Meta shipped Muse Spark 1.3 on 2 September 2026, four weeks after 1.2 landed on 5 August. There was no keynote and no launch blog post — as of 3 September, the ai.meta.com blog index still lists "Introducing Muse Spark 1.1" (9 July 2026) as the only Muse Spark announcement. 1.2 and 1.3 both arrived as documentation-and-API point releases.
That makes the upgrade question harder than usual: the two versions are indistinguishable on every number a pricing page can show you. Here is what actually changed, what independent measurement says so far, and the one concrete reason some teams should stay on 1.2.
What actually changed between Muse Spark 1.2 and 1.3?
On the published specification: almost nothing. Here is the head-to-head, taken from Meta's pricing and rate-limits page and the live OpenRouter model records for both SKUs. The identical rows are the finding.
| Dimension | Muse Spark 1.2 | Muse Spark 1.3 |
|---|---|---|
| Released | 5 Aug 2026 | 2 Sep 2026 |
| Model ID | muse-spark-1.2 | muse-spark-1.3 |
| Input price | $1.25 / 1M | $1.25 / 1M |
| Output price | $4.25 / 1M | $4.25 / 1M |
| Cached input | $0.15 / 1M | $0.15 / 1M |
| Context window | 1,048,576 tokens | 1,048,576 tokens |
| Max output | 943,718 tokens | 943,718 tokens |
| Input modalities | Text, image, video, audio, PDF | Text, image, video, audio*, PDF |
reasoning_effort levels | minimal → xhigh | minimal → xhigh |
| Web search grounding | $2.50 / 1k queries | $2.50 / 1k queries |
| Standard rate limit | 3,000 RPM / 4M TPM | 3,000 RPM / 4M TPM |
| Contributor SKU | Yes | Yes |
| Endpoints | Responses, Chat Completions, Messages | Responses, Chat Completions, Messages |
* See the audio caveat below — it is the only spec-level regression in 1.3.
Meta states the position plainly in its model docs: the three Muse Spark versions each share "the same modalities and context window, and differing only by capability." There is no long-context premium on either version, and no version-specific parameter to learn.
What does Meta say is different in 1.3?
Meta's Models page gives one sentence of claim, and it is worth reading closely. Muse Spark 1.3 is described as "the latest version, tuned for agentic workflows (multi-step tool, browser, and long-horizon tasks) with improved coding over 1.2," and is marked "Recommended for new work." It is also now the default model in every code sample across the documentation, which 1.2 previously was.
The OpenRouter descriptions track the same shift in emphasis. 1.2 is described as "a reasoning model designed for complex agentic tasks" that "supports multi-agent workflows, whether as either a main agent that plans and delegates or as a subagent executing in parallel." 1.3 is described as being "for long-running agentic, multi-agent, and coding workflows," designed "to keep track of information across extended tasks, work through conflicting inputs, and request clarification or confirmation when needed, with an emphasis on concise execution."
Two of those phrases are new and specific enough to be testable: work through conflicting inputs and request clarification or confirmation when needed. If the tuning is real, 1.3 should ask a question rather than guess when a long agent run hits contradictory context — the exact failure mode that makes long-horizon agents expensive. Meta has published no eval backing it, so treat it as a stated design goal, not a measured result.
What do independent benchmarks show for 1.3 vs 1.2?
A month on, the independent picture has filled in — and it is more interesting than the headline gain suggests. Note that all figures below are Artificial Analysis Intelligence Index v4.3.2, captured 5 October 2026. AA rebased the Index from v4.1.1 to v4.3.2 during September, swapping the eval basket (it now runs AA-Briefcase, AA Long-Context Reasoning, AutomationBench, CritPt, GDP-PDF, GDPval, Humanity's Last Exam, Omniscience, SciCode and Terminal-Bench 4.0). Every score on the Index moved as a result, so a v4.1.1 number and a v4.3.2 number are not comparable and there is no conversion between them. The rows below are all on the current basket.
