Muse Ember vs Claude Haiku 5.5: Meta's low-cost model for coding

AutoRouter · Nov 9, 2026 · Updated Nov 10, 2026 · 10 min read

Muse Ember is the smallest model in Meta's Muse family, released on November 9, 2026 for the tool calls, coding subtasks and routing that agents repeat. On Meta's API it costs $0.10 per million input tokens and $0.50 per million output tokens for requests up to 200,000 input tokens, as of November 10, 2026. Claude Haiku 5.5 and GPT-6 Luna list the same two prices on prompts up to 100,000 tokens. Meta says it matches Haiku 5.5's best score on the Artificial Analysis Intelligence Index at about a third of the cost per task. Meta's own docs give multi-file changes to its largest model, so hard coding work still belongs on a frontier model such as Claude Opus 5.5. AutoRouter, the Mac app for Codex and Claude Code, has Muse Ember in its model picker.

AutoRouter wrote this post. Every outside figure below comes from the maker's or publisher's own page, linked where it is used and checked on November 10, 2026, one day after the launch.

What is Muse Ember?

Muse Ember is Meta's smallest and lowest-priced reasoning model, built to run the bounded steps that a larger model hands out. As of November 10, 2026, Meta's model page gives these details:

• Where it runs: Meta Model API, as muse-ember-1.0, and Muse Code, Meta's coding agent, on pay-as-you-go and every subscription plan. OpenRouter lists it as meta/muse-ember at the same input and output prices.
• Size: a context window of 1,048,576 tokens and up to 524,288 output tokens.
• Input: text, image, video, audio and PDF. It outputs text.
• Reasoning: effort runs from minimal to max, and max is the default. Reasoning can't be switched off, and reasoning tokens bill as output tokens.

What does Muse Ember cost next to Claude Haiku 5.5 and GPT-6 Luna?

Muse Ember, Claude Haiku 5.5 and GPT-6 Luna all list $0.10 per million input tokens and $0.50 per million output tokens as of November 10, 2026. They differ in the price of cached input and in the prompt size where a higher price tier starts: 100,000 tokens for Haiku 5.5, 200,000 for Muse Ember and 272,000 for GPT-6 Luna. On all three, a request over the threshold pays the higher prices on the whole request.

The arithmetic below prices one assumed agent request, the same one this blog's other pricing posts use: 45,000 cached input tokens, 5,000 new input tokens and 2,000 output tokens, thinking included. On Haiku 5.5 the new tokens bill at the 5-minute cache-write price. Meta lists no cache-write price, so on Muse Ember they bill at the input price.

Assumed request: 45,000 cached input, 5,000 new input, 2,000 output tokens
List prices per million tokens as of November 10, 2026

Muse Ember (first tier, up to 200,000 input tokens)
  45,000 cached input  x  $0.004  =  $0.00018
   5,000 new input     x  $0.10   =  $0.00050   (no cache-write price listed)
   2,000 output        x  $0.50   =  $0.00100
  Total                              $0.00168

Claude Haiku 5.5 (first tier, prompts up to 100,000 tokens)
  45,000 cached input  x  $0.01   =  $0.00045
   5,000 new input     x  $0.125  =  $0.000625  (5-minute cache write)
   2,000 output        x  $0.50   =  $0.00100
  Total                              $0.002075

Longer prompt: 135,000 cached input, 15,000 new input, 2,000 output tokens

Muse Ember (still first tier)
  $0.00054 + $0.00150 + $0.00100   =  $0.00304

Claude Haiku 5.5 (higher tier: $0.05 cached, $0.625 cache write, $2.50 output)
  $0.00675 + $0.009375 + $0.00500  =  $0.021125

At the assumed size, Muse Ember costs about $0.0017, which is 19% less than Haiku 5.5's $0.0021. GPT-6 Luna also comes to $0.0021, because OpenAI bills cache writes at 1.25 times the input price. Past 100,000 prompt tokens the gap widens: the longer request costs about $0.021 on Haiku 5.5, on its higher tier, and $0.003 on Muse Ember, a seventh as much.

The arithmetic gives both models the same token counts, and real counts differ: Muse Ember reasons at max effort unless you lower it, while Haiku 5.5's default effort is medium.

What does Meta claim about Muse Ember, and who measured it?

