Muse Ember Scores 36, Costs $0.13 a Task [2026]

Meta AI released Muse Ember on November 9, 2026, a model it calls its smallest and most efficient, priced at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200K input tokens. Meta's developer blog reports a score of 36 on the Artificial Analysis Intelligence Index for Muse Ember, which carries a 1,048,576-token context window, a 524,288-token maximum output, and a default reasoning effort of max, and is available through the Model API and Muse Code. Meta has not released parameter counts or model weights for the model, which accepts text, image, video, audio and PDF input and returns text.

Meta AI announced Muse Ember on November 9, 2026, and billed it as the company’s smallest and most efficient model. In its own words, the model is “our smallest and most efficient model designed for complex reasoning execution and the work agents repeat all day: tool calls, coding subtasks, routing, extraction, and long document parsing.” Coverage from TechCrunch, MarkTechPost, The Economic Times, Zee Business and Times Now confirms the launch, and third-party reports credit Meta Superintelligence Lab with the release. The numbers behind the pitch are blunt. Muse Ember costs $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 200K input tokens, it accepts a 1,048,576-token context window, and Meta’s developer blog gives it a score of 36 on the Artificial Analysis Intelligence Index. We covered the launch pricing in our earlier Muse Ember price report https://shattered.io/meta-muse-ember-agent-execution-model-0-10-per-million-tokens-2026/ . This piece asks a different question: what do those figures mean for a team that has to budget an agent fleet? Everything in the spec and price sections comes from Meta AI’s announcement and developer blog. The cost scenarios further down are our own arithmetic on Meta’s published rates, and we label them that way. Meta has not released parameter counts or model weights, so nothing here speculates about the architecture. Don't miss new tech stories on Google Add Shattered once in the Google app and our stories show up in your news suggestions. Add now https://www.google.com/preferences/source?hl=en&q=https%3A%2F%2Fshattered.io%2F Muse Ember Specs at a Glance The spec sheet is short, which suits a model built to sit inside loops. Muse Ember takes text, image, video, audio and PDF input and returns text. That input range matters for the document-parsing and extraction jobs Meta names in its description, because a single model call can read a scanned PDF without a separate OCR or transcription step in front of it. | Spec | Muse Ember per Meta | |---|---| | Release date | November 9, 2026 | | Releasing organization | Meta AI official post , Meta Superintelligence Lab third-party reports | | Input modalities | Text, image, video, audio, PDF | | Output | Text | | Context window | 1,048,576 tokens | | Maximum output | 524,288 tokens | | Model ID | muse-ember-1.0 | | Reasoning effort | minimal through max, default max | | Availability | Model API and Muse Code | Two details stand out. First, the output ceiling of 524,288 tokens is exactly half the context window, which leaves room for very long generated reports or code files. Second, the default reasoning effort is max, not a middle setting. Developers who copy a quickstart snippet and never touch the parameter will run the most expensive thinking mode every time. Why a Small Model Matters for Agents Agent software repeats the same few operations thousands of times. A coding agent reads a file, proposes a patch, calls a linter, reads the result and tries again. A support agent routes a ticket, extracts an order number, looks up a record and drafts a reply. Each step is small, but the steps stack up fast, and the bill scales with the number of calls rather than the difficulty of any single one. Meta’s description targets exactly that pattern. It lists tool calls, coding subtasks, routing, extraction and long document parsing, which are the high-frequency, low-glamour tasks in an agent loop. The pitch is not that Muse Ember beats the largest frontier models at hard reasoning. The pitch is that it handles the repetitive middle of the workflow at a price that lets teams stop rationing calls. Teams building multi-tool agents already feel this pressure. Our walkthrough on pooling MCP servers in OpenCode https://shattered.io/opencode-mcp-servers-pooling-12-steps-2026/ shows how quickly tool calls multiply once an agent has more than a handful of connected services. A cheap, fast model for those calls changes the arithmetic of the whole design. Muse Ember Pricing, Tier by Tier Meta publishes two pricing tiers, split at 200K input tokens. For prompts up to 200K input tokens, the rate is $0.10 per million input tokens and $0.50 per million output tokens. For prompts above 200K input tokens, the rate rises to $0.30 per million input tokens and $1.50 per million output tokens. Meta also lists a cached-input price of $0.004 per million tokens for the shorter-context tier. The cached rate is the most interesting line. At $0.004 against $0.10, cached input costs 25 times less than fresh input. Agents that resend a long system prompt, a tool catalog or a large reference document on every turn stand to gain the most. A loop that replays a 30,000-token preamble on each call would see that portion