Muse Spark 1.3: Meta's Model Ties Opus 5 on Coding Benchmarks

Update — September 14, 2026: Chief AI Officer Alexandr Wang confirmed on X that the entire Spark 1 through 1.3 line was built with one target in mind: Muse, Meta's personal-agent product (they share a name "for a reason," per Wang). His framing: "each of these steps were secretly to set the stage for our personal agent," and Meta Superintelligence Labs (MSL) has been chasing "personal superintelligence" — AI superpowers for everyone — since it started last year. That reframes 1.3's coding and agentic gains documented below not as isolated model improvements, but as the last checkpoint before productizing them into Muse. See Wang's full statement analyzed below.

Meta's press blog called Muse Spark 1.3 a model "optimized for competitive coding performance." Its own benchmark table backs that up in a way most model-launch claims don't: on DeepSWE v1.1, a long-horizon agentic coding benchmark, Muse Spark 1.3 scores 75.4 — ahead of Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0). On Terminal-Bench 2.1, it ties GPT-5.6 Sol at 88.8 and edges past Opus 5's 86.7. Those aren't marketing rounding errors; they're the two coding benchmarks Meta chose to headline, and the numbers hold up against the field's best.

Muse Spark 1.3 rolled out September 3, 2026, replacing Muse Spark 1.2 as the model behind Muse Code, Meta's terminal-based coding agent. Mark Zuckerberg announced it on X: "Muse Spark 1.3 is rolling out today with frontier performance almost too cheap to meter. This is the biggest jump we've made so far on coding and agentic work." He also teased what's next — open weights for the Spark line, and a mysterious 🍉-codenamed release, both "coming soon" with no further detail.

The launch also comes with a pricing structure worth understanding before you touch the API: a "contributor" tier that's up to 21x cheaper in exchange for letting Meta train on your usage.

TL;DR

What's actually new in Muse Spark 1.3

Meta's introducing post describes 1.3 as delivering "max reasoning for challenging reasoning and agentic tasks and improved real-world usability," alongside native multimodal perception. In practice, the benchmark table shows the jump concentrated in three places: long-horizon coding, agentic terminal use, and long-context retrieval — not a uniform lift across every category.

Long-context retrieval is where the biggest jump shows up

On MRCR (a multi-round coreference retrieval benchmark that stresses whether a model can actually use a long context window rather than just accept one), Muse Spark 1.3 posts a huge improvement over its own predecessor:

Going from 55.5 to 98.1 on the 512K-1M band isn't an incremental tune — it's a different regime of long-context reliability. Opus 5 wasn't reported on MRCR in Meta's table, so there's no direct three-way comparison there, but Spark 1.3 clearly outpaces GPT-5.6 Sol at the longest context lengths.

Coding: where Spark 1.3 leads or ties the field

This is the headline result for a technical audience:

Muse Spark 1.3 leads outright on DeepSWE v1.1 and SWEAtlas, and ties GPT-5.6 Sol for the top Terminal-Bench 2.1 score — with both ahead of Opus 5 on that one. Terminal-Bench itself is a benchmark explainx.ai has covered in depth since its 2.0 release: it measures whether an agent can actually operate a real terminal session end to end, not just produce a diff.

Agent benchmarks: a more mixed picture

Outside coding, Opus 5 keeps its lead on most agent-focused evaluations:

Spark 1.3 improves on every one of these versus Spark 1.2 — often by double digits, as on OSWorld 2.0 (47.6 → 66.9) — but it doesn't top Opus 5 or, on DeepSearchQA and Agentic IF Index, GPT-5.6 Sol either. The honest read: Meta closed most of the gap to the frontier on general agentic work, and specifically pulled ahead on coding.

The pricing table, and what "contributor" actually means

This is the part worth reading carefully before you point production traffic at the API. Meta ships Muse Spark 1.3 as two SKUs with the same 1M-token context window and wildly different prices:

Do the math and the gap is stark: input is roughly 12.5x cheaper on the contributor tier, output roughly 21x cheaper. For a team running high-volume agentic workloads — the kind Muse Code's fan-out sub-agent pattern generates — that's not a rounding difference, it's the line between "affordable" and "not affordable at scale."

Is this actually transparent, or just a discount with strings attached?

It's both, and it's worth naming plainly: muse-spark-1.3-contributor means Meta can use what you send it to improve its own models and products. The muse-spark-1.3 tier means it won't. Compared to how some other providers handle training-data consent — an opt-in toggle buried in account settings, easy to leave on by accident, or off by default with a separate paid tier — Meta's version is unusually legible: the two SKUs are named for the trade, priced for the trade, and sit next to each other in the same table. You can't miss it.