| Measurement | Muse Spark 1.2 | Muse Spark 1.3 |
|---|---|---|
AA Intelligence Index v4.3.2 — xhigh | 39.58 | 45.07 |
AA Intelligence Index v4.3.2 — max | Not measured | 48.09 |
Cost per Index task — xhigh | $0.97 | $1.37 |
Cost per Index task — max | — | $1.60 |
| Cost to run the whole Index suite | $1,385.40 | $1,655.27 (xhigh) / $2,000.35 (max) |
Median output speed — xhigh | 270.1 tok/s | 186.6 tok/s |
Median output speed — max | — | 145.9 tok/s |
Median time to first token — xhigh | 11.7 s | 35.8 s |
Median time to first token — max | — | 47.9 s |
| Input modalities per AA | Text, image, speech, video | Text, image, video |
OpenRouter vgi_bench | 53.4% | 61.4% |
| Arena text leaderboard (3 Sep 2026) | #5 — 1499 ±10, 3,240 votes | Not listed at launch |
The capability gain is real, and it is bigger than a point release usually delivers. Compared at the same reasoning effort — xhigh against xhigh — Artificial Analysis puts 1.3 at 45.07 against 39.58 for 1.2 on Intelligence Index v4.3.2, a 5.5-point move. Turn 1.3 up to max, an effort tier AA has not published for 1.2, and it reaches 48.09 — 8.5 points above 1.2's best measured run, and within 2.7 points of Claude Opus 5 at max effort (50.78) for roughly a quarter of Opus 5's cost per task ($1.60 against $5.86). OpenRouter's vgi_bench moves the same direction, 53.4% to 61.4%.
What the fuller data set now shows is the price of that gain, and it is not money — the token rates are identical. It is speed. At matched xhigh effort, 1.3 outputs 186.6 tok/s against 1.2's 270.1, and waits 35.8 seconds before the first chunk against 1.2's 11.7. Push 1.3 to max and it drops to 145.9 tok/s with a 47.9-second time to first token — roughly half the throughput and four times the initial latency of 1.2. For a batch job that is irrelevant; for anything a human watches, it is the whole story.
The second cost is reasoning tokens. AA's cost-per-Index-task figures isolate it cleanly: at identical per-token prices, 1.3 costs $1.37 per Index task at xhigh against 1.2's $0.97 — 41% more for the same work — rising to $1.60 at max, 65% more. The published price did not change; the number of tokens the model emits to reach an answer did.
Two earlier rows have been removed rather than updated. AA retired its separate Coding Index and Agentic Index composites during the v4.3.2 rebase; the figures we previously quoted for 1.2 (72.2 and 49.3) were v4.1.1-era and have no current equivalent, so they are not convertible and we do not carry them forward. Human-preference data also remains thin: 1.2's #5 Arena placement rests on 3,240 votes, and 1.3 had no Arena listing at launch.
The honest read: 1.3 is the better model and it is not close, but it is not a strict upgrade. You are trading latency and reasoning tokens for accuracy at a fixed token price. Our earlier look at whether Muse Spark 1.2's benchmark claims held up against Claude Opus 5 is the pattern to expect again — headline indices move first, agentic numbers settle later.
Why should audio users stay on Muse Spark 1.2?
This is the one place where 1.3 is a downgrade, and it is documented by Meta rather than inferred. Both the Models page and the Video and audio understanding guide carry the same note:
Audio understanding in Muse Spark 1.3 is currently not fully supported, and response quality for requests including audio content may be degraded.
Meta's recommendation is to use Muse Spark 1.2 for audio, or the dedicated muse-voice-transcribe-1.0 model for speech-to-text. The audio guide follows its own advice — every audio example in it is written against muse-spark-1.2, not the 1.3 default used everywhere else in the docs.
Third parties have started encoding the caveat as a hard fact. Artificial Analysis's model pages now list Muse Spark 1.2's input modalities as "text, image, speech, and video" but 1.3's as "text, image, and video" — speech dropped entirely. OpenRouter lists no audio input modality for any Muse Spark version. Meta's own docs still show audio on 1.3 with the asterisk, so the registries are stricter than the vendor; either way, nobody is treating 1.3 as an audio model.
This applies to embedded audio in video too. Muse Spark reads a video's soundtrack alongside its frames, so a pipeline summarising meeting recordings or narrated screen-captures is an audio workload even if you think of it as video. The word "currently" implies a fix is intended; Meta has published no timeline.
What is the Contributor tier, and what do you trade for it?
