta says its max reasoning modes for 1.3 follow after additional safety testing; the launch configuration is what ships on Brainbase today. These are model-token prices before sandbox runtime, tool calls, and other infrastructure usage.

#The Contributor tier

Meta also ships a second endpoint, muse-spark-1.3-contributor, at $0.10 input and $0.20 output per million tokens — roughly a twelfth of the standard rate. The difference is data: prompts and completions sent to the Contributor tier are used to improve Meta's products. The Brainbase identifier tracks the standardmuse-spark-1.3 endpoint, where they are not.

For internal tooling, evaluation runs, and workloads with no proprietary context, the Contributor pricing is hard to argue with. For production agents that see customer code or data, it is a decision to make deliberately rather than by default.

#Run Muse Spark 1.3 on Brainbase

Select the model in the agent spec and leave the rest of your stack unchanged. This example runs Muse Spark 1.3 inside the OpenCode harness on a managed sandbox:

curl https://api.brainbaselabs.com/v2/threads \
  -H "Authorization: Bearer $BRAINBASE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "agent": {
      "harness": "opencode",
      "model": "@meta/muse-spark-1.3"
    },
    "input": "Migrate the billing service off the deprecated client, keep the public API stable, and open a PR."
  }'

#Where Muse Spark 1.3 fits

Muse Spark 1.3 is the strongest case yet for a sub-frontier-priced model as a primary coding agent. It ties or beats GPT-5.6 Sol on terminal work, beats Opus 5 on long-horizon repository tasks, and is the only model in its comparison set whose million-token context holds recall past the halfway mark. Where it still trails is open-ended research and strict instruction following, where Sol keeps a clear lead.

Start with it where long sessions, large repositories, and cost all matter, then evaluate it against your production traces. Brainbase keeps the harness, sandbox, event stream, and observability layer constant so the model is the only variable.