The practical guidance: use the contributor tier for hobby projects, prototypes, internal tooling, and anything without proprietary code, credentials, or customer data flowing through it — as some developers on OpenCode already did with Muse Spark 1.2, running personal projects at close to zero marginal cost. Use the non-contributor tier for anything touching production code, client work, or regulated data, where the ~12.5-21x premium buys you a clean answer to "does this vendor train on our code."

What people are asking

Does it act like a tool, or does it improvise? Developers who used Spark 1.2 for real coding work consistently noted it followed instructions precisely, without unprompted scope creep or gratuitous refactors — a trait some larger, more "creative" models lack. Early impressions on 1.3 point the same direction, reinforced by the DeepSWE and Terminal-Bench numbers, which specifically reward an agent that completes the stated task rather than wandering into adjacent changes.

Is the contributor tier actually usable for real work, or a bait price? At $0.10/M input tokens, it's cheap enough that several developers reported running hobby and personal projects on Spark 1.2's contributor tier for effectively no cost. The trade-off is real, not fine print, but for non-sensitive workloads it's a genuinely good deal — not a teaser rate that reverts after a free trial.

How does it compare to GPT-5.6 Sol and other budget-friendly coding models? On the coding benchmarks specifically, Spark 1.3's best value shows up relative to cost — leading DeepSWE and tying Terminal-Bench 2.1 while its contributor-tier price undercuts most frontier alternatives. See explainx.ai's broader look at budget coding subscriptions under $20 and the four-way Fable 5 vs Grok 4.6 vs GPT-5.6 Sol vs Qwen3-8 Max comparison for how it stacks up against the rest of the field on cost per output.

What's the 🍉 Zuckerberg mentioned? Unconfirmed. His post named it directly alongside a promise of open-weight Muse Spark releases, but Meta hasn't published details on either. Treat it as a genuine forward-looking signal, not a shipped feature — worth watching, not worth planning around yet.

Update, September 8-9: Meta launched Muse, the personal agent built on this model

Meta shipped Muse, its consumer personal-agent app, running on Muse Spark 1.3. The bigger story there isn't the model — it's the security architecture: a per-user Secure VM, a Sentinel process that brokers every network request and connector action, and a public bug bounty paying up to $130,000 for a working prompt-injection exploit. See the full architecture breakdown.

Update, September 4-5: Meta is now pushing everyone toward the "max" reasoning tier

Meta's Chief AI Officer Alexandr Wang posted on X the day after this launch, publicly recommending the max reasoning tier specifically — the same tier this post's benchmark tables above are drawn from — over the high and xhigh levels some users had already been trying:

"we just publicly released Muse Spark 1.3 max! we see significantly stronger coding and agentic performance on muse spark 1.3 max, so would strongly recommend trying it out even if you've already tried muse spark 1.3 high or muse spark 1.3 xhigh."

He also stated the release completed safety testing before shipping ("we are launching this after completing our safety testing") and gave the fastest path to try it via Meta's terminal coding agent:

curl -fsSL https://dev.meta.ai/install.sh | bash

Then select muse spark 1.3 and set the reasoning level to max inside Muse Code, or pass max as the reasoning parameter directly through the Meta Model API or OpenRouter.

What's actually new here versus the September 3 launch: no new benchmark numbers accompanied Wang's post — the DeepSWE v1.1, Terminal-Bench 2.1, and MRCR figures in this post's tables were already run at the max reasoning level. What changed is availability and framing: max is now the tier Meta is actively steering users toward as the default worth paying the extra reasoning cost for, with an explicit public statement that safety testing was completed prior to this wider push. If you tried Spark 1.3 at high or xhigh and found it underwhelming against Opus 5 or GPT-5.6 Sol, Meta's own claim is that the gap closes at max — which lines up with the benchmark tables above, all of which already reflect that tier.

Alexandr Wang: the Spark line was built for Muse all along

On September 14, 2026, Wang posted a retrospective that reframes everything documented above. His claim: Meta Superintelligence Labs (MSL) never treated Spark 1 through 1.3 as a coding-model line that happened to also power an agent later — the models were built specifically to become Muse from the start:

"we have been building our models (muse spark 1 through 1.3) specifically to be exceptional for muse over many months. in each of our releases, we made big gains on agentic and multimodal capability. each of these steps were secretly to set the stage for our personal agent, @Muse. (they're named the same thing for a reason… we planned ahead!)"

