Muse Ember vs GPT-6 Luna: Complete Comparison and Report on Batch Discounts, Long-Context Rates, Agentic Benchmarks, and Speed
Muse Ember is Meta's smallest and most efficient reasoning model, released on November 9, 2026 under the API identifier muse-ember-1.0 and sold through the Meta Model API and Muse Code. GPT-6 Luna is OpenAI's efficiency model for focused, high-volume work, released in the API on September 22, 2026 together with GPT-6 Sol, and also the model that Free, Go and Plus users meet in ChatGPT. The two sit in exactly the same price slot, with a list rate of $0.10 per 1M input tokens and $0.50 per 1M output tokens for both, a 1M-token class context window for both, and a reasoning-effort control that runs up to a "max" setting on both. Meta used GPT-6 Luna as one of the two named rivals in its own launch charts, which makes this pairing the most direct test of whether Meta's newest small model changes the choice for developers already buying OpenAI's most efficient GPT-6 model.
On the measurements that both models share, Muse Ember leads clearly. Artificial Analysis scores it 36 on its Intelligence Index v5.0.1 against 28 for GPT-6 Luna, both at maximum effort, and records 454 output tokens per second against 109, with a time to first token of 4.94 seconds against 80.13 seconds. On Terminal-Bench 4.0 the gap is the widest of any shared test, 40.9% for Ember against 12.6% for Luna in the Artificial Analysis run. Luna keeps a narrow edge on one physics reasoning test (CritPt v2 Hard Subset, 42.5% against 39.0%) and ties Ember on GDP.pdf at 22.8%.
The money favors GPT-6 Luna in two specific situations. OpenAI sells Batch and Flex processing at 50% of standard rates, so offline Luna work costs $0.05 input and $0.25 output, while Meta lists no batch tier for Ember. On very long prompts, Meta switches Ember to $0.30 input and $1.50 output for the whole request once the prompt passes 200K tokens, while OpenAI's long-context surcharge starts only above 272K tokens and adds less (2x input, 1.5x output). Muse Ember is the better default for latency-sensitive agents, coding subagents and multimodal inputs that include audio or video; GPT-6 Luna is the better choice for discounted batch pipelines, prompts between 200K and 1M tokens, and teams that need EU data residency or OpenAI's wider set of hosted tools. All prices in this comparison were verified on the official Meta and OpenAI pricing pages on November 11, 2026.
··········
MUSE EMBER AND GPT-6 LUNA RELEASE TIMELINE AND MODEL IDENTITY.
The two models arrived seven weeks apart and come from different product lines. GPT-6 Luna launched in the OpenAI API on September 22, 2026 alongside GPT-6 Sol, at the same $0.10/$0.50 rate it carries today. OpenAI's changelog records a September 25 fix for an image-encoding bug that had degraded image understanding in both Sol and Luna, followed by the Decisions API, which entered beta with gpt-6-luna on October 6 and became generally available on October 22 with support for up to 400 questions per request. In ChatGPT, OpenAI shipped an October version of GPT-6 Luna on October 7 that replaced GPT-5.6 Luna for Free and paid users, and a further update dated October 29 that produces shorter answers and less agreeable language. OpenAI's October system card states that Codex and ChatGPT Work users stay on the September version, while the API model page lists a single snapshot, gpt-6-luna, without saying which of those versions it serves.
Muse Ember is the first model Meta trained using Muse Wonder Full, its largest model, and Meta presents it as the company's smallest multimodal and smallest reasoning model. It was announced on November 9, 2026 on Meta's research blog and became available the same day in the Meta Model API and in Muse Code. Meta positions it for bounded, repeated work such as tool calls, coding subtasks, routing, extraction and document summarization, and for subagent roles under a larger planner such as Muse Wonder or Muse Spark. Our launch report on Meta's Muse Ember pricing and 1M context covers the announcement itself.
........
