Muse Spark 1.3

Meta's frontier reasoning model, shipped September 2, 2026: a 1M-token multimodal agent model at $1.25/$4.25 per million tokens.

Last reviewed September 6, 2026

At a glance

Overview

Muse Spark 1.3 is Meta’s frontier reasoning model, released by Meta Superintelligence Labs (MSL) on September 2, 2026. It is the fourth Muse Spark release in five months, and the first one Meta has been able to put next to Anthropic’s and OpenAI’s flagships without the comparison being embarrassing.

Meta positions 1.3 for work that runs for hours rather than turns: long-running agentic sessions, multi-agent orchestration, and coding across large repositories. Three behaviours carry that framing — keeping track of requirements across an extended task without dropping constraints, working through conflicting or incomplete inputs, and asking a clarifying question rather than guessing when a request is ambiguous. The third is the one that matters most in practice and the one no benchmark measures.

The line has moved fast enough that the version you are quoting matters more than usual:

Muse Spark remains closed-weights. Meta’s open-weight work now runs on a separate track, in the 30B Muse Glimmer model released on August 10, 2026, and Mark Zuckerberg has committed to eventually opening the weights of Muse Spark 1.2, not this release.

Capabilities

• Long-horizon agentic execution — the model runs several workflows in parallel inside one session, gathers its own context by calling tools across sources, notices gaps in its own plan and patches them, then carries what it learned forward to a finished result.
• Clarification and escalation — 1.3 asks when a request is ambiguous, escalates when stuck, and asks for confirmation before irreversible actions. On long tasks it adapts to whether the user wants regular check-ins or quiet background work.
• Turn routing in long conversations — better at deciding whether a new message amends an earlier request or interrupts the current one.
• Coding — trained on large engineering tasks; Meta reports fewer wasted moves, less commentary and cleaner diffs, at roughly 20% fewer tool calls and 25% fewer tokens than 1.2 on its own coding comparisons.
• Long-context retrieval — the strongest published result on the model, and the one category where Muse Spark leads every number Meta shows. See context window for why retrieval quality and window size are different things.
• Multimodal input — text, images, video and documents including PDFs. Meta notes that audio understanding is not fully supported and that responses to audio content may be degraded.
• Tool use and orchestration — parallel tool calls, zero-shot generalisation to new native tools and Model Context Protocol servers, structured output, and built-in web search grounding. The shape is multi-agent, not single-turn chat.
• Calibration — Meta claims the model says it has hit a dead end rather than inventing a result. That is a claim about hallucination behaviour that only sustained use can settle.

Technical Specifications

• Developer: Meta Superintelligence Labs (MSL), part of Meta
• Model type: multimodal reasoning model, closed weights
• Model IDs: muse-spark-1.3 (standard), muse-spark-1.3-contributor (discounted, data-retaining). Both were published on 2026-09-02.
• Context window: 1,048,576 tokens
• Maximum output: 943,718 tokens, reported by API catalogues (OpenRouter, llm-stats) rather than by Meta’s own model page. Output and input share the one budget, so a full-length completion leaves roughly 105,000 tokens for the prompt.
• Input modalities: text, image, video, documents/PDF. Audio is accepted but explicitly flagged as not fully supported.
• Output modality: text
• Reasoning effort: xhigh is the configuration that shipped; a max configuration was withheld at launch pending further safety testing. Reasoning tokens bill at the output rate.
• Knowledge cutoff: not disclosed by Meta.
• Architecture and parameter count: not disclosed. No model card, no mixture-of-experts confirmation, no layer configuration.
• API: Meta Model API at dev.meta.ai, OpenAI SDK-compatible. Also distributed via OpenRouter. No batch API and no batch discount.

Use Cases

• Long-running coding agents — repository-scale refactors, migrations and bug hunts where the session outlives the context of a single prompt. This is the workload Meta trained 1.3 for and the one its efficiency claim is measured on.
• Multi-agent orchestration — decomposing a task across parallel sub-agents inside one session rather than spawning a process per task. See multi-agent systems.
• Whole-corpus retrieval and analysis — the MRCR results are the reason to reach for this model over a cheaper one when the input genuinely is hundreds of thousands of tokens.
• Computer use — navigating across applications and services over an extended session; computer use benchmarks are where 1.3 gained most over 1.2.
• Cost-sensitive agentic pipelines — at a quarter of the per-token price of the top-tier OpenAI and Anthropic models, Muse Spark is the cheapest model at its measured intelligence level.
• Experimentation on the Contributor tier — prototyping where a 21x discount is worth handing Meta your prompts. Never for regulated or customer data.

Performance / Benchmarks

Two provenance notes govern every number here. First, unless stated otherwise these are Meta-reported figures from the 1.3 launch scorecard. Second, Meta ran Muse Spark 1.3, Claude Opus 5 and GPT-5.6 Sol at max reasoning effort, but Muse Spark 1.2 at xhigh — so the competitor columns are like-for-like and the generational ones are not.

Coding (Meta-reported, September 2026):

Long context — MRCR v2 multi-needle retrieval, run inside the 1M-token window (Meta-reported):

Anthropic publishes no MRCR figure for Claude Opus 5, so this is a lead over what OpenAI reports, not over the field.

