Meta Muse Spark 1.1 Review: Still the Fast One, No Longer the Cheap One

For four weeks the case for Meta Muse Spark 1.1 was arithmetic. $1.25 per million tokens in, $4.25 out, for a natively multimodal model that ran tools better than almost anything near its price. Then on August 5, 2026, Meta launched Muse Spark 1.2 alongside Muse Code — its first terminal coding agent — and attached something to the new model that 1.1 never got: a contributor tier at $0.10 in and $0.20 out. Twelve times cheaper on input, twenty-one times cheaper on output, in exchange for letting Meta train on your prompts. Muse Spark 1.1's price did not move a cent. Its position did.

That is the honest frame for reviewing this model now. Muse Spark 1.1 has not been deprecated, has not been repriced, and is still the faster and better-documented of Meta's two reasoning checkpoints — but it is no longer the cheapest way to rent a Meta model, and the agent Meta is pushing hardest does not run on it. This review covers what 1.1 is, what it is measurably good at, what the August launch changed, and the narrower set of jobs where it is still the right call. Where a number comes from Meta's own evaluations, the sentence says so; where it comes from Artificial Analysis or from production traffic, it says that too.

Quick verdict

• What it is — a natively multimodal reasoning model (text, image, video, PDF and audio in, text out) with configurable reasoning effort, built to run tools and operate a computer. Closed weights, served by Meta.

• Price — $1.25 input / $4.25 output per million tokens, $0.15 on a cache hit. Unchanged since launch, and still roughly a quarter of GPT-5.6 Sol's output price ($5/$30) and a sixth of Claude Opus 4.8's ($5/$25).

• Intelligence — 51 on the Artificial Analysis Intelligence Index, rank #22 of 185 in its class against a class median of 32. An independent measurement, not a Meta claim.

• Strength — tool use and agentic reasoning, plus genuine speed: 213.5 output tokens per second, which Artificial Analysis ranks #2 of 185.

• Weakness — pure reasoning and raw coding are mid-pack, long-context recall is not frontier, and it is verbose: 94M output tokens to complete the Intelligence Index against a 65M median.

• The new caveat — Muse Spark 1.2 now scores three points higher and, on its contributor tier, costs a twentieth as much. 1.1's remaining edge is latency, and it is a big one.

What Meta shipped on August 5 — and what it did to 1.1

Muse Code is a terminal coding agent, currently in beta, that installs from a one-line shell script on macOS and Linux (no Windows build) and runs as the muse binary. It takes complete software-engineering tasks across large repositories: planning the change, writing the code, validating the result. Meta's pitch is architectural cohesion — the agent and the model were trained together rather than bolted together — and the feature list is aimed squarely at Claude Code and Codex: background agents that keep working after you close the terminal, event-log resume, parallel sub-agent worktrees, and an OS sandbox with approvals enabled by default. The demo Meta published for it is a GPU kernel optimisation loop running more than 1,000 tool calls over as much as 24 hours on NVIDIA Hopper hardware.

Muse Code runs Muse Spark 1.2, not 1.1. That is the first practical consequence for anyone reading this review: if you want Meta's agent, you are on the new checkpoint.

The second consequence is pricing, and it is the more interesting one. Meta shipped 1.2 with two distinct model IDs rather than one:

• Standard (muse-spark-1.2) — $1.25 input, $0.15 cached input, $4.25 output per million tokens. Identical to Muse Spark 1.1. Meta commits that prompts and completions on this tier are not used to improve its products.

• Contributor (muse-spark-1.2-contributor) — $0.10 input, $0.002 cached input, $0.20 output. In exchange, you grant Meta permission to train future models on your prompts and completions.

• Muse Spark 1.1 — no contributor tier exists. It stays at standard pricing, and Meta announced no deprecation.

Meta's own framing for the contributor tier is that it is "more than 10 times cheaper than even the pay-as-you-go tier," which understates it: the real multiples are 12.5× on input and 21.25× on output, with cached input at a near-free $0.002. That is the "low price" headline the launch generated, and it belongs to 1.2 alone.