Meta claims that Muse Ember matches Claude Haiku 5.5's best score on the Artificial Analysis Intelligence Index at about a third of the cost per task, and that it beats every GPT-6 Luna setting (Meta's launch post). Artificial Analysis ran the index, and its own pages list these results as of November 10, 2026:

The three models share the same token prices, so the difference in cost per task comes from how many tokens each one used. Three things limit what this tells you about coding:

• The comparison is at max effort. Haiku 5.5's default is medium, and Meta's chart shows it costing less per task, with a lower score, at lower efforts.
• Meta's coding results are mixed. Muse Ember leads on two benchmarks below, and Haiku 5.5 leads on FrontierCode 2.0-AA, which Meta describes as engineering tasks judged on whether the change could be merged.
• No frontier model is in the comparison. The results rank small models against each other.

Which coding work suits Muse Ember, and which needs a frontier model?

Muse Ember suits bounded subtasks with a result you can check, and hard, multi-file or ambiguous work belongs on a frontier model such as Claude Opus 5.5. Meta and Anthropic each draw that line for their own small model, as of November 10, 2026:

Anthropic's guide says small models do well on high-volume work with checkable outputs and are "less suited to long agentic loops". Meta says the same about Muse Ember: for hard single-shot coding problems it points to Muse Spark or Muse Wonder, and in its example coding workflow the larger model plans and reviews while Muse Ember works on one bounded subtask. A small model's result on bounded tasks does not carry over to hard ones.

Which model to use for which coding task maps the rest of your work to a model tier, and How to send quick coding tasks to a cheaper model shows how to move the small jobs.

How do you try Muse Ember in Claude Code or Codex?

Point either app at Meta's API with a Meta Model API key, following Meta's coding agents guide, or pick Muse Ember in the model picker if you use AutoRouter. Meta's guide runs the main loop on Muse Wonder and gives Muse Ember the small jobs.

Claude Code. Meta's guide sets these environment variables before you start claude. The haiku alias and the subagent model both go to Muse Ember:

export ANTHROPIC_BASE_URL="https://api.meta.ai"
export ANTHROPIC_AUTH_TOKEN="$MODEL_API_KEY"
export ANTHROPIC_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_OPUS_MODEL="muse-wonder-1.0"
export ANTHROPIC_DEFAULT_SONNET_MODEL="muse-spark-1.3"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="muse-ember-1.0"
export CLAUDE_CODE_SUBAGENT_MODEL="muse-ember-1.0"
export ENABLE_TOOL_SEARCH="true"

With these set, the whole session runs on Meta's models, because Meta's API serves no Claude models. To try Muse Ember alone, set all five model variables to the same ID, as the guide says. While ANTHROPIC_AUTH_TOKEN is set, Claude Code sends it in place of your claude.ai login, and those requests bill per token to the key's owner.

Codex. Meta's guide adds a [model_providers.meta] table to ~/.codex/config.toml (Codex provider settings), or runs Codex from a scratch CODEX_HOME that leaves your existing setup alone. Its examples use Muse Wonder and Muse Spark. Muse Ember's model ID is muse-ember-1.0.

Try a new model in Codex or Claude Code without another API key covers the general routes to any new model.

How does AutoRouter handle a new model like Muse Ember?

Muse Ember is in AutoRouter's model picker, where you can choose it yourself or name it in a plain-English rule. AutoRouter, the Mac app for Codex and Claude Code, puts new models in the picker you already use right away, with no keys or endpoints to set up, and the AutoRouter team tests every new model as it becomes available.

• Choose it. Pick Muse Ember from the model menu in Codex, or with /model in Claude Code, and your prompts go straight to it.
• Name it in a rule. Rules are plain English, for example "Use Muse Ember when a subagent reads logs." Your rules come before AutoRouter's own.
• Or select AutoRouter. It works out how hard each task is and only considers models that clear the quality bar for it. Hard work goes to the most capable model your plans cover, and unclear work is never treated as a quick task. Untested models are never picked on their own.
• See what happened. Every request shows its model, what it cost and why.

Models outside your ChatGPT and Claude plans come with AutoRouter Pro, at $50 a month with $20 of included usage each month. Paid models bill at the provider's price plus 10%, which AutoRouter earns. AutoRouter runs on macOS 14 or later and is in beta. Requests pass through its servers on their way to each provider (AutoRouter privacy).

Questions

Is Muse Ember cheaper than Claude Haiku 5.5?

Per input and output token it costs the same: both list $0.10 and $0.50 per million tokens as of November 10, 2026 (Meta, Anthropic). Muse Ember's cached input costs $0.004 per million tokens against Haiku 5.5's $0.01, and its higher price tier starts at 200,000 input tokens where Haiku 5.5's starts at 100,000