Large Language Models

Muse Spark 1.2

Meta's coding-focused flagship — co-trained with Muse Code terminal agent, TB 2.1 82.9%, $1.25/$4.25 standard pricing, 1M context, open weights announced.

Muse Spark 1.2 is Meta Superintelligence Labs’ August 5, 2026 update — a coding-focused model launched alongside Muse Code, Meta’s first terminal coding agent. It’s the third Muse release in four months, and the pitch this time is clear: compete on price, not peak performance.

On Meta’s own charts, Spark 1.2 trails Claude Opus 5 on every coding benchmark published. But at $1.25/$4.25 per million tokens (or a jaw-dropping $0.10/$0.20 on the contributor tier), it undercuts most frontier models on cost per task — and that’s where this release actually matters.

Architecture

SpecValue
Model familyMuse (Meta Superintelligence Labs)
PredecessorMuse Spark 1.1 (July 9, 2026)
Context window1M tokens
Input modalitiesText, image
Output modalitiesText
API model IDsmuse-spark-1.2 (standard) / muse-spark-1.2-contributor

Pricing

TierInput / 1MCached / 1MOutput / 1MRPM cap
Standard$1.25$0.15$4.253,000
Contributor$0.10$0.002$0.2060

The contributor tier trades dramatically lower pricing for permission to train future Meta models on your prompts and completions. Web search grounding adds $2.50 per 1,000 queries on top of token costs.

Benchmarks

Meta published its own head-to-head vs Opus 5 and Codex (GPT-5.6 Terra). The pattern: competitive, but second.

BenchmarkMuse Spark 1.2Claude Opus 5Codex (GPT-5.6 Terra)
Terminal-Bench 2.182.9%86.7%81.8%
DeepSWE 1.159.3%65.0%64.8%
Meta Internal Coding70.6%79.4%Not published

Artificial Analysis scored Muse Spark 1.2 at 54 on the Intelligence Index (xhigh effort) — up 3 points from Spark 1.1 (51), effectively tied with GPT-5.5 (55) and Grok 4.5 (54). At $0.40 per task, it’s near the Pareto frontier of intelligence vs cost.

Muse Code

The real product story. Muse Code is a terminal coding agent (macOS and Linux) co-trained with Spark 1.2 — not a generic wrapper, but a model-agent pair optimized together. Key features:

  • Multi-step coding: writes code, runs it, reads results, iterates
  • Sub-agent parallelism for complex tasks
  • Checkpoint logging — resumes after crashes instead of restarting
  • Installed via a single command; bills per token through Meta Model API

How it compares

  • vs Claude Opus 5: Opus 5 wins coding benchmarks by 4–9 points. But Spark 1.2 costs roughly $0.40/task vs Opus 5’s higher per-task cost — for many workflows, the price gap matters more than the score gap.
  • vs GPT-5.6 Sol: Sol scores higher (AA Index 59 vs 54) at $5/$30 — 4× more expensive on input, 7× on output.
  • vs Grok 4.5: Tied on AA Index (54 each). Grok 4.5 at $2/$6 is somewhat pricier; Spark 1.2’s contributor tier is 20× cheaper.
  • vs Inkling: Both Apache 2.0 positioned (Inkling already, Spark soon). Inkling at $1.87/$4.68 is priced similarly to Spark standard; Spark’s contributor tier undercuts heavily.
  • vs Sakana Fugu: Fugu orchestrates multiple models (including Spark). Different abstraction level — Fugu is a system, Spark is a model.

Open weights

On August 10, Meta announced plans to release Muse Spark 1.2 open weights “in the coming weeks.” No date, license, or checkpoint format has been confirmed. If Meta delivers, this becomes the most capable downloadable foundation model from a US lab — a significant strategic move after months of proprietary Muse releases.

Limitations

  • Not the coding ceiling. Opus 5 leads by 4–9 points on every coding benchmark Meta tested.
  • Contributor tier fine print. The $0.10/$0.20 pricing requires sharing all prompts and completions for training. Enterprise users with data sensitivity constraints should stay on standard.
  • Muse Code is beta. No desktop app. macOS/Linux only. Billing is per-token with no subscription option.
  • Open weights timeline is vague. “Coming weeks” is not a date.

For the full AI model landscape, see the 2026 LLM guide.

Best for

  • Terminal-based coding agents via Muse Code
  • Cost-effective frontier inference for agentic workflows
  • Developer tooling integration (planning, multi-agent orchestration)
  • Fine-tuning base once open weights ship

Pros & cons

Strengths

  • Co-trained with Muse Code — model and agent harness optimized together, not bolted on
  • TB 2.1 82.9% beats Codex/GPT-5.6 Terra (81.8%) at roughly half the cost per task
  • AA Intelligence Index 54 — ties GPT-5.5 and Grok 4.5, costs less per task than both
  • Contributor tier ($0.10/$0.20) is the cheapest frontier-class inference available — if you accept the training-data trade

Limitations

  • TB 2.1 82.9% trails Claude Opus 5 (86.7%) — not the coding ceiling
  • DeepSWE 59.3% trails Opus 5 (65.0%) and Codex (64.8%) — gap widens on harder repo-level tasks
  • Meta's own internal benchmark shows a 9-point gap vs Opus 5 (70.6% vs 79.4%)
  • Contributor tier shares your prompts and completions with Meta for training — read the fine print

Frequently asked questions

When was Muse Spark 1.2 released?

August 5, 2026, alongside Muse Code, Meta's first terminal coding agent. It's the third Muse model in four months (Muse Spark 1.0 in April, 1.1 in July).

How does Muse Spark 1.2 compare to Claude Opus 5?

Opus 5 wins on every coding benchmark Meta published: 86.7% vs 82.9% on TB 2.1, 65.0% vs 59.3% on DeepSWE, 79.4% vs 70.6% on Meta's internal coding bench. But Spark 1.2 costs $0.40 per AA task vs higher costs for Opus 5 — this is a price story, not a capability story.

What is Muse Code?

A terminal coding agent (macOS/Linux) co-trained with Spark 1.2. It handles multi-step coding tasks, runs sub-agents in parallel, and maintains a checkpoint log so it can resume after crashes. Available via Meta Model API credentials.

What is the contributor tier?

muse-spark-1.2-contributor: $0.10 input / $0.20 output (12.5× and 21.25× cheaper than standard). In exchange, Meta can use your prompts and completions to train future models. 60 RPM cap vs 3,000 RPM on standard.

Will Muse Spark 1.2 be open weight?

Meta announced on August 10 that open weights will be released 'in the coming weeks.' No exact date, license, or checkpoint details have been published yet.

Sources

Last updated: 2026-08-11 · Specs and pricing change fast — verify on the vendor's site before relying on them.