Meta shipped Muse Spark 1.3 on September 2, 2026, one day after Anthropic’s Claude Fable 5.1 and on the same day as Google’s Gemini 3.8 Flash, three frontier-class releases in two days. The model is available immediately inside Muse Code, Meta’s coding agent, and through the Meta Model API as a drop-in upgrade: change the model ID, and endpoints, SDKs, and prices stay the same.
Meta’s announcement pitches 1.3 for the agentic builds developers actually ship, meaning long-running, multi-agent workflows where a model has to track context across steps, cope with messy or conflicting inputs, and ask for help instead of guessing. Independent testing supports the narrower claim that the model is broadly frontier-competitive: the version customers can use today ties GPT-5.6 Sol, Grok 4.6, and Claude Opus 5 on Artificial Analysis’s Intelligence Index at roughly half their cost per task on that index, though Meta’s advertised advantages in sustained multi-agent production work have not yet been independently tested. The catch sits in the pricing table, where the cheapest rate, $0.10 per million input tokens, comes with permission for Meta to use your prompts and outputs to improve its models, and with lower service limits. (Individuals named in this article are referenced for news context only; they are not affiliated with Coursiv and do not endorse it.)
What Meta Released
Muse Spark 1.3 is a proprietary, closed-weights multimodal reasoning model that accepts text, images, and video across a context window of about one million tokens. Meta’s previously available reasoning settings, including xhigh, work today; the max reasoning setting, used for the entire Muse Spark 1.3 column in Meta’s own scorecard, is in limited preview for partners and, in Meta’s words, coming shortly after additional safety testing.
What Muse Code is
Muse Code, the other half of the announcement, is Meta’s terminal-based coding agent, introduced in beta alongside Muse Spark 1.2 on August 5, 2026, which makes 1.3 a four-week follow-up. It plans changes, writes code, and validates results across large repositories, running persistent subagents in isolated copies of a repository. Version 1.3 is the model underneath it; Meta did not announce corresponding changes to the Muse Code runtime.
What changed from 1.2
Meta trained 1.3 for long-horizon coding and better instruction following, with the explicit goal of fewer unnecessary turns and cleaner outputs. The company claims about 20 percent fewer tool calls and about 25 percent fewer tokens on coding tasks compared with 1.2. Behaviorally, Meta says the model asks clarifying questions when prompts are ambiguous, invokes help when stuck, confirms before consequential actions, and has a better sense of what it cannot do rather than hallucinating outcomes; it also claims stronger resistance to prompt injection and training across several agent frameworks so behavior depends less on any one tool. These are company claims, and they describe exactly the behaviors independent benchmarks do not yet measure.
Muse Spark 1.3 Benchmarks
Source: Meta’s published evaluation scorecard. Configurations and evaluation origins vary by row; these are not uniform independent reruns. The scorecard compares 1.3 at max effort with 1.2 at xhigh, and with GPT-5.6 Sol and Claude Opus 5 both at max. Per Meta’s methodology note, the competitor figures mix Meta’s own runs with official leaderboard results and providers’ self-reported numbers, and Meta cautions that rival models may not be optimized for its prompts, tools, or runtime.
| Benchmark | Muse Spark 1.3 (max) | Muse Spark 1.2 (xhigh) | GPT-5.6 Sol (max) | Claude Opus 5 (max) |
|---|---|---|---|---|
| Long-context retrieval, MRCR 256K to 512K | 98.5 | 66.3 | 91.5 | n/a |
| Long-context retrieval, MRCR 512K to 1M | 98.1 | 55.5 | 73.8 | n/a |
| DeepSWE v1.1 (software engineering) | 75.4 | 55.0 | 73.0 | 74.0 |
| Terminal-Bench 2.1 (agentic coding) | 88.8 | 82.9 | 88.8 | 86.6 |
| GDPval-AA v2 (knowledge work) | 1754 | 1615 | 1710 | 1824 |
| OSWorld 2.0 (computer use) | 66.9 | 47.6 | 62.7 | 68.3 |
| AutomationBench (business workflows) | 49.4 | 38.2 | 46.7 | 50.3 |
| JobBench | 64.9 | 61.6 | 45.4 | 65.7 |
| DeepSearchQA | 89.4 | 85.9 | 93.0 | 90.4 |
| Agentic IF Index (instruction following) | 57.8 | 46.2 | 60.5 | 59.1 |
Where it wins and where it does not
The pattern is unusually clean. Muse Spark 1.3 leads or ties on long-context retrieval and coding, with near-perfect scores on million-token retrieval where GPT-5.6 Sol drops to the seventies. On the broader agent benchmarks, knowledge work, computer use, business workflows, and instruction following, Claude Opus 5 leads on four of six and Sol leads on search and instruction following, mostly by margins of a point or two that may sit within evaluation noise, since Meta publishes no confidence intervals. One more distinction: MRCR measures whether a model can retrieve specified information from a long context, which is not the same as reasoning reliably across a million-token codebase.
The comparison caveat
Two things soften the table. First, the 1.3 column uses max effort, which customers cannot access yet, while the 1.2 column uses xhigh, so part of the generational jump reflects a higher reasoning tier rather than a better model. Second, Terminal-Bench 2.1 is the older version of that benchmark; Anthropic’s Fable 5.1 release this week reported on the much harder version 4.0, where every model scores far lower, so the two tables cannot be compared with each other.