of the bill nearly vanish, provided the caching behavior matches what Meta’s pricing page implies. The long-context tier deserves a careful read. Every rate triples once a prompt crosses 200K input tokens. Our sources do not spell out whether the higher rate applies to the whole request or only to the tokens past the threshold. Check Meta’s API documentation before you size a budget around very long prompts, because the difference changes the cost of a 250K-token request by a wide margin. Cost Math: What Real Agent Jobs Would Cost Published per-million rates are hard to feel. The table below converts them into per-job costs for six agent tasks. These are our calculations using Meta’s list prices, the token counts are illustrative assumptions, and none of the figures come from Meta. They ignore reasoning tokens, retries and any volume terms, so treat them as a floor rather than a forecast. | Job illustrative | Input tokens | Output tokens | Tier | Estimated cost | |---|---|---|---|---| | Routing call | 2,000 | 200 | Up to 200K | $0.0003 | | Coding subtask, no caching | 40,000 | 3,000 | Up to 200K | $0.0055 | | Coding subtask, 30,000 input tokens cached | 40,000 | 3,000 | Up to 200K | about $0.0026 | | Long document extraction | 150,000 | 4,000 | Up to 200K | $0.017 | | Very long document | 600,000 | 8,000 | Above 200K | $0.192 | | Maximum-length generation | 1,000 | 524,288 | Up to 200K | about $0.26 | Routing and extraction at scale A routing call of 2,000 input tokens and 200 output tokens costs about three hundredths of a cent. At those rates, one million routing decisions run roughly $300. That is the territory where a team can put a model in front of every inbound message instead of running keyword rules and sending only the hard cases to a larger model. Extraction jobs scale the same way. Parsing a 150,000-token document into structured fields costs about 1.7 cents, so 10,000 such documents come to roughly $170. The calculation assumes the model gets the extraction right, which is the part Meta’s single benchmark figure cannot confirm for any specific workload. Where the bill jumps The very long document row shows the tier effect. A 600,000-token input at the higher rate costs about $0.19 per job, which is more than ten times the cost of the 150,000-token extraction even though the input is only four times larger. Teams that can chunk documents below the 200K line, or cache the shared portion, will pay far less than teams that feed the full window every time. The maximum-output row is a useful ceiling. A single response that fills all 524,288 output tokens costs about 26 cents on the shorter tier and about 79 cents at the higher output rate. Few real jobs approach that length, but a runaway agent loop that keeps generating can. A hard cap on output tokens belongs in every production deployment. The $0.13 Per Task Benchmark Figure Meta’s developer blog reports an estimated cost of about $0.13 per Intelligence Index task at maximum reasoning effort. That figure includes the cost of reasoning tokens, which the model generates before it answers. It tells you more about real spend than the per-million rates do, because reasoning models can burn far more output tokens than the visible answer suggests. The number also sits far above our routing example. A task in the index is a hard problem, not a routing decision, so a 13-cent average implies long reasoning traces. Run 1,000 such tasks and the bill reaches about $130. Run 100,000 and it reaches about $13,000. That is cheap next to a large frontier model, but it is not free, and it argues for dialing the reasoning effort down on easy work. The gap between the two numbers is the central lesson for budgeters. Per-token prices set the floor, and reasoning effort sets the multiplier. A team that leaves the default at max for a classification job pays for deliberation the task never needed. Reading the 36 Score on the Intelligence Index Muse Ember scored 36 on the Artificial Analysis Intelligence Index, according to Meta’s developer blog. A single score does not say much without peers beside it, and our sources do not provide independent numbers for rival models, so we will not invent a leaderboard. What the figure does show is that Meta chose to publish an index result alongside a cost estimate, which frames the release as a price-performance story rather than a raw capability story. That framing fits a small model. Nobody buys a budget agent model to write the best essay or solve the hardest proof. They buy it to hit an acceptable quality bar at the lowest cost per completed task. The score of 36 is Meta’s evidence for the quality side, and the $0.13 estimate is its evidence for the cost side. Readers should note one limit. The index result comes from Meta’s own developer blog, so it is a company-reported figure until Artificial Analysis or other testers publish their own runs. Independent replication usually lands within a few points, but it can move the story either way. Meta’s Claim Against Claude Haiku 5.5 and GPT-6 Luna Meta’s headline comparison is aimed at two small-tier rivals. The company says Muse Ember delivers “more intelligence per dollar” than Claude Haiku 5.5 and GPT-6 Luna “at any reasoning setting.” That is a company claim, not an independently established fact, and Meta has not published the full methodology in the material we reviewed. The