Item | Muse Ember | GPT-6 Luna |
Developer | Meta | OpenAI |
API model ID | muse-ember-1.0 | gpt-6-luna |
API release | November 9, 2026 | September 22, 2026 |
Later updates | None yet | Sep 25 image fix; ChatGPT versions Oct 7 and Oct 29 |
Vendor positioning | Smallest, most efficient Muse model | Most efficient model for focused, high-volume tasks |
Training note | First model trained using Muse Wonder Full | Smaller sibling of GPT-6 Sol |
Weights | Closed | Closed |
........
··········
MUSE EMBER VS GPT-6 LUNA API PRICING, CACHE RATES AND BATCH DISCOUNTS.
The headline rates are identical, and the differences sit in the secondary lines. Cached input is cheaper on Ember, at $0.004 per 1M tokens against $0.01 for Luna, which matters for applications that resend a large fixed system prompt on every call. OpenAI also bills cache writes, at $0.125 per 1M tokens for its 30-minute cache and $0.25 for its 12-hour cache, while Meta's pricing page lists no cache-write charge for Ember. In the opposite direction, OpenAI's Batch and Flex tiers halve every Luna rate, and its Fast mode doubles them for lower latency; Meta lists neither a batch tier nor a priority tier for Ember, and its Contributor mode, which discounts Muse Spark in exchange for training rights, is not offered for Ember.
........
Price per 1M tokens | Muse Ember | GPT-6 Luna Standard | GPT-6 Luna Batch/Flex | GPT-6 Luna Fast |
Input | $0.10 | $0.10 | $0.05 | $0.20 |
Cached input | $0.004 | $0.01 | $0.005 | $0.02 |
Cache write (30 min) | Not listed | $0.125 | $0.0625 | $0.25 |
Cache write (12 h) | Not listed | $0.25 | $0.125 | $0.50 |
Output | $0.50 | $0.50 | $0.25 | $1.00 |
Batch tier | Not listed | 50% of Standard | Included | Not applicable |
Regional processing | Not listed | +10% uplift | +10% uplift | +10% uplift |
........
The extra charges that sit outside token rates also differ. Meta bills web search grounding at $2.50 per 1,000 queries on top of Ember's token costs. OpenAI applies a 10% uplift to regional processing (data residency) endpoints for models released on or after March 5, 2026, which includes GPT-6 Luna, and the same uplift to FedRAMP endpoints. For a broader view of OpenAI's current price list, our overview of all OpenAI API models and their prices places Luna against Sol, Terra and Astra.
··········
LONG-CONTEXT RATES FOR MUSE EMBER AND GPT-6 LUNA ABOVE 200K AND 272K TOKENS.
Both vendors charge more for very long prompts, and both apply the higher rate to the whole request once the threshold is crossed, but the thresholds and multipliers are different. Meta moves Ember to its long-context tier when total input, including cached tokens, exceeds 200K tokens, and the rates triple to $0.30 input, $0.012 cached input and $1.50 output. Meta also counts the stored conversation as input on requests that use previous_response_id, which can push multi-turn agent sessions into the higher tier without any single large document. OpenAI applies its surcharge only above 272K input tokens and charges 2x the input and cache rates and 1.5x the output rate for the full request.
........
Prompt size | Muse Ember input / output | GPT-6 Luna input / output (Standard) |
Up to 200K tokens | $0.10 / $0.50 | $0.10 / $0.50 |
200K to 272K tokens | $0.30 / $1.50 | $0.10 / $0.50 |
Above 272K tokens | $0.30 / $1.50 | $0.20 / $0.75 |
Cached input above threshold | $0.012 | $0.02 |
Threshold counts cached tokens | Yes | Yes (full request repriced) |
........
The band between 200K and 272K tokens is where the gap is largest, since Ember already pays three times its base rate while Luna is still at standard pricing. Above 272K, Ember's input still costs 50% more than Luna's and its output costs twice as much. Teams that routinely send whole codebases, long contracts or multi-hour transcripts into a small model will see the difference on every invoice.
··········
CONTEXT WINDOW, OUTPUT LIMITS, MODALITIES AND EFFORT CONTROLS.