Both 1.2 and 1.3 ship a second SKU — muse-spark-1.2-contributor and muse-spark-1.3-contributor — at a fraction of standard pricing.
| Per 1M tokens | Standard tier | Contributor tier |
|---|---|---|
| Cached input | $0.15 | $0.002 |
| Input | $1.25 | $0.10 |
| Output | $4.25 | $0.20 |
| Trains on your data | No | Yes |
| Requests per minute | 3,000 | 100 |
| Tokens per minute | 4,000,000 | 3,000,000 |
The trade is explicit in Meta's docs: Standard means "your prompts and completions are not used to train Meta models"; Contributor is "heavily discounted token pricing in exchange for permission to use your prompts and completions to train future Meta models." On a realistic 1M-in / 200k-out job that is $2.10 on Standard against $0.14 on Contributor — about 15× cheaper.
The catch most write-ups miss is the throughput ceiling. Contributor is capped at 100 requests per minute against Standard's 3,000 — a 30× reduction, applied per team rather than per API key. Contributor is built for prototyping and burst experiments, not for serving a chatty production agent, and no amount of extra keys will raise it.
The timing tells you something about Meta's intent. 1.2 shipped on 5 August and its Contributor SKU only appeared 16 days later, on 21 August. 1.3's Contributor SKU appeared 53 minutes after the base model, on the same day. Muse Spark 1.1 has no Contributor variant at all. The tier started as a bolt-on experiment and is now a standing part of the product line. One further detail: muse-spark-1.2-contributor carries the canonical build date 20260805 — the same snapshot as base 1.2, despite being listed two weeks later. Contributor is a billing-and-terms tier over identical weights, not a different or smaller model.
The full terms deserve more room than this comparison gives them — we cover what the data-sharing grant actually permits in what Contributor pricing really costs you, and the 1.3 SKU specifically in the Muse Spark 1.3 complete guide.
Is Muse Spark 1.2 deprecated or still available?
Still available, and not deprecated. Meta's Models page lists 1.3 as "the latest version," 1.2 as "the previous version" and 1.1 as "the original version" — three neutral labels with no sunset notice, no end-of-life date and no migration deadline attached to any of them. All three remain on the Standard tier at the same price.
Serving status backs that up. As of 5 October 2026 — a month after 1.3 shipped — both muse-spark-1.2 and muse-spark-1.2-contributor are still live on OpenRouter's model registry with no expiry date on the record. Meta has kept 1.1 addressable nearly two months after 1.2 replaced it — a reasonable prior for how long 1.2 stays reachable. Pin explicit version strings rather than relying on a default and you control your own timing.
Do you need to change any code to upgrade?
For nearly every application, the change is one string. The two versions share the same endpoints (Responses, Chat Completions and Messages), the same 1,048,576-token context, the same 943,718-token output ceiling, the same reasoning_effort ladder from minimal to xhigh, the same tool-calling and structured-output surface, and the same prices. OpenRouter reports byte-identical supported-parameter lists for both.
- model="muse-spark-1.2"
+ model="muse-spark-1.3"Three things to check before you flip it:
- Audio paths. If any request can contain an
input_audiopart — or a video with a soundtrack you rely on — keep those calls on 1.2 and route only text/image/PDF traffic to 1.3. - Output token budget. Artificial Analysis's per-task cost on its Intelligence Index v4.3.2 is $1.37 for 1.3 at
xhighversus $0.97 for 1.2, and $1.60 atmax— a 41–65% increase at identical per-token prices, purely from extra reasoning tokens. Re-baseline cost on your own traffic rather than assuming parity. - Latency budget. 1.3 is measurably slower: 186.6 tok/s and 35.8 s to first token at
xhigh, against 270.1 tok/s and 11.7 s for 1.2. If a user is waiting on the stream, test perceived latency before you flip anything interactive. - Prompt-cache warmth. Switching model IDs starts a fresh cache. Expect a short window paying $1.25 rather than $0.15 on prefixes that were previously hitting cache.
Neither version supports reasoning_effort: "none", so there is no thinking-disabled path on either — that has not changed.
Who should switch to Muse Spark 1.3 — and who should wait?