The Independent Numbers
Artificial Analysis, the independent evaluator, scored the customer-available xhigh tier at 61 on its Intelligence Index, up from 57 for Muse Spark 1.2, tying GPT-5.6 Sol at max, Grok 4.6 at high, and Claude Opus 5 at high, and ranked the gated max tier at 62, sixth of more than 600 models tracked. The cost to complete the index was $0.55 per task for Muse Spark 1.3 against $0.95 for Sol and $0.94 for Grok 4.6.
The same evaluation surfaced regressions Meta’s scorecard does not show: long-context reasoning accuracy fell from 83 to 79 percent versus 1.2, a knowledge benchmark dropped three points for xhigh, and cost per task actually rose from $0.40 to $0.55 because the new model consumes about 57 percent more input tokens per task on agentic evaluations, with output tokens up only about 8 percent. Cheap per-token pricing and more expensive per-task behavior can coexist, and here they do.
Muse Spark 1.3 Pricing: Standard vs Contributor
| Tier | Input per million tokens | Cached input | Output per million tokens | Your data | Rate limit |
|---|---|---|---|---|---|
| Standard | $1.25 | $0.15 | $4.25 | Not used for training, per Meta’s tier description | 3,000 requests per minute |
| Contributor | $0.10 | $0.002 | $0.20 | Permission to train future Meta models on prompts and completions | 100 requests per minute |
Rates and limits are from Meta’s Model API pricing and rate-limit documentation.
Meta describes the contributor tier as the same model at heavily discounted pricing, in exchange for permission to train future models on your prompts and completions, alongside lower rate limits. Its documentation positions the tier for prototyping, testing integrations, and scaling experiments where training on your data is acceptable. Two cautions apply to both rows. A no-training description is not a no-retention or no-logging promise, so review Meta’s current terms on retention, abuse monitoring, and human review before sending anything sensitive. And because the tiers also differ in rate limits, the discount is not a clean valuation of your data; it is a bundle of trade-offs.
Why the contributor tier exists
Meta’s discount suggests that real-world developer interactions are valuable enough to subsidize, though the company has not disclosed how contributor data enters its training process or how much of the discount reflects data permission rather than lower service limits. The offer appears to be working: Meta AI chief Alexandr Wang told Axios that a meaningful double-digit percentage of developers choose the contributor option, and described the unchanged pricing as aggressive. The launch thread shows both reactions: one developer called it an epic model that was not worth sharing their data, while others asked whether the price is permanent. Meta has not said.
Availability and the Rollout Questions
Muse Spark 1.3 is live in Muse Code and the Meta Model API, and third-party routers listed it within hours. Launch replies raised questions Meta’s thread did not answer, including whether every Muse Code plan gets 1.3, why some users still saw version 1.1, and whether the model will reach Cursor; Meta had not explained those individual availability reports at publication time.
How It Compares to This Week’s Other Releases
Base API list prices only; caching, batch discounts, reasoning-token use, and promotional expiry all change real costs.
| Model | Input / output per million tokens (list) | Vendor positioning |
|---|---|---|
| Muse Spark 1.3 (standard) | $1.25 / $4.25 | Long-context and coding agents |
| Muse Spark 1.3 (contributor) | $0.10 / $0.20 | Same model, training permission required |
| Gemini 3.8 Flash (promo to Dec 31) | $0.75 / $3.75 | Google’s newly released Flash model |
| GPT-5.6 Sol (promo to late Nov) | $4 / $20 | OpenAI’s broad frontier model |
| Claude Fable 5.1 | $10 / $50 | Anthropic’s top model for long-running agentic work |
Three companies shipped in two days, and each drew a different line. Google competes with a new low-priced Flash model, Anthropic on endurance at premium prices, and Meta on a frontier-tied model at mid-range prices with an unusually steep contributor discount. For developers the practical consequence is that a frontier-competitive model now costs $0.55 per task on Artificial Analysis’s index, which is not the same as a real coding ticket or a production agent run but is a useful relative signal.
What It Means for You
If you build
Muse Spark 1.3 is worth testing wherever long context or coding dominates: repository-scale reasoning, document-heavy agents, multi-step coding tasks. Watch tokens per task, not price per token, because higher input-token use can erode the discount. Before submitting proprietary or personal data on any tier, review Meta’s retention, security, and contractual terms and follow your organization’s normal approval process. Use the contributor tier only for non-sensitive material you own or have permission to share for training purposes.
If you do not
The contributor tier makes explicit a bargain most AI products leave implicit: cheaper AI in exchange for permission to train on what you submit. The discount is not a precise price for that permission, since the tiers differ in other ways too, but it is a useful reference point the next time a product offers you a free tier.
The Skill Underneath the Model Wars
Three frontier releases in two days is the new pace, and no one can evaluate each one from scratch. The people who keep up are the ones who can read a benchmark table with its caveats, match a model to a task, and judge what a data trade is worth. Those habits transfer across every model that ships next week. Coursiv builds them with step-by-step guides, short daily lessons, and hands-on practice with AI tools, designed for busy people without a technical background. Check the official site for current course details and pricing.
What to Watch Next
Watch whether the max tier reaches general availability and whether its safety testing changes the scores. Watch for the token-efficiency question to be settled by independent per-task measurements on real coding workloads. Watch the contributor tier’s terms and adoption, and whether any other lab offers a comparable data-use discount.
The Bottom Line
Muse Spark 1.3 makes Meta a credible competitor for agentic and coding workloads. Its standard tier is comparatively inexpensive; its contributor tier is cheaper still because developers grant Meta permission to train on what they submit. The benchmark gains are real enough to justify testing, and not enough to replace it.