The price story, rewritten

Before you rewrite your budget around $0.20 output tokens, read the rest of the contributor tier's terms, because they are what makes it cheap. Rate limits drop from a documented 3,000 requests per minute on the standard tier to 60 on the contributor tier — a cap that permits a developer experimenting on a laptop and forbids a production agent fleet. And the data trade is not a formality: your prompts and your model's completions become training material. For anyone handling customer data, proprietary source, or anything under a confidentiality obligation, the contributor tier is not a cheaper version of the same product; it is a different product. Meta acknowledged as much at the same launch by adding a zero-data-retention option for enterprise customers on the paid path.

So the honest comparison is not "1.1 at $4.25 versus 1.2 at $0.20." It is: at standard pricing, the two models cost exactly the same, and 1.2 is the better-scoring one. The $0.20 tier is a real and aggressive offer, but it is aimed at individuals and experiments, and it is only available on the newer model.

What actually drives the bill on either model is caching, not the sticker rate. Take a representative agent step: 60,000 tokens of repository or document context in, 3,000 tokens of plan-and-answer out. Cold, that is $0.075 of input plus $0.013 of output — about 8.8 cents, or $88 across a thousand steps. Keep the context prefix stable so it hits the cache at $0.15 per million and input collapses to $0.009, putting the step at roughly 2.2 cents and the thousand-step run at $22. A 75% swing, controlled entirely by whether your agent framework reuses a stable prefix or rebuilds the prompt every turn. If you take one engineering action after reading this review, make it cache-key stability rather than model selection.

One structural note on where you rent it. Muse Spark 1.1 is in the OrcaRouter catalog at $1.25 and $4.25 with a $0.15 cache read — the same numbers Meta charges, because OrcaRouter runs 0% markup and passes provider list price straight through, so a vendor price change lands here the day it lands there. Muse Spark 1.2 is not in the catalog yet and is served by Meta directly. That matters for the comparison you are probably about to run: GPT-5.6 Sol, Claude Opus 4.8, Grok 4.5 and Gemini 3.1 Pro all sit behind one key at list price, so you can benchmark them against Muse Spark 1.1 on a single evaluation set without a second contract, and the number in your spreadsheet is the number on the invoice.

What Muse Spark 1.1 actually is

Muse Spark is the flagship reasoning line from Meta Superintelligence Labs, the group Mark Zuckerberg stood up under Alexandr Wang. It marks a strategic reversal: where Meta's Llama models were open-weight, the Muse family — Spark, plus the Image and Video generators — is closed and delivered as a product and an API. Spark 1.0 arrived April 8, 2026; 1.1 followed on July 9 alongside the public preview of the Meta Model API. The practical change for developers is that Meta is now a first-party API vendor competing directly with OpenAI and Anthropic, not just a weights publisher — and the August 5 launch of a coding agent is the clearest confirmation of that yet.

Version 1.1 was a substantive step up rather than a point release. On Artificial Analysis it gained eight Intelligence Index points in three months, from 43 to 51. Its coding index rose 12 points to 71, SciCode improved to 58%, and Humanity's Last Exam rose to 45%. Meta said the biggest gains were in tool use, computer use, coding and multimodal understanding — consistent with a model tuned to act rather than merely answer.

On context, the earlier confusion has settled: Artificial Analysis and OrcaRouter both list the window at 1,048,576 tokens, and Meta describes a window the model actively compacts. Holding a million tokens is not the same as recalling from them, though, and 1.1's long-context retrieval score of about 54.1 says the recall is ordinary. Treat the window as a capacity, not a guarantee.

Reasoning effort: Instant, Contemplating, and the dial underneath

On Meta's consumer surface, Muse Spark exposes two named modes: Instant answers immediately without extended reasoning, like a standard chat turn; Contemplating runs multiple agents in parallel using a "thought compression" technique borrowed from reinforcement learning, and Meta positions it against DeepMind Deep Think and GPT-5.4 Pro on demanding scientific and analytical work. Through the API the same capability appears as a configurable reasoning-effort parameter — Artificial Analysis publishes its scores at the model's xhigh setting, the most expensive one it exposes.

