Muse Spark 1.3 is built for agentic coding and long-context work, not audio or consumer chat: Meta reports roughly 20% fewer tool calls and 25% fewer tokens than its predecessor on equivalent tasks. Teams already running GPT-5.6 Sol or Claude Opus 5 have limited reason to switch unless long-context recall specifically matters.
Muse Spark 1.3 is Meta AI Research's fourth Muse Spark release, shipped September 2, 2026 with a 1,000,000-token context window for agentic coding and long-document work. Its generally available xhigh tier scores 94% on GPQA Diamond, among the strongest published results of any September 2026 frontier model. The release drops audio input and stays developer-first via the Meta Model API.
Where it sits
- $2.00/M$ per 1M tokensBlended price (3:1)Lower is better#34 / 64peer median $1.70/Mvendor price, checked by HokAI
- 182 tok/stokens/sOutput speedHigher is better#12 / 39peer median 90 tok/scited: Artificial Analysis
- --% solvedSWE-bench VerifiedHigher is better-- / 28peer median 78.3%per source, see benchmark scores
- 94%% correctGPQA DiamondHigher is better#7 / 44peer median 88.3%per source, see benchmark scores
Priced around the middle of the 64 GA models with a published price (rank 34), in the top third on GPQA Diamond (rank 7 of 44), and one of 65 whose vendor states it does not train on customer data.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: Meta AI · Family: Muse
Context window: 1,000,000 tokens · Max output: 131,072
Input modalities: text, image, video, pdf, tool-calls · Output: text, tool-calls
About Muse Spark 1.3
Meta AI Research shipped Muse Spark 1.3 on September 2, 2026, the fourth entry in the Muse Spark line after the original release (April 8, 2026), Muse Spark 1.1 (July 9, 2026), and Muse Spark 1.2 (August 5, 2026), all built by the same Meta Superintelligence Labs team. Where the April launch targeted Meta's own consumer apps first, 1.3 is pitched at developers: it ships through the Meta Model API and Muse Code, Meta's terminal-based coding agent, rather than through the Meta AI consumer app. Architecture and parameter count remain undisclosed, consistent with every prior Muse Spark release.
The Artificial Analysis Intelligence Index puts the publicly available xhigh reasoning-effort tier at 61. A higher max tier scores 62 and ranks 6th of 643 tracked models, behind only Claude Fable 5.1 and Claude Opus 5, though max stays in limited partner preview. On GPQA Diamond, the xhigh tier scores 94%, up sharply from the 89.5% the original April model posted. Meta does not publish results on SWE-bench Verified, AIME, MMLU-Pro, or ARC-AGI-2 for any Muse Spark release, so those figures stay unconfirmed rather than estimated from adjacent benchmarks.
The context window grew from 262,144 tokens at launch to 1,000,000 tokens in 1.3, close to a fourfold increase. Meta has not published a max output token limit for this release. On Meta's own MRCR long-context recall test, the max variant scores 98.5% across the 256K-512K token span.
Muse Spark 1.3 accepts text, images, video, and documents as input. It does not currently support audio input, a capability the original model had; Meta's own documentation points developers to Muse Spark 1.2 or a separate tool, Muse Voice Transcribe, for speech-to-text work instead. Reasoning effort is selectable across five published levels (minimal, low, medium, high, xhigh), and Contemplating Mode, the multi-agent parallel-reasoning feature introduced at launch, carries over unchanged. A sixth level, max reasoning, was announced but had not shipped publicly as of the September 2 release, pending additional safety testing.
Two pricing tiers apply to the publicly available xhigh model: a Standard tier billed per token with a modest cache discount, and a Contributor tier priced far lower for developers willing to let Meta train on their prompts and completions. The pricing FAQ below has the exact per-token rates. Pricing for the max tier has not been published, since it remains a limited partner preview.
Access runs through the Meta Model API, which is OpenAI-SDK compatible, plus Muse Code and third-party routing through OpenRouter. Meta has not announced availability through AWS Bedrock, Google Vertex AI, or Azure. Unlike the original April launch, which rolled out through Meta's consumer apps immediately, the equivalent consumer rollout for 1.3 was not live at launch.
Meta reports stronger adversarial robustness and prompt-injection resistance in 1.3 compared with 1.2, plus better calibration around irreversible agentic actions such as deleting a file or sending a message on a user's behalf. The model's safety evaluation is documented in Meta's Safety and Preparedness Report, following the company's Advanced AI Scaling Framework. Training data cutoff, parameter count, and full architecture remain undisclosed across every Muse Spark release to date.
Muse Spark 1.3 is best suited to agentic coding work through Muse Code or the Meta Model API, tasks that need to recall facts from very long documents, and teams comfortable with the Contributor tier's data-training tradeoff for cheaper access. It is a weaker fit for anyone needing audio input, dropped from this release, and for teams that need guaranteed access to the strongest max variant, which stays invite-only. Teams already committed to GPT-5.6 Sol or Claude Opus 5 pipelines have less reason to switch unless long-context recall or coding-token efficiency specifically matters to their workload.
Versus Muse Spark 1.2, Meta says 1.3 needs about a fifth fewer tool calls and roughly a quarter fewer tokens on equivalent coding tasks, plus shows better adherence to long, multi-step instructions. The Contributor pricing tier, introduced alongside Muse Code, carries into 1.3 unchanged.