Agentic (Meta-reported) — the picture reverses: Claude Opus 5 leads four of six, and OpenAI’s model takes the other two.

Third-party: Artificial Analysis measured the shipping xhigh configuration at 61 on Intelligence Index v4.1 — level with GPT-5.6 Sol at max reasoning, Grok 4.6 at high and Claude Opus 5 at high, and behind Claude Fable 5.1 at 66. The withheld max configuration scored 62. The same v4.1 measurement across four releases: 1.1 at 53, 1.2 at 57, 1.3 at 61.

Artificial Analysis then replaced the index. Index v4.2 shipped on September 4, 2026 — it adds the AA-Briefcase and GDP.pdf evaluations, retires GPQA Diamond as too easy, and doubles the private held-out weighting to 40%. Under v4.2 the Muse Spark 1.3 model page reports 53 at max effort, against Gemini 3.8 Flash at 59. The v4.1 and v4.2 numbers are not comparable, and the drop from 61 to 53 is a change of ruler, not of model — the same re-scoring moved every model AA tracks. Quote whichever you like, but never the two side by side.

Cost per task rose against that: Artificial Analysis put 1.2 at $0.40 per Index task and 1.3 (xhigh) at $0.55, a 37% increase at identical per-token pricing, attributed to heavier input-token consumption on agentic evaluations. Meta’s “20% fewer tool calls” holds inside a coding loop; it does not hold across a wider agentic mix. Our own write-up of the release goes through this in detail: Meta Ships Muse Spark 1.3 With Big Long-Context Gains.

Limitations

• The headline scores use a mode you cannot call. Max reasoning was still in safety testing at launch. The gap is not cosmetic: Artificial Analysis measured OSWorld 2.0 at 57.2 for xhigh against 66.9 for max, and GDPval-AA v2 at 1709 against 1754.
• The generational comparison is unequal. Every gain over 1.2 on Meta’s scorecard is 1.3’s best configuration against 1.2’s second-best, and Meta publishes no 1.2 max column to separate model improvement from reasoning-effort improvement.
• Cost per finished task went up. A lower per-token rate is not a lower bill. Independent measurement found tasks costing 37% more than on 1.2.
• Closed weights. No self-hosting, no fine-tuning, no offline deployment, no inspection. The open-weights commitment covers 1.2, not this release.
• The Contributor tier trains on your data. The 21x discount is paid for in prompts and completions. It is unsuitable for anything regulated, confidential or customer-owned.
• Output and input share one budget. The 943,718-token output ceiling is not additive to the context window; asking for a very long completion leaves almost no room for the prompt.
• Audio is degraded. Meta flags audio understanding as not fully supported.
• No published knowledge cutoff, parameter count or architecture. There is no model card with the specifications an integrator would expect.
• No batch API. There is a background submission mode with its own rate cap, but no discounted batch tier — which matters if your cost model assumed one.
• Fast deprecation risk. Four releases in five months is a good problem for capability and a bad one for anything you have pinned to a model ID.

Pricing & Access

Rates are per million tokens on the Meta Model API and have not moved since Muse Spark 1.1 in July 2026.

• Web search grounding: $2.50 per 1,000 search queries, on top of token cost.
• Free credits: $20 for new Meta Model API accounts.
• Cached input at $0.15 is 12% of the base input rate and undercuts several competitors’ uncached input price — the reason prompt caching is worth engineering for on this model.
• Reasoning tokens bill as output at $4.25, which is where an agentic bill actually accumulates.
• The Contributor tier is the same checkpoint at a 21x discount, priced by the data-retention right rather than by capability. Standard-tier traffic is not retained for product improvement.
• Max reasoning: no published price; it was not generally available at launch.
• Consumer access: free in Meta AI on meta.ai and across Meta’s apps.

Ecosystem & Tools

• Muse Code — Meta’s terminal coding agent for macOS and Linux, launched with Muse Spark 1.2 on August 5, 2026 and updated to 1.3. It runs concurrent sub-agents and keeps a local append-only event log of every model call, tool run, approval and edit.
• Meta Model API — the self-serve, OpenAI SDK-compatible endpoint at dev.meta.ai, with built-in web search grounding.
• OpenRouter — third-party routing for both the standard and Contributor model IDs.
• Muse Glimmer — Meta’s open-weight 30B model, released Apache 2.0 on August 10, 2026, with a quantized build that runs offline on a single 24 GB consumer GPU. It is the self-hostable counterpart to Muse Spark, not a smaller version of it.
• Muse Image and Muse Video — MSL’s media generation models, launched July 7, 2026. Muse Image integrates with Muse Spark, invokes search and coding tools to improve accuracy, and embeds Content Seal watermarking.
• Llama 4 — Meta’s earlier open-weight family, no longer the top of Meta’s lineup.

Community & Resources

• Introducing Muse Spark 1.3 — the Meta AI Research launch post
• Muse Spark 1.3 model page and scorecard — Meta for Developers, including the pricing table
• Meta Model API pricing and rate limits — the official rate card
• Build with Muse Spark on Meta Model API — Meta’s developer guide
• Muse Spark 1.3: Meta reaches the frontier — independent evaluation
• Meta Ships Muse Spark 1.3 With Big Long-Context Gains — our release analysis
• Meta Launches Muse Spark 1.1, Its First Paid AI Model — the release that opened the API