Treat that dial as a cost lever, not a quality slider. Parallel-agent reasoning burns far more tokens than a single pass, and the benchmark numbers you read were produced at maximum effort. Default to the low end for routine calls and reserve high effort for the cases where a wrong first answer is expensive.

How it scores: strong with tools, mid-pack without

The clearest way to read Muse Spark 1.1 is tools-on versus tools-off. With tools, it leads the agentic tests: MCP Atlas 88.1 against Claude Opus 4.8's 82.2, JobBench 54.7 against 48.4, Legal Agent Bench 20.0 against Grok 4.5's 12.9, and Humanity's Last Exam with tools 62.1 against 57.9. Without tools it is ordinary — plain Humanity's Last Exam lands around 45, well behind Gemini 3.1 Pro's roughly 77 on the same test.

On execution precision it also trails the leaders: OSWorld-Verified computer use 80.8 against Opus 4.8's 83.4, SWE-Bench Pro coding 61.5 against 69.2, DeepSWE 1.1 long-horizon coding 53.3 against GPT-5.5's 67.0, and BabyVision visual reasoning 76.3 against 83.6. Many of these figures are vendor-reported, several from Meta's own evaluations, and they run on different harnesses — read them as directional and re-test on your own tasks.

The independent picture from Artificial Analysis is more compact and more useful. Intelligence Index 51, rank #22 of 185, against a class median of 32. Speed 213.5 output tokens per second, rank #2 of 185 — this is a genuinely fast model. Price rank #31, cache-hit price $0.15 for an 88% discount. Verbosity 94M output tokens to get through the index against a 65M median, which is where a chunk of the real bill comes from. Total spend to evaluate it on the whole suite: $548.07.

One caveat worth carrying: Meta advertises text, image, video, audio and PDF input, but Artificial Analysis's technical-specification card for 1.1 records only text and image as evaluated. Both can be true — the API accepts more than the benchmark harness exercised — but nobody outside Meta has published results on a video or audio workload. If mixed-media analysis is why you are here, budget time to verify it yourself.

Should you move to 1.2? The latency answer

On coding, Meta says yes, and its numbers are internally consistent: Terminal-Bench 2.1 rising from 76.2% to 82.9%, DeepSWE v1.1 from 53.0% to 59.3%, and its internal coding benchmark from 68.3% to 70.6%. These are Meta-run evaluations, scored pass@1 over five attempts in Meta's own harness — the standard caveat applies, which is that rival models in a vendor's harness are frequently under-tuned. Independently, Artificial Analysis moved Muse Spark 1.2 to 54 on the Intelligence Index, rank #13 of 185, three points above 1.1.

What Meta bought those points with is deliberation, and it shows up in every timing measurement. Artificial Analysis clocks time to first token at 26.12 seconds for 1.2 against 2.90 seconds for 1.1 — both at each model's highest reasoning effort. Output speed fell from 213.5 to 165 tokens per second. Verbosity rose slightly, 95M against 94M tokens, and the cost of running the identical benchmark suite rose from $548.07 to $637.85 on identical per-token pricing. That last figure is the cleanest statement of the trade: same price list, 16% more tokens burned, three more index points.

Read the 26-second figure carefully, because it is easy to misread. It is a maximum-effort benchmark measurement, and the API defaults to a middle setting. For a real-traffic reference point, Muse Spark 1.1 on OrcaRouter — serving whatever effort callers actually request — shows a p50 time to first token of 1.84 seconds and a p95 of 6.00 seconds across a week of production traffic, with a zero error rate. Benchmark latency at maximum effort and production latency at default effort are different quantities, and the gap between them is usually more than an order of magnitude.

So the decision resolves cleanly by workload shape. If you are running long-horizon coding agents that hold a repository in context and work for minutes at a time, 1.2 is the better model and the latency penalty amortises over a task that was never going to feel instant. If you are serving anything interactive — a chat surface, a tool-calling loop with a human waiting, a latency SLO measured in seconds — Muse Spark 1.1 remains the better-shaped model in the family, and its speed rank is not a rounding error. Nothing about the August launch changes that.