"when we started MSL last year, the goal was always to develop personal superintelligence. that always meant giving AI superpowers to everyone in the world, and the personal agent is the perfect manifestation of that." — @alexandr_wang, September 14, 2026

Two claims worth separating. First, the technical one: Wang says each Spark release's "big gains on agentic and multimodal capability" — the same benchmark jumps documented in the TL;DR above — weren't generic model improvements but were specifically selected to serve Muse's use case (long-horizon tool use, multimodal perception, terminal-agent workflows). That's consistent with what the numbers already showed: Spark 1.3's biggest gains over 1.2 landed on DeepSWE v1.1 and Terminal-Bench 2.1, both long-horizon agentic benchmarks rather than raw knowledge or single-turn coding tests — exactly the profile a personal agent needs and a chat-only assistant doesn't.

Second, the strategic one: "personal superintelligence" as MSL's stated goal since its founding "last year" places Muse as the intended vehicle for that mission from day one, not a pivot. If accurate, that means every Spark release's roadmap — model size, reasoning-tier structure (high/xhigh/max), even the contributor pricing tier that trades cheaper tokens for training data — was shaped by what a consumer personal agent would eventually need, not by what would top a coding leaderboard. That's a different mental model for anyone evaluating Spark purely against Opus 5 or GPT-5.6 Sol on coding benchmarks: those wins are a byproduct of Muse-oriented training, not the primary target.

How to access Muse Spark 1.3

Three supported paths, per Meta's cookbook and quickstart docs:

1. Muse Code — Meta's terminal-based, multi-agent coding tool, purpose-built for Spark models. Now out of beta as of September 1, 2026, with inter-session messaging and a workflows engine that composes naturally with 1.3's coding gains.
2. Meta Model API — direct, self-serve access, OpenAI SDK-compatible. Point your existing OpenAI client at Meta's endpoint and swap the model string; no new SDK to learn.
3. OpenRouter — for teams that want to try 1.3 inside tools they already use without touching their integration code.

The cookbook also documents several agentic patterns worth knowing about regardless of which access path you pick: multi-agent orchestration (a four-profile agentic team turning a one-line product idea into a working SaaS app), agent fan-out into isolated git worktrees for parallel work, computer use, an autonomous GitHub agent built on the Meta Model API plus OpenCode, and search grounding via a web tool for live information.

Practitioner takeaways

• If you're cost-conscious and your workload isn't sensitive, the contributor tier is a legitimate default, not just a teaser. At $0.10/M input and $0.20/M output with a 1M context window, it's among the cheapest frontier-adjacent coding access available anywhere right now.
• Route anything proprietary through the non-contributor tier. The 12.5-21x premium is the actual, named cost of keeping your code and data out of Meta's training pipeline — budget for it the same way you'd budget for a private-cloud SKU elsewhere.
• For long-horizon agentic coding specifically, Spark 1.3 is now a legitimate Opus 5 alternative, not just a cheaper also-ran. The DeepSWE v1.1 and Terminal-Bench 2.1 numbers are Meta's own published table, not a vendor's cherry-picked demo — and they hold up against the two other frontier options in the same chart.
• Don't overweight the general-agent numbers. Opus 5 still leads GDPVal-AA v2, JobBench, OSWorld 2.0, and AutomationBench. If your workload is broad computer-use or knowledge-work automation rather than coding specifically, the calculus is different.
• Watch for the open-weights announcement. A self-hostable Spark model would materially change the calculus for teams currently choosing between API-hosted Spark and Meta's existing open line — see explainx.ai's Llama 4 open-source guide for how Meta's open and hosted tracks have diverged so far.

Related reading

• Meta One: subscription tiers bundling AI features up to $499/month — Sept 2026; how Meta is now monetizing Muse-powered generation directly
• Meta Launches Muse: Personal Agent With Sentinel Security Architecture
• Meta Muse Code: Terminal Coding Agent Powered by Muse Spark 1.2
• Muse Code Exits Beta: Workflows and Inter-Session Messaging
• Muse Spark 1.1: Meta Model API, 1M Context, Agentic Coding Upgrade
• Muse Spark and the "Personal Superintelligence" Product Thesis
• Gemini 3.8 Flash: Benchmarks and Pricing
• Fable 5 vs Grok 4.6 vs GPT-5.6 Sol vs Qwen3-8 Max Comparison
• Best AI Coding Subscription Under $20
• Meta Llama 4: Open-Source Models Guide
• Terminal-Bench 2.0: AI Agent Benchmark Evaluation
• Official source: research.meta.ai/blog/introducing-muse-spark-1-3

Benchmark figures, pricing, and access details in this post reflect Meta's official announcement as of September 3, 2026. Pricing tiers and model availability may change — check Meta's Model API documentation for current rates before deploying.