The two context windows are practically equal, at 1,048,576 tokens for Ember and 1,050,000 tokens for Luna, but the output ceilings are far apart. Meta allows up to 524,288 output tokens per Ember response, four times Luna's 128,000-token limit, which affects very long generations such as full-file rewrites or bulk extraction in a single call. Ember also accepts audio, video and PDF input natively alongside text and images, while GPT-6 Luna accepts text and images and produces text only.
The reasoning controls look similar on paper and behave differently by default. Ember's effort setting runs from "minimal" to "max", and max is the default; GPT-6 Luna supports none, low, medium, high, xhigh and max, with medium as the default. A developer who swaps one model for the other without setting effort explicitly will therefore compare Ember at its most expensive setting against Luna at a middle setting, which changes token usage, latency and cost. OpenAI also notes that on Chat Completions, Luna supports function calling only when effort is set to none, with built-in tools and function calling otherwise handled through the Responses API.
........
Specification | Muse Ember | GPT-6 Luna |
Context window | 1,048,576 tokens | 1,050,000 tokens |
Maximum output | 524,288 tokens | 128,000 tokens |
Input modalities | Text, image, video, audio, PDF | Text, image |
Output | Text | Text |
Effort levels | minimal to max | none, low, medium, high, xhigh, max |
Default effort | max | medium |
Knowledge cutoff | Not published | May 18, 2026 |
Fine-tuning | Not listed | Not supported |
API endpoints | Responses, Chat Completions, Messages | Responses, Chat Completions, Batch, Realtime and others |
........
··········
WHERE MUSE EMBER AND GPT-6 LUNA ARE AVAILABLE: API, APPS, CLOUDS AND RATE LIMITS.
GPT-6 Luna has the wider distribution by a large margin. Beyond the OpenAI API, it is listed on Amazon Bedrock as openai.gpt-6-luna and in private preview on Google Cloud, and it supports EU data residency on the Standard, Fast, Flex and Batch tiers. In ChatGPT, OpenAI's pricing page lists GPT-6 Luna for Free and Go users at standard speed in the desktop app subject to rollout, and alongside GPT-6.1 Sol on the $20 Plus plan, where the usage table estimates 350 to 3,000 local messages per five-hour period. Muse Ember is sold through the Meta Model API and through Muse Code, where Meta's documentation says it is available on every plan, from the $5 Everyday tier to the $200 Maximum tier; its model page also links a Google Cloud Marketplace listing for enterprise use. Meta does not mention Ember in the consumer Meta AI app.
........
Access route | Muse Ember | GPT-6 Luna |
First-party API | Meta Model API | OpenAI API |
Coding product | Muse Code, all plans ($5 to $200 a month) | Codex (September version) |
Consumer app | Not mentioned for Meta AI | ChatGPT Free, Go, Plus and other plans |
Amazon Bedrock | Not listed | openai.gpt-6-luna |
Google Cloud | Marketplace listing | Private preview |
EU data residency | Not stated | Standard, Fast, Flex, Batch |
Base rate limit | 3,000 RPM, 4M TPM | Build tier 5,000 RPM, 2M TPM |
Highest listed tier | Single tier only | Grow tier 30,000 RPM, 180M TPM |
........
The rate limits favor Ember for a new account and Luna for a scaled one. Meta lists a single Ember tier of 3,000 requests and 4 million tokens per minute, which is twice the token throughput of OpenAI's entry Build tier. OpenAI's limits rise with spend, to 10,000 requests and 10 million tokens per minute on Launch and 30,000 requests and 180 million tokens per minute on Grow, plus batch queues of up to 15 billion tokens. Our guide to Meta AI API access and Muse model prices covers the rest of the Meta Model API.
··········
BENCHMARK SOURCES FOR MUSE EMBER AND GPT-6 LUNA AND HOW RELIABLE THEY ARE.
