Muse Spark 1.3 Is Out: Meta Claims Frontier Parity, and Its Own Table Mostly Backs It

§ 01What shipped#

Meta’s announcement leads with agentic workflows: a model trained across “a diverse set of harnesses” to sustain long tasks, juggle several workflows in one thread, ask clarifying questions when a prompt is ambiguous, and confirm before consequential actions. The model page carries the pricing table and the benchmark graphic.

The release cadence is the first thing to notice. Muse Spark launched in April, 1.1 arrived in July with the first paid API access, 1.2 shipped August 5 alongside Muse Code, and 1.3 landed four weeks later. That is a monthly cadence from the lab that, a year ago, was releasing open Llama weights on a much slower clock.

§ 02The benchmarks: Meta’s full table#

Meta publishes an 11-row comparison against three models: its own Muse Spark 1.2, OpenAI’s GPT-5.6 Sol, and Anthropic’s Claude Opus 5. The evaluation methodology states the rules: max effort for Muse Spark 1.3, Opus 5, and Sol; xhigh for 1.2; for each model, the highest comparable primary-metric value from Meta’s evaluation, the official leaderboard, or the provider’s self-reported result. Refusals and ungradable answers score zero and stay in the denominator.

Three readings, one chart each.

First, the generation jump. Every row moved up from 1.2, and the moves are large where 1.2 was weakest: retrieval deep into the context window, long-horizon coding, and driving a desktop.

Second, the parity claim. Put 1.3 next to Opus 5 and Sol on the seven rows where all three have a score, and the claim Alexandr Wang made to Bloomberg, “competitive” with Anthropic’s best and “better than” OpenAI’s, reads as mostly true on Meta’s own numbers. Against Opus 5 it is within a point on JobBench, AutomationBench, and DeepSearchQA, within 1.4 on OSWorld, and ahead on the three coding rows. Against Sol it wins everything except agentic browsing and Meta’s own instruction-following index. The one clear gap is GDPVal-AA, where Opus 5 holds a 70-point Elo lead.

Third, the row Meta wants you to see. The long-context numbers are the outliers in the whole table. On MRCR v2’s hardest band, 512K to 1M tokens with 8 needles, Muse Spark 1.3 scores 98.1 against 73.8 for GPT-5.6 Sol, and Meta reports no Opus 5 figure at all. If it holds up in independent runs, that is the difference between a model that can use a million-token window and a model that merely accepts one.

§ 03Two vendor tables, one week apart#

Here is a check anyone can run. Google published its own comparison table on September 2 for Gemini 3.8 Flash, and it also scored Claude Opus 5. On three benchmarks both vendors ran, the Opus 5 numbers do not match.

DeepSWE agrees because both vendors took it from Datacurve’s public leaderboard. OSWorld disagrees by seven points; Meta’s methodology notes it ran version 08.08 in an internal framework with its own computer-control tool, and Google’s page reports a partial score under its own setup. Terminal-Bench disagrees by 2.4 points, and Meta’s note says it ran each task “with the coding agent named in the result,” while Sol’s score came straight from OpenAI’s model card. None of this is bad faith. It is what happens when every vendor picks harnesses, versions, and effort levels, then reports the highest comparable number it can find. The practical rule: trust a vendor’s generation-over-generation deltas, treat its cross-vendor rows as a range, and wait for the independent index.

That index exists. Artificial Analysis scores Muse Spark 1.3 (max) at 62 on its Intelligence Index, sixth of 636 models, behind Claude Fable 5.1 (high) at 62.5 and Claude Opus 5 at 63.1, and ahead of GPT-5.6 Sol (max) at 60.9. The xhigh variant scores 60.8. That is a narrower and more defensible version of Meta’s claim: not ahead of the frontier, but on it, at a quarter to a fifth of Opus 5’s output price.

§ 04The pricing, and the second door#

The standard tier is straightforward, and Wang told Bloomberg developers pay no more for 1.3 than they did for 1.2. Against the models in Meta’s own table, the output price is the argument.

The second door deserves a plain reading. Meta’s pricing table lists two model ids. muse-spark-1.3 at $1.25 and $4.25 carries the label “Not used to improve our products.” muse-spark-1.3-contributor at $0.10 and $0.20 carries the label “Used to improve our products.” That is roughly a 92% discount in exchange for your prompts and outputs becoming training material. Meta is not hiding it; the two labels sit one row apart. For a hobby project it is close to free. For anything touching customer data, contracts, or code you would not publish, the standard tier is the only one to consider, and the discount is the price of that rule being easy to forget at 2am.