wording is careful in a useful way. It says intelligence per dollar, not raw intelligence. A model can lose on absolute quality and still win that contest if its price is low enough. The phrase “at any reasoning setting” widens the claim, since it covers cheap low-effort runs and expensive max-effort runs alike. Proving it takes a head-to-head test across every effort level, and that is the test we expect independent evaluators to run first. Buyers can run their own version of that test. Our guide to a local LLM-as-a-judge evaluation harness https://shattered.io/llm-as-a-judge-local-eval-harness-12-steps-2026/ walks through building a repeatable scorecard on your own prompts. Run Muse Ember, Claude Haiku 5.5 and GPT-6 Luna through the same 200 real tasks, record cost and pass rate, and you will have an answer that matters more than any vendor chart. Reasoning Effort: Minimal to Max Muse Ember exposes a reasoning effort control that runs from minimal through max, with max as the default. This control is the main cost lever beyond the price card. Low effort should cut reasoning tokens sharply on simple jobs, while high effort buys accuracy on harder ones. Meta’s claim of better intelligence per dollar “at any reasoning setting” implies the company expects developers to move the dial. A sensible deployment pattern follows from the design. Send routing, classification and field extraction at minimal or low effort. Reserve max for multi-step coding subtasks and ambiguous document questions. Log the token counts per effort level for a week, then compare the cost per accepted result rather than the cost per call. Teams that skip the logging step often discover that their cheapest setting is not the setting with the lowest cost per success. The max default also has a safety angle for budgets. A developer who prototypes at default settings and then ships without changing them carries prototype-grade reasoning costs into production. Set the effort explicitly in code, not in a config someone forgets. Where to Use It: Model API and Muse Code Muse Ember is available through Meta’s Model API and inside Muse Code. The standard model ID is muse-ember-1.0, which developers will pass in API requests. A minimal request shape looks like the sketch below. It is illustrative only, so confirm the exact field names against Meta’s API reference before you use it. { "model": "muse-ember-1.0", "reasoning effort": "minimal", "max output tokens": 2048, "input": "Extract the invoice number and total from the attached PDF." } Muse Code is Meta’s coding product, and the cited official page lists subscription tiers for it. The page does not establish a subscription price specific to Muse Ember, so we cannot say what a Muse Code user pays to access the model. Anyone who wants per-token billing should use the Model API, where the rates above apply. The split between the two surfaces mirrors a familiar pattern. Individual developers who want an assistant in their editor choose the subscription product, and platform teams who want to control spend per request choose the API. Our comparison of self-hosted LLMs versus API pricing https://shattered.io/self-hosted-llm-vs-api-kimi-k3-1-pro-glm-5-3-deepseek-v4-1-mistral-large-4-2026/ covers the broader trade-off, including the point where running your own weights beats paying per token. That option stays closed for Muse Ember today, since Meta has not released weights. What Meta Has Not Confirmed A clean launch can hide gaps. The table below separates what the sources establish from what they leave open, so readers do not mistake silence for a fact. | Question | Status | What the sources say | |---|---|---| | Parameter count | Not confirmed | Reporting says Meta has not disclosed it | | Open weights | Not confirmed | Reporting says Meta has not released weights | | Muse Ember subscription price | Not confirmed | The official page lists Muse Code tiers but no Muse Ember-specific price | | Beating Claude Haiku 5.5 and GPT-6 Luna | Company claim | Meta says it wins on intelligence per dollar, no independent test cited | | Intelligence Index score of 36 | Company-reported | Published on Meta’s developer blog | | API list prices | Published | $0.10 and $0.50 per million tokens up to 200K input | The missing parameter count matters more than it looks. Without it, outsiders cannot judge whether the low price reflects a genuinely small model, aggressive distillation, subsidized pricing or all three. Subsidized pricing in particular can change after a launch window, which is why long contracts built on day-one rates carry risk. Market Impact on Small-Model Pricing A $0.10 input price sets a new reference point for the budget tier from a company with Meta’s distribution reach. Small-model price cuts rarely stay isolated. When one lab publishes a low price and a self-reported efficiency lead, rivals face a choice between matching the price, matching the claim with their own benchmarks, or pointing to quality differences the index misses. We expect the competition to play out on three fronts. The first is list price per million tokens. The second is cost per completed task, where reasoning-token usage matters as much as the rate. The third is cached-input pricing, where Meta’s $0.004 rate leaves little room for rivals to undercut without hurting their own margins. Developers who read only the first