Almost every capability figure for this pair comes from two places, Meta's launch materials and Artificial Analysis. OpenAI has published safety and dangerous-capability evaluations for GPT-6 Luna in its system cards, but no general capability scores for Luna that can be set against Ember's. Meta ran Ember, Claude Haiku 5.5 and GPT-6 Luna at maximum reasoning effort, and several of its rows are labelled "-AA", meaning they reuse Artificial Analysis evaluations. Two of Meta's tests, OSWorld 2.1 and PerceptionBench, are Meta-run, and its Personal Superintelligence Index is an internal composite of eight evaluations that no third party reproduces.
The two sources do not always agree, and the differences should be read before quoting any number. On Terminal-Bench 4.0, Meta reports 16.4% for GPT-6 Luna using OpenAI's Codex agent framework, while Artificial Analysis reports 12.6% in its own run; both sources give Ember 40.9%, with Meta specifying the mini-swe-agent framework. On AutomationBench-AA, Meta's chart gives 37.8 for Ember and 20.2 for Luna on a strict pass@1 basis, while the Artificial Analysis comparison page lists 68.1% and 53.2%, which points to a different scoring method on the same benchmark. Meta's launch blog and its developer blog also give slightly different cost-per-task figures for Ember ($0.14 and about $0.13), while Artificial Analysis lists $0.14.
··········
[[ADPLUS_300x250_1]]
MUSE EMBER VS GPT-6 LUNA ON DIRECTLY COMPARABLE BENCHMARKS.
The table below keeps the two sources apart. Every row compares both models under the same evaluator and the same maximum effort setting, so each row is internally consistent even where the two sources differ from each other.
........
Benchmark (max effort) | Source | Muse Ember | GPT-6 Luna |
Intelligence Index v5.0.1 | Artificial Analysis | 36 | 28 |
Terminal-Bench 4.0 | Artificial Analysis | 40.9% | 12.6% |
Terminal-Bench 4.0 | Meta (best published framework) | 40.9 | 16.4 |
AutomationBench-AA | Artificial Analysis | 68.1% | 53.2% |
AutomationBench-AA (strict pass@1) | Meta | 37.8 | 20.2 |
Coding Agent Index-AA | Meta | 46 | 31 |
FrontierCode 2.0-AA | Both sources | 27.2% | 25.2% |
GDP.pdf-AA | Both sources | 22.8% | 22.8% |
HLE-Diamond | Artificial Analysis | 25.1% | 22.2% |
CritPt v2 (Hard Subset) | Artificial Analysis | 39.0% | 42.5% |
Terminal-Bench-Science 0.1 | Artificial Analysis | 20.0% | 8.6% |
AA-Omniscience v1.1 | Artificial Analysis | 7 | -8 |
AA-Briefcase v1.1 | Artificial Analysis | 1521 | 1337 |
........
The pattern is consistent across agentic and terminal work, where Ember leads by 24.5 to 28.3 points on Terminal-Bench 4.0 depending on the source and by 15 points on Meta's Coding Agent Index. On knowledge and science reasoning, the margins shrink to a few points, and on CritPt v2 Luna scores higher. The AA-Omniscience result, 7 for Ember against -8 for Luna, measures knowledge reliability with a penalty for wrong answers, so a negative score means Luna gave more incorrect answers than correct ones on that test.
··········
MUSE EMBER RESULTS THAT ONLY META PUBLISHED.
Several of Ember's strongest numbers come from evaluations that Meta designed or ran itself. On OSWorld 2.1, a computer-use benchmark, Meta reports 69.5 partial credit and 33.8 binary success for Ember against 50.5 and 17.0 for GPT-6 Luna. On PerceptionBench with tools, Ember scores 80.6 against 62.8. On Agents' Last Exam (Linux subset) the gap is smaller, 61.6 against 57.1, and on the internal Personal Superintelligence Index Ember scores 46.8 against 37.0.
........
Meta-run or internal evaluation | Muse Ember | GPT-6 Luna | Independent replication |
OSWorld 2.1 (partial credit) | 69.5 | 50.5 | None published |
OSWorld 2.1 (binary) | 33.8 | 17.0 | None published |
PerceptionBench (with tools) | 80.6 | 62.8 | None published |
Agents' Last Exam (Linux subset) | 61.6 | 57.1 | None published |
Personal Superintelligence Index | 46.8 | 37.0 | Internal composite |
........