The efficiency claim matters as much as the unit price, and it points the opposite way from Google’s caveat a day earlier. Where Google warned that Gemini 3.8 Flash “works harder” and may spend more tokens per task, Meta says 1.3 spends fewer: about 20% fewer tool calls and 25% fewer tokens than 1.2 in comparisons by its own engineers. If that holds on your workload, the same job costs less on 1.3 than on 1.2 at an unchanged price. As with Google’s claim, the right move is a measurement on your real traffic, not a migration on a press release.

§ 05What Meta says about behavior and safety#

Three claims in the announcement bear directly on anyone handing this model real work.

• Confirmation before consequential actions. Meta says 1.3 has “better calibration on what constitutes irreversible actions and proceeds accordingly,” and Wang put it more simply: when it is about to do something irreversible, it asks. That is the right default for an agent with tool access, and it is a product decision as much as a model property.
• Clarifying questions and asking for help. The model “asks clarifying questions when prompts are ambiguous” and “invokes help from the user when stuck,” rather than guessing. Meta also says it has a better sense of what it does and does not know, and flags hurdles instead of hallucinating outcomes.
• Prompt-injection resistance, without a number. Meta claims “stronger adversarial robustness, with improved resistance to adversarial inputs and prompt injections” but publishes no figure. For scale, Google’s Gray Swan chart from September 2 put Muse Spark 1.2 at a 24.2% attack success rate within 15 attempts, against 4.8% for Claude Opus 5 and 5.5% for Gemini 3.8 Flash. Whether 1.3 closed that gap is exactly the number this release does not give you.

§ 06The max mode that is not shipping yet#

The announcement’s availability paragraph is careful: “Previously available reasoning modes are available today with max reasoning coming shortly after we finish additional safety testing.” Every number in the comparison table is from the max mode. Artificial Analysis describes Muse Spark 1.3 (max) as being in limited preview for Meta’s partners, and scores the publicly available xhigh mode more than a point lower on its index.

So the model you can call today through Muse Code or the API is not the model in the table. That is a normal staging pattern, and Meta says the gap is safety testing, not capability. It still means a developer benchmarking 1.3 this week should expect xhigh-class results, and the max column becomes real on the day Meta flips it on. We will update this page that day.

§ 07Who should care, and what to do#

If you build on Muse Spark 1.2 or Muse Code: 1.3 is already the default in Muse Code, and the API id changes to muse-spark-1.3. Re-run your evals at the effort level you actually use, check the token counts Meta says should fall, and read the two pricing labels before you pick a model id.

If you are choosing an AI agent or AI employee platform: count the week. Fable 5.1 on September 1, Gemini 3.8 Flash and Qwen3.8-Max-0902 on September 2, Muse Spark 1.3 the same day, GLM-5.3-Flash’s promotional pricing ending September 9. The model layer now rotates weekly. The question to ask a platform is not which model it runs today but whether the work you build on it survives the next swap: the roles, the memory, the accumulated context, the tasks in flight.

§ 08What it means for CellCog users#

CellCog does not run on Muse Spark 1.3. Its Flash tiers route to Gemini 3.8 Flash, its Max tiers to Claude Fable 5.1, and its Creative tier to Claude Opus 5, each swapped in on the day the model shipped. Every frontier release, this one included, is evaluated against that routing, and if 1.3’s public modes earn a place, the switch happens underneath your employees without a change on your side. Your pricing is not tied to any of it: a full shift of real work runs about $25, and the cost depends purely on how much work you assign. How the tiers fit together is in Flash, Core, Max.

§ 09The record#

As of September 3, 2026: Muse Spark 1.3 is available in Muse Code and the Meta Model API at $1.25/$4.25 per million tokens (contributor tier $0.10/$0.20); max reasoning is pending additional safety testing; Meta’s 11-row comparison table is published with its methodology; Artificial Analysis ranks the max preview sixth on its Intelligence Index; consumer rollout to Meta AI, Facebook, and Instagram is promised within days; open weights for Muse Spark are on Meta’s stated roadmap without a date. If max reasoning goes public, if independent replications land materially different from Meta’s numbers, if the consumer rollout completes, or if the open-weights release is dated, the update happens here, same day.