number will miss the other two. The pricing pressure is already visible across the sector. Our earlier look at DeepSeek API pricing against Mistral Large 4 https://shattered.io/deepseek-api-pricing-vs-mistral-large-4-0-60-vs-2-10-2026/ showed how wide the gap between low-cost and premium API rates has become. Muse Ember adds a large Western lab to the low end of that spread, which narrows the argument that only Chinese open-weight providers can offer rock-bottom rates. Historical Context: How Small Models Became Agent Workhorses Frontier labs used to treat their small tiers as afterthoughts, a cheaper option for chatbots and autocomplete. The rise of agents reversed that. Once software started calling models in loops, the cheapest competent model became the one handling most of the traffic, and the largest model became the escalation path for hard cases. Each major lab now sells a compact tier built around that role. Anthropic’s Haiku line and OpenAI’s Luna tier are the two rivals Meta named in its own comparison. Our coverage of Anthropic’s Fable 5.2 and Mythos 5.2 launch https://shattered.io/anthropic-fable-5-2-mythos-5-2-30-percent-faster-8-40-pricing-2026/ shows the same dynamic from the other direction, with speed and price positioned as headline features rather than footnotes. Meta’s entry fits that history. Muse Ember carries the Muse brand into the cheapest slot, with a long context window and multimodal input that earlier budget tiers often lacked. The 1,048,576-token window on a small model is the part that would have sounded unusual a few years ago, and it is the feature that makes the long document parsing claim plausible. Five Predictions for the Next 90 Days These are our forecasts, not reported facts, and we have grounded each one in the published numbers. - Independent index runs will land. Artificial Analysis and other evaluators typically test new releases within weeks. Expect third-party results that either confirm the 36 score or move it by a few points. - Rivals will answer on cached pricing. The 25x gap between fresh and cached input is the easiest lever to match, so expect at least one competing small tier to cut its cached-input rate. - Developers will publish effort-level cost tables. Because the default is max, early adopters will post measurements showing how much dropping to minimal saves on routing and extraction. - Meta will face pressure on weights and size. The undisclosed parameter count and absent weights invite questions from enterprise buyers, and a partial disclosure is plausible. - Long-context abuse will show up in bills. Teams that stuff the full 1,048,576-token window into every call will hit the higher tier and report surprise invoices, which will push the chunk-and-cache pattern into standard practice. What Builders Should Do This Week Start with a measurement, not a migration. Pick one repetitive agent job, such as ticket routing or invoice extraction, and run 200 real examples through Muse Ember at two effort levels. Record accepted results, total tokens and total cost. Do the same against whichever small model you use today. Then set the guardrails Meta’s defaults will not set for you. Pin the reasoning effort in code, cap output tokens per request, and split documents so most calls stay under the 200K input line. Structure prompts so the shared preamble is cacheable. Those four habits do more for your invoice than any choice between vendors. Finally, avoid long-term commitments built on launch pricing. Meta has not disclosed model size, and a price this low from a new release can shift. Keep your integration thin enough to swap models, and keep following our AI and machine learning coverage https://shattered.io/ai-machine-learning/ for independent test results as they appear. FAQ: Muse Ember What is Muse Ember? Muse Ember is Meta’s smallest and most efficient model, announced by Meta AI on November 9, 2026. Meta designed it for repetitive agent work such as tool calls, coding subtasks, routing, extraction and long document parsing. How much does Muse Ember cost? For prompts up to 200K input tokens, it costs $0.10 per million input tokens and $0.50 per million output tokens. Above 200K input tokens, the rates are $0.30 and $1.50. Cached input costs $0.004 per million tokens on the shorter-context tier. How large is the context window? The context window holds 1,048,576 tokens, and the maximum output is 524,288 tokens. The model accepts text, image, video, audio and PDF input and produces text. Is Muse Ember better than Claude Haiku 5.5 or GPT-6 Luna? Meta claims it delivers more intelligence per dollar than both at any reasoning setting. That is a company claim and not an independently established fact, so wait for third-party tests or run your own evaluation before you switch. What did Muse Ember score on the Intelligence Index? Meta’s developer blog reports a score of 36 on the Artificial Analysis Intelligence Index, with an estimated cost of about $0.13 per index task at maximum reasoning effort. Where can developers use it? Muse Ember runs on Meta’s Model API under the ID muse-ember-1.0 and inside Muse Code. Meta has not confirmed a Muse Ember-specific subscription price. Are the weights or parameter count public? No. Reporting says Meta has not disclosed the parameter count or released model weights, so self-hosting is not an option at this point.