These figures are vendor claims. They point in the same direction as the independent Artificial Analysis results, which supports their general direction, but the exact margins have not been checked by another organization. OpenAI did publish dangerous-capability figures for GPT-6 Luna in its October 29 addendum, such as 46.19% on ExploitBench at maximum effort and 51.75% on SEC-Bench Pro, but Meta published no equivalent cyber scores for Ember, so those rows cannot be compared.
··········
SPEED, LATENCY AND TOKEN USE MEASURED BY ARTIFICIAL ANALYSIS.
Speed is the largest measured difference between the two models. In Artificial Analysis measurements at maximum effort, Ember generated 454 output tokens per second against 109 for GPT-6 Luna, and its time to first token was 4.94 seconds against 80.13 seconds. The mean end-to-end time for a single request was 13.12 seconds for Ember and 170.06 seconds for Luna, and the average time per Intelligence Index task was 3.1 minutes against 11.5 minutes.
........
Measurement (max effort) | Muse Ember | GPT-6 Luna |
Output speed | 454 tokens/s | 109 tokens/s |
Time to first token | 4.94 s | 80.13 s |
End-to-end time (single request, mean) | 13.12 s | 170.06 s |
Time per Intelligence Index task | 3.1 min | 11.5 min |
Output tokens per task | 84.5K | 75.3K |
Reasoning tokens per task | 52.2K | 57.6K |
Cost per task | $0.14 | $0.15 |
Cost to run the full Index | $117 | $136 |
........
Token use is close. Ember produced more output tokens per task (84.5K against 75.3K) but fewer of them were reasoning tokens (52.2K against 57.6K), and its cheaper cache rate brought its cost per task to $0.14 against $0.15. Meta's partners add their own latency claims for Ember, including a p95 latency of 2.2 seconds measured by You.com and tail latency up to 30% lower than a comparable model in Cloudflare's tests, but those partners did not name the models they compared against, and none of these partner figures involve GPT-6 Luna. The Luna latency figures above apply to maximum effort; Artificial Analysis also tracks Luna at lower settings, and OpenAI's default for the model is medium.
··········
WHAT MUSE EMBER AND GPT-6 LUNA COST ON THREE TYPICAL WORKLOADS.
The cost examples below use only the official list prices verified on November 11, 2026 and were computed in Python. Cache-write charges are excluded from both sides because Meta lists none and OpenAI's depend on how often the cache refreshes; regional-processing uplifts and web search fees are also excluded.
The first workload is a high-volume support or classification service handling 1 million requests a month, each with a 2,000-token cached system prompt, 1,000 fresh input tokens and a 300-token answer. The second is long-document analysis, 10,000 requests a month at 400,000 fresh input tokens and 4,000 output tokens each, with two variants at 250,000 and 150,000 input tokens to show the threshold effect. The third is a coding-subagent fleet running 1,000 tasks a day for 30 days, priced at the Artificial Analysis cost per task for each model at maximum effort.
........
Workload (monthly) | Muse Ember | GPT-6 Luna Standard | GPT-6 Luna Batch |
Support/classification, 1M requests | $258 | $270 | $135 |
Long documents, 10,000 x 400K tokens | $1,260 | $830 | $415 |
Long documents, 10,000 x 250K tokens | $810 | $270 | $135 |
Long documents, 10,000 x 150K tokens | $170 | $170 | $85 |
Coding subagents, 30,000 tasks (AA cost per task) | $4,200 | $4,500 | Not measured |
Coding subagents, total task time | about 1,550 hours | about 5,750 hours | Not measured |
........
