Meta Ships Muse Spark 1.3 With Big Long-Context Gains

Introduction

Meta Superintelligence Labs released Muse Spark 1.3 on September 2, 2026, an update aimed squarely at work that runs for hours rather than turns. The model ships immediately in Meta's Muse Code client for macOS and Linux and through the Meta Model API, at unchanged pricing of $1.25 per million input tokens and $4.25 per million output tokens.

The headline claim is efficiency: on Meta's own engineering comparisons, Muse Spark 1.3 finishes coding tasks with about 20% fewer tool calls and 25% fewer tokens than version 1.2. Two caveats sit underneath it. Meta's best benchmark numbers come from a maximum reasoning configuration developers cannot yet call, and independent measurement across a wider task mix found cost per task going up, not down.

The release lands four weeks after Muse Spark 1.2, which shipped August 5, 2026 alongside Muse Code, Meta's terminal-based coding agent. That cadence — a point release a month — is the context for reading the gains below.

What Changed in the Model's Behaviour

Meta describes 1.3 as built for a single long thread rather than a single long answer. The model runs several workflows in parallel inside one session, gathers its own context by calling tools across different sources, notices gaps in its own plan and patches them, and carries what it has learned forward to a finished result.

The second cluster of changes is about working with a person rather than around one. Muse Spark 1.3 asks clarifying questions when a request is ambiguous, escalates when it is stuck, and asks for confirmation before irreversible actions. On long tasks it adapts to the user's preference for either regular check-ins or quiet background work. In long conversations it is also better at deciding which task a new message belongs to — whether the user is amending an old request or interrupting the current one.

Meta also claims better calibration about the model's own limits: on hitting a dead end it should say so rather than invent a result. That is a direct response to the failure mode described in AI hallucinations, and it is the kind of claim that only holds up under sustained use, not a benchmark run.

How to Read Meta's Scorecard

One methodology detail governs every table below. On Meta's published scorecard, Muse Spark 1.3, Claude Opus 5 and GPT-5.6 Sol were all run at max reasoning effort — but Muse Spark 1.2 was run at xhigh.

So the competitor comparisons are like-for-like, and the generational ones are not. Every jump over 1.2 in these tables is 1.3's best configuration measured against 1.2's second-best. The gains are real, but they are not purely model improvement, and Meta does not publish a 1.2 max column to separate the two. Blank cells below are benchmarks Meta did not publish for that model.

Coding

Meta trained 1.3 on large engineering tasks and reports fewer wasted moves, less commentary and cleaner code. All numbers are Meta-reported, from the launch scorecard.

The jump over 1.2 on DeepSWE — 55.0 to 75.4 — is far larger than the margin over Claude Opus 5 and GPT-5.6 Sol, which is the more honest way to read the table: Meta closed a large gap to its own previous version and now sits level with the frontier on coding rather than above it.

Long Context Is the Real Result

The sharpest gains are on long-context retrieval, and it is the one category where Muse Spark leads on every published number. On MRCR, a multi-needle retrieval benchmark run inside the model's 1M-token context window:

Anthropic has published no MRCR result for Claude Opus 5, so this is a lead over what OpenAI reports, not a measured lead over the whole field. The 512K–1M row is still the striking one: a model that scored 55.5 one version ago now scores 98.1, and the nearest published competitor is 24 points behind.

Agentic Work Is a Closer Fight

On agentic benchmarks the picture reverses. Claude Opus 5 leads four of six, and GPT-5.6 Sol takes the other two.

Every gain over 1.2 is large. Every gap to a competitor is small. Both things are true at once, and Meta's own scorecard says so.

The Configuration Caveat

The numbers above are max reasoning results, and max reasoning is not what shipped this week. Meta says the maximum reasoning modes arrive "shortly after we finish additional safety testing." What developers can call now is the xhigh configuration, and the difference is not cosmetic: Artificial Analysis measured OSWorld 2.0 at 57.2 for xhigh against 66.9 for max, GDPval-AA v2 at 1709 against 1754, and JobBench at 61.2 against 64.9.

Artificial Analysis scored the shipping xhigh configuration at 61 on its Intelligence Index at the time of publication, tying 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. Meta has not published pricing for the max configuration.

The Efficiency Claim Does Not Survive a Wider Test

Meta's "20% fewer tool calls, 25% fewer tokens" is a coding-workflow measurement made by Meta's own engineers. Artificial Analysis ran a broader suite of reasoning and agentic tasks and found the opposite where it counts:

That is a 37% increase in the cost of finishing a task, at identical per-token pricing, attributed primarily to heavier input-token consumption on agentic evaluations. Both findings can hold: the model can be leaner inside a coding loop and hungrier across a wider mix of agentic work.

The practical lesson is that a per-token rate is not a price. Token rates, reasoning effort, number of turns, tool calls and retries all determine what a finished task actually costs — which is why $0.55 per task still made Muse Spark 1.3 the cheapest model at its score level even as its own cost per task rose.

Safety and Availability

Meta reports that 1.3 is more robust to adversarial inputs and prompt injection, and better calibrated on which actions are irreversible during complex agentic tasks — a pairing that matters more as AI agents are given tools that write to real systems.

The model carries the same 1M-token context window and closed weights as Muse Spark 1.1. Meta says a larger model is coming, and Mark Zuckerberg has committed to an open-weights release of Muse Spark 1.2 — the previous version, not this one.

Conclusion

Muse Spark 1.3 is the version where Meta's model stops trailing the frontier and starts trading with it. The long-context results are ahead of everything OpenAI has published, the coding results are level with Anthropic and OpenAI, and the agentic results sit a point or two behind Claude Opus 5 across most of the suite — at a quarter of Sol's input-token price.

Three things temper it. The scorecard advertises a max reasoning mode still in safety testing, so the model developers can call this week is the weaker one. The generational gains compare that unreleased max mode against 1.2 at a lower reasoning setting. And the efficiency headline reverses under independent testing, where finishing a task got 37% more expensive. Meta has closed the capability gap; the claim that it did so while getting cheaper is the part that does not hold up.

Sources

• Introducing Muse Spark 1.3 (Meta AI Research)
• Muse Spark 1.3 model page and benchmark scorecard (Meta for Developers)
• Muse Spark 1.3: Meta reaches the frontier (Artificial Analysis)
• Muse Spark 1.2: Benchmarks and analysis (Artificial Analysis)
• Meta says Muse Spark 1.3 has frontier performance — but its best results come from a model developers can't broadly use yet (VentureBeat)
• Meta AI Released Muse Spark 1.3 (MarkTechPost)
• Meta Releases AI Model Muse Spark 1.3, Edges Closer to OpenAI, Anthropic (Bloomberg)
• Meta debuts Muse Spark 1.3 as personal agent work continues (Axios)