Developer experience and limits

The Meta Model API remains a public preview, though the August launch expanded global access and added the enterprise zero-data-retention option. Meta ships an OpenAI-style interface and SDK access, and Spark is live to consumers in Meta AI's "Thinking" mode. Rate limits are now published rather than guessed at — 3,000 requests per minute documented for the standard tier — but preview status is still preview status, so keep a fallback route for the days capacity is constrained. Automatic failover across providers is exactly the kind of infrastructure a preview-grade endpoint argues for, and it is the reason routing a preview model through a gateway costs you nothing and buys you a second path.

Through OrcaRouter the model ID is meta/muse-spark-1.1 and both /v1/chat/completions and /v1/responses are available, so an existing OpenAI client needs a base URL and a model string and nothing else:

from openai import OpenAI

client = OpenAI(base_url="https://api.orcarouter.ai/v1", api_key="YOUR_API_KEY")

response = client.chat.completions.create(model="meta/muse-spark-1.1", messages=[{"role": "user", "content": "Plan and run the tools to fix this issue."}])

print(response.choices[0].message.content)

Where it fits — and where it does not

Good fit: tool-running and computer-use automation where response time is visible to a person; cost-sensitive reasoning at scale on data you cannot hand to a training set; workflows mixing text with images, video, PDFs or audio, with your own verification on the media path; and teams who want a fast, cheap default with a premium model held in reserve for the hardest steps.

Weak fit: top-of-leaderboard coding accuracy and the hardest greenfield engineering, which still belong to Claude Opus 4.8 and GPT-5.6 Sol; pure-reasoning problems without tools; visual-precision tasks where Gemini 3.1 Pro leads; long-horizon repository work, which is now explicitly 1.2 and Muse Code's job; and anything where you need Meta's cheapest possible token, because that tier does not exist on this model.

FAQ

Is Muse Spark 1.1 still available now that 1.2 has shipped?

Yes. Meta announced no deprecation and no repricing at the August 5 launch — 1.1 remains on the Meta Model API at $1.25 and $4.25 per million tokens, and it is in the OrcaRouter catalog at the same numbers. Coverage of the launch consistently describes it as continuing at existing API pricing. That said, Meta's engineering attention has visibly moved: the coding agent, the new benchmarks and the cheap tier all attach to 1.2.

Can I get Muse Spark 1.1 on the $0.10 contributor tier?

No. The contributor tier is a separate model ID on the newer checkpoint, muse-spark-1.2-contributor. There is no 1.1 equivalent, so the cheapest Meta reasoning tokens require moving versions — and accepting a 60-requests-per-minute cap plus permission for Meta to train on your prompts and completions.

Should I upgrade to Muse Spark 1.2?

If your workload is long-running coding or repository-scale agent work, yes: it scores higher on both Meta's own coding evaluations and Artificial Analysis's independent index, at the same standard price. If your workload is interactive or latency-sensitive, no: 1.2 trades roughly nine times the time to first token and 23% of the output speed for three index points, and its reasoning cannot be switched off. Both sit on the same API, so the switch is a model string either way — which is the cheapest possible A/B test and worth running on your own traffic before deciding.

Is Muse Spark 1.1 open source?

No, and that is the point of the line. Unlike Meta's Llama models, the Muse family is closed-weight and delivered as a product and an API. There are no weights on Hugging Face, no self-hosting path, and no third-party providers — Meta is the only party serving this model, which makes a routing layer with failover the practical answer to single-provider risk rather than a second vendor.

The verdict, one version later

Muse Spark 1.1 is a fast, cheap, genuinely multimodal agentic model that is strong at tool use and honestly mid-pack at pure coding and reasoning. Nothing measured about it got worse in August. What changed is the shape of the family around it: Meta now has a better-scoring model at the same standard price, a dramatically cheaper tier for developers willing to trade their data, and an agent product that runs neither on 1.1 nor on anything you can point at 1.1.

What is left is a narrower but real recommendation. If you need Meta's reasoning quality at a human-perceptible speed — a second to first token in production, 213 tokens per second after that — 1.1 is still the version to call, and the three index points 1.2 adds are not worth a nine-fold wait. If you do not need that speed, there is no longer much reason to stay. Either way it is one model string, so the honest advice is to run both on your own evaluation set for a day and let your latency budget decide.