On the support workload, Ember's cheaper cached input saves about 4% against Luna at standard rates, or $12 on a $270 bill, but moving the same traffic to Luna's Batch tier halves the bill to $135 for any job that can wait for asynchronous processing. On long documents, the 250K-token case shows the full weight of the threshold difference, with Ember costing three times as much because it is already on long-context rates while Luna is not. At 150K tokens the two are identical. For the subagent fleet, the cost gap is small at $300 a month, while the task-time gap is close to 4,200 hours of wall-clock time, which decides throughput for any pipeline that waits on the model.
··········
SAFETY FRAMEWORKS, DATA USE AND RESTRICTIONS.
The two companies publish safety results under different frameworks, so the classifications cannot be placed on one scale. OpenAI's October system card and its October 29 addendum treat GPT-6 Luna as High capability in Biological and Chemical and in Cybersecurity under its Preparedness Framework, below Critical in cyber and below High in AI Self-Improvement, with the same safeguards as GPT-5.6 Sol and Luna. The addendum reports a 99.91% score on instruction-hierarchy prompt-injection evaluations and notes higher factual-error rates than the October 7 version on all six hallucination metrics, while still lower than GPT-5.6 Luna on each. Meta says Muse Ember stayed below the high-risk thresholds of its Meta Superintelligence Scaling Framework in the categories it assessed, including chemical and biological, cybersecurity and loss-of-control risks, and that it received the same core safety training and testing as Meta's larger models, without publishing per-evaluation numbers.
On data use, Meta's pricing page states that content sent in Standard mode is not used to train Meta models, and Ember is offered only in Standard mode. OpenAI offers EU data residency for Luna on all main processing tiers at a 10% uplift. Meta's Ember pages do not state a data-retention period or regional processing options, and Meta notes that Muse availability can vary by country under its Geographic Use Policy.
··········
GAPS IN THE PUBLISHED EVIDENCE FOR BOTH MODELS.
Several facts that buyers would want are missing from the official pages. Meta has not published a knowledge cutoff, a batch tier, cache-write pricing, a data-retention period or tiered rate limits for Ember, and its benchmark charts do not map each data point to a named effort level. OpenAI has not published general capability benchmarks for GPT-6 Luna on its own pages, and its API model page lists one snapshot without saying whether it serves the September version used in Codex or the October versions shipped in ChatGPT. Neither model appears in an independent computer-use or cyber evaluation that covers both, and no third party has yet reproduced Meta's OSWorld 2.1 or PerceptionBench figures.
··········
WHICH ONE TO CHOOSE: MUSE EMBER OR GPT-6 LUNA BY WORKLOAD.
With identical list prices, the decision rests on latency, prompt length, batch eligibility and ecosystem. Our earlier report on Muse Ember vs Claude Haiku 5.5 covers the other rival Meta named at launch.
........
Workload or requirement | Better choice | Reason |
Interactive agents and coding subagents | Muse Ember | 454 vs 109 tokens/s, 4.94 s vs 80.13 s to first token at max effort |
Terminal and computer-use tasks | Muse Ember | 40.9% vs 12.6% on Terminal-Bench 4.0 (Artificial Analysis) |
Audio, video or PDF input | Muse Ember | Native audio, video and PDF input; Luna takes text and images |
Outputs longer than 128K tokens | Muse Ember | 524,288-token output limit |
Offline batch classification or extraction | GPT-6 Luna | Batch and Flex at 50% of standard rates |
Prompts between 200K and 1M tokens | GPT-6 Luna | Surcharge starts at 272K and adds less |
EU data residency or Bedrock deployment | GPT-6 Luna | Listed on both; not stated for Ember |
Very high sustained throughput | GPT-6 Luna | Grow tier up to 30,000 RPM and 180M TPM |
Large fixed system prompts, real-time traffic | Muse Ember | Cached input at $0.004 vs $0.01 |
........
For a developer building real-time agents, tool-calling loops or coding subagents, the independent measurements point to Muse Ember, which scores eight Intelligence Index points higher than GPT-6 Luna at the same list price and returns its first token about 16 times faster at maximum effort. For a team running large asynchronous jobs, or sending prompts of 200K tokens and more, GPT-6 Luna costs less on every request, by half with the Batch tier and by two thirds in the 200K to 272K band.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
