
Muse Spark 1.3: Specs, Price, and When to Switch
By Alex Morgan
MyClaw Editorial
MyClaw
Run Best-in-Class AI Agents Now
Run OpenClaw or Hermes Agent on managed MyClaw hosting, so your AI agent stays online, updated, and ready for real work.
AI Takeaway
- What is Muse Spark 1.3? Meta's current Muse Spark checkpoint for agentic and coding work. API ID
muse-spark-1.3. Shipped 2 September 2026 in Muse Code and Meta Model API. Official docs: recommended for new work. - What does it cost? Standard: $1.25 / $4.25 per 1M tokens (cached $0.15), same list as 1.2. Contributor
muse-spark-1.3-contributor: $0.10 / $0.20 (cached $0.002) if Meta may train on your prompts. - Is it better than 1.2? At matched
xhigh, Terminal-Bench 2.1 is 89.2 vs 82.9, with a large long-context lift. JobBench is a wash. DeepSWE 75.4 is 1.3 max only. Audio is a step back. - Should you switch? Yes for new Muse Spark work. Keep 1.2 for audio and jobs that already pass. Pin
reasoning_effortbefore you compare bills.
Muse Spark 1.3 at a Glance
Muse Spark 1.3 is the current Spark workhorse, not a cheaper SKU and not a 1.2 rename. Official Meta Model API pages, checked 3 September 2026:
| Detail | Muse Spark 1.3 |
|---|---|
| Release date | 2 September 2026 |
| API model IDs | muse-spark-1.3 (Standard), muse-spark-1.3-contributor |
| Context window | 1,048,576 tokens |
| Max output (docs example) | 131,072 tokens |
| Input | Text, image, video, audio*, PDF |
| Output | Text |
| Reasoning | minimal, low, medium, high, xhigh (none returns HTTP 400) |
| Standard price | $1.25 / $4.25 per 1M; cached input $0.15 |
| Contributor price | $0.10 / $0.20 per 1M; cached input $0.002 |
| Extra | Web search grounding $2.50 per 1,000 queries; no long-context premium |
*Audio understanding is not fully supported; quality may be degraded. Docs default to muse-spark-1.3 and reasoning_effort: "high". Reasoning tokens bill as output.
The context window is not new. 1.1 and 1.2 already held about a million tokens at the same Standard rate. The upgrade is what 1.3 does inside that window.
On MyClaw there is no 1.3 model page. The live listing is still the original Muse Spark family card. Run the IDs the provider actually exposes, and write down any substitution.
What Changed From Muse Spark 1.2

Same Standard sticker. Efficiency is the claim.
Meta kept the 1.2 Standard rate. 1.3 is the next Spark checkpoint, not a discount. If both finish the same job in the same tokens, there is no price reason to stay on 1.2. Meta is not promising the same token count. Launch materials claim fewer unneeded turns, less verbose code, and, in Meta engineer comparisons, about 20% fewer tool calls and 25% fewer tokens. That comparison is vendor-internal, not an independent rerun.
A cheaper call is still expensive if it retries. Count:
(model spend + retries + review time + repair time) ÷ accepted tasks
Coding and long-horizon agents
Official docs position Muse Spark 1.3 for multi-step tool, browser, and long-horizon work, with improved coding over 1.2. It shipped the same day in Muse Code and Meta Model API. The launch post says it handles messy sources better, asks when a prompt is ambiguous, confirms before consequential actions, and knows when it is stuck.
Meta's methodology table is the vendor evidence. Columns: 1.3 max, 1.3 xhigh, 1.2 xhigh, GPT-5.6 Sol max, Opus 5 max. Compare 1.3 with 1.2 at matched xhigh. A 1.3 max score is not the same test as 1.2 xhigh. GDPVal-AA v2 is Elo (human baseline 1,000), not a percent.
| Benchmark | 1.3 xhigh | 1.2 xhigh |
|---|---|---|
| Terminal-Bench 2.1 | 89.2 | 82.9 |
| SWE-Atlas Codebase QnA | 54.0 | 46.2 |
| GDPVal-AA v2 | 1709 | 1615 |
| OSWorld 2.0 | 57.2 | 47.6 |
| DeepSearchQA | 89.4 | 85.9 |
| MRCR 256K–512K | 97.6 | 66.3 |
| MRCR 512K–1M | 93.1 | 55.5 |
| JobBench | 61.2 | 61.6 |
Terminal-Bench 2.1 is the cleanest coding read: 1.3 xhigh is 89.2. 1.3 max is 88.8, tying GPT-5.6 Sol max. Long context is the large step: 1.2 xhigh drops to 66.3 and 55.5 on those MRCR bands. JobBench is a wash. OSWorld 2.0 is 57.2 versus 47.6, but 1.2 ran version 06.24 and everyone else ran 08.08.