Pricing
Standard (xhigh) tier: $1.25 per 1M input tokens, $4.25 per 1M output, $0.15 per 1M cached input. Contributor tier: $0.10 input and $0.20 output per 1M tokens; Meta trains on submitted prompts and completions in return for that lower rate. Max-tier pricing is unpublished, since that variant is invite-only for Meta's partners.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.037 | $0.0043 | $0.042 |
| Support reply | $0.0025 | $0.0013 | $0.0038 |
| One coding agent run | $0.250 | $0.085 | $0.335 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- Expanded Context Window: Big enough to hold a full multi-file codebase in one prompt, close to four times the original Muse Spark's context window from April 2026.
- Five Reasoning-Effort Levels: Choose minimal, low, medium, high, or xhigh per request; a higher max tier exists but was not open to the public at launch.
- Contemplating Mode: Runs multiple reasoning agents in parallel on one problem and synthesizes their output into a single answer, carried over unchanged from the original Muse Spark.
- Muse Code Terminal Agent: A terminal-based coding agent co-trained with the model that can run multiple coding sub-agents concurrently on a single task.
- Contributor Pricing Tier: A steeply discounted per-token rate for developers who let Meta train on their prompts and completions, unchanged since its debut alongside Muse Code.
Pros
- The max tier's composite Artificial Analysis score of 62 lands a rank matched only by Claude Opus 5 and Claude Fable 5.1 among 643 tracked models.
- Cost-efficient at its intelligence tier: Artificial Analysis measured $0.55 per Intelligence Index task-point for the xhigh tier, cheaper than GPT-5.6 Sol at $0.95 or Grok 4.6 at $0.94 for the same score.
- Contemplating Mode's parallel multi-agent reasoning remains a genuine architectural differentiator: few widely deployed frontier models run multiple reasoning agents at inference time the same way.
Cons
- Dropped audio input entirely; Meta points developers to a separate transcription tool or an earlier Muse Spark release for speech-to-text work.
- The highest-capability max tier is invite-only for Meta partners, with no public pricing as of September 2026.
- No confirmed consumer app rollout at launch, unlike the original Muse Spark, which shipped inside Meta's consumer apps immediately.
Benchmarks
- GPQA Diamond: 94% vendor-reported · 02 Sep 2026 — PhD-level science questions that are hard to search for, % correct.
- Deepswe V1 1 Max: 75.4 vendor-reported · 02 Sep 2026
- Mrcr 256k 512k Max: 98.5 vendor-reported · 02 Sep 2026
- Humanity's Last Exam: 47% vendor-reported · 02 Sep 2026 — Expert-written questions across many fields, % correct.
- Gdpval Aa V2 Elo Max: 1754 vendor-reported · 02 Sep 2026
- Terminal Bench 2 1 Max: 88.8 vendor-reported · 02 Sep 2026
- AA Intelligence Index: 61 cited: Artificial Analysis · 02 Sep 2026 — Composite of 10 evaluations run by Artificial Analysis, 0 to 100.
- Output speed: 182 tok/s cited: Artificial Analysis · 02 Sep 2026 — Median tokens written per second as measured by Artificial Analysis.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
How much do you pay for Muse Spark 1.3?
The publicly available xhigh tier runs $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15 per million. A cheaper Contributor tier costs $0.10 input and $0.20 output per million tokens, in exchange for letting Meta train on your prompts and completions. Pricing for the higher-capability max tier has not been published; it remains a limited partner preview as of September 2026.
Does Muse Spark 1.3 have a free plan?
Not through the API: every tier, Contributor included, charges per token, and Meta has not published a free developer quota for 1.3. Earlier Muse Spark models were free through the Meta AI consumer app, but Meta says that consumer rollout for 1.3 was still pending at launch, so there was no confirmed free access route as of September 2, 2026.
What should you use instead of Muse Spark 1.3?
For a model with a fully public API and no data-training tradeoff, GPT-5.6 Sol and Claude Opus 5 are both generally available today. Gemini 3.1 Pro is worth considering if raw science-reasoning benchmarks matter more than agentic coding efficiency. Anyone who needs audio input, which Muse Spark 1.3 dropped, should look at Muse Spark 1.2 instead.
What separates Muse Spark 1.3 from GPT-5.6 Sol?
On long-context recall, Muse Spark 1.3's max variant leads clearly: 98.1% versus GPT-5.6 Sol's 73.8% on Meta's 512K-1M token MRCR test. On coding, Muse Spark 1.3 also edges ahead on DeepSWE v1.1 (75.4 versus 73.0) and ties on Terminal-Bench 2.1 at 88.8. GPT-5.6 Sol still wins some agent benchmarks, including DeepSearchQA, so the better pick depends on which workload matters more.
How long does it take to get going with Muse Spark 1.3?
Developers can start in minutes with an API key through the Meta Model API, which uses an OpenAI-compatible SDK, or by routing through OpenRouter. Muse Code, Meta's terminal coding agent, adds a ready-made agentic coding workflow on top of the base model. There is no confirmed consumer app path for 1.3 yet, unlike earlier Muse Spark versions.
Top Alternatives
- GPT-5.6 Sol: Pick Muse Spark 1.3 for long-context recall and coding efficiency; pick GPT-5.6 Sol if you need a fully public API with no partner-preview gating.
- Claude Opus 5: Pick Claude Opus 5 for the strongest overall Artificial Analysis Intelligence Index score; pick Muse Spark 1.3 if million-token recall or the Contributor discount matters more.
- Gemini 3.1 Pro: Pick Gemini 3.1 Pro for a fully public API today; pick Muse Spark 1.3 if you're already building agentic coding workflows around Muse Code.
HokAI guides covering Muse Spark 1.3
- What Meta AI Is Actually For, Now Its Model Keeps Changing Its License: Meta AI's model went closed in April 2026, then reopened in August. In between, Meta added $7.99-$19.99 subscriptions and $2-per-million-token business billing.