Switch now if you are starting new work (Meta's own recommendation), you run long-horizon or multi-agent loops where the "conflicting inputs" and "ask for confirmation" tuning targets your actual failure mode, or you are coding-heavy and want the measured 5.5-to-8.5-point index gain at zero extra list price. With prices identical, the downside is bounded: if 1.3 underperforms for you, reverting is the same one-line change.
Wait if audio is anywhere in your pipeline — that is a documented regression, not a rumour. Also wait if latency is user-visible — 1.3's 35.8-second time to first token at xhigh is three times 1.2's, and its throughput is about 30% lower — or if you are cost-sensitive on output-heavy workloads until you have measured 1.3's extra reasoning tokens against your own prompts. Interactive chat and streaming UIs are the clearest case for staying on 1.2.
The one thing not worth doing is agonising over it. This is not a pricing decision or an architecture decision — it is a model string, both versions stay served, and a two-hour A/B on your own eval set will tell you more than any leaderboard will. If you want the wider context on where Muse Spark sits against the frontier, start with our Muse Spark guide and the 1.3 vs Claude Opus 5 comparison. And if you are staffing the engineers who have to run these evaluations properly, that is the kind of work Codersera places for.
FAQ
What's new in Muse Spark 1.3?
Meta describes 1.3 as tuned for agentic workflows — multi-step tool, browser and long-horizon tasks — with improved coding over 1.2. OpenRouter's description adds that it tracks information across extended tasks, works through conflicting inputs, and asks for clarification when needed. Price, 1M context, max output and the parameter surface are all unchanged from 1.2.
Is Muse Spark 1.3 better than 1.2?
On accuracy, clearly yes. Artificial Analysis's Intelligence Index v4.3.2 (October 2026) scores 1.3 at 45.07 against 39.58 for 1.2 at matched xhigh effort, rising to 48.09 at max; OpenRouter's vgi_bench shows 61.4% against 53.4%. On speed, no: 1.3 runs at 186.6 tok/s against 1.2's 270.1 and takes 35.8 seconds to first token against 11.7. It is also documented as worse at audio.
Is Muse Spark 1.3 more expensive?
No. Both versions are $1.25 per million input tokens, $4.25 per million output, and $0.15 per million cached input on Meta's Standard tier, with web-search grounding at $2.50 per 1,000 queries on either. There is no long-context premium on either version. Your bill can still rise, though: on Artificial Analysis's Intelligence Index v4.3.2, 1.3 cost $1.37 per task at xhigh against 1.2's $0.97 — 41% more at identical token rates, entirely from extra reasoning tokens.
What is the Contributor tier?
A discounted SKU — muse-spark-1.3-contributor — priced at $0.10/1M input and $0.20/1M output in exchange for permission to train future Meta models on your prompts and completions. Standard tier never trains on your data. Contributor is also rate-limited to 100 requests per minute against Standard's 3,000, so it suits prototyping rather than production serving.
Is Muse Spark 1.2 deprecated?
No. Meta's docs label it "the previous version" with no sunset date, end-of-life notice or migration deadline. Both the standard and Contributor 1.2 SKUs were serving normally with full uptime as of 3 September 2026, and 1.1 is still addressable two months after being superseded. Meta recommends 1.3 for new work but does not require you to move.
Do I need to change my code to upgrade?
For most applications, only the model string. The two versions share endpoints, context window, max output, reasoning-effort levels, tool-calling and structured-output support, and pricing. The exception is audio: keep input_audio requests and audio-bearing video on 1.2. Also expect a cold prompt cache immediately after switching model IDs.
Does Muse Spark 1.3 support audio?
It accepts audio input, but Meta states audio understanding "is currently not fully supported" and response quality "may be degraded." The official recommendation is Muse Spark 1.2 for audio understanding, or the dedicated muse-voice-transcribe-1.0 model for speech-to-text. Meta's own audio documentation examples use 1.2. This includes the embedded soundtrack of an uploaded video.
Why did Meta not publish a blog post for Muse Spark 1.3?
Meta announced Muse Spark 1.1 with a blog post on 9 July 2026, but shipped both 1.2 and 1.3 as documentation-and-API point releases. As of 3 September 2026, the ai.meta.com blog index carries no 1.2 or 1.3 announcement. The authoritative sources for what changed are Meta's Models and Pricing pages, plus the live model registry.