DeepSWE v1.1 is the easy mistake. 75.4 is 1.3 max only. The 1.3 xhigh cell is a dash. 1.2 xhigh is 55.0. That is not a same-effort upgrade. The 1.3 max DeepSWE run used mini-swe-agent; other DeepSWE cells came from the public leaderboard. It is not a Muse Code score.
Opus 5 max still leads several agent rows on that chart. Treat them as context, not a reason to skip 1.3 inside Spark. The Muse Spark vs Gemini page still uses 1.2 versus 3.7 Flash.
Audio is a step back
1.2 lists audio as a supported input. 1.3 marks it with a first-party warning. Keep 1.2, or Muse Voice Transcribe, when the prompt depends on hearing the file.
Pin reasoning effort before you read a score
Live reasoning_effort values are minimal, low, medium, high, and xhigh. none returns HTTP 400. Omit the parameter and the model still reasons at a model-determined level. Docs examples use high.
The 2 September launch post said previously available modes ship today, and max reasoning comes after more safety testing. The eval table still scored 1.3 at max. That is not the default API call. Pin effort before you read a vendor row.
Reasoning tokens count toward the output cap and billed completion. The Responses API can replay encrypted reasoning across turns. Chat Completions cannot for external keys. For multi-step agent loops, use Responses.
Muse Spark 1.3 Pricing: Standard vs Contributor
Same checkpoint, two contracts. Contributor is not a smaller model and not a weaker reasoning setting.
| Usage | Standard (muse-spark-1.3) | Contributor (muse-spark-1.3-contributor) |
|---|---|---|
| Input / 1M | $1.25 | $0.10 |
| Output / 1M | $4.25 | $0.20 |
| Cached input / 1M | $0.15 | $0.002 |
| Data terms | Not used to train Meta models | Prompts and completions may train future Meta models |
| Rate limits | 3,000 RPM / 4,000,000 TPM | 100 RPM / 3,000,000 TPM |
Contributor is a data-rights discount. Fine for throwaway prototypes. Wrong for proprietary code unless legal already said yes. Aggregators often lead with $0.10 because that is the Contributor floor. Quote Standard first.
Public API pricing is out. The July Meta Muse Spark review still describes the original family, before 1.3 and before list rates.
Cached input and search grounding ($2.50 per 1,000 queries) are extra line items. There is no long-context premium. MyClaw hosting and model tokens bill separately.
Who Should Use Muse Spark 1.3
| Choose 1.3 first | Keep 1.2 for now | Use Contributor only when |
|---|---|---|
| New Muse Code or Meta Model API work, long-horizon coding, browser and tool agents | Audio-heavy jobs; already-passing 1.2 loops that have not had a matched 1.3 run | Non-sensitive prototypes where training-on-data is acceptable and 100 RPM is enough |
Do not migrate a working 1.2 harness because the docs example ID changed. Confirm the ID the provider actually exposes, and write down substitutions.
Run Muse Spark 1.3 on a Real Agent Workflow

The model stops if the laptop sleeps, the terminal dies, or the files live in another chat. OpenClaw is the runtime: channels, tools, memory, and scheduled jobs. MyClaw hosts that runtime so the workspace stays up while you swap 1.2 and 1.3 IDs.
Step 1: Freeze one measurable Spark job
Pick a single outcome you already ran on 1.2: a failing test, a bounded refactor, or a sourced research memo. Write the pass/fail rule before either model starts. Skip audio jobs on 1.3.
Step 2: Run 1.2, then 1.3 at the same reasoning effort
Keep files, tools, permissions, and instructions identical. First pass is 1.2. Second pass is 1.3 at the same high or xhigh setting. Track tokens, tool calls, retries, leftover cleanup, and whether the ticket closed. A coding-agent workflow makes failure visible because tests and diffs do the scoring.
Step 3: Keep the hosted workspace so you can switch
Leave logs, diffs, notes, and artifacts in the same hosted workspace. The next run should start from the work, not a blank chat. MyClaw pricing covers the always-on host. Model tokens still bill separately. Use Standard unless you explicitly chose Contributor.
Final Verdict: Default 1.3. Keep 1.2 for Audio and Proven Jobs
Default new Spark work to muse-spark-1.3 Standard. Keep 1.2 for audio and for loops that already close. Confirm the win on one matched ticket before you migrate the rest.
Skip the Setup, Run Best-in-Class AI Agents Now
Launch a managed OpenClaw or Hermes Agent workspace in minutes, with always-on hosting, updates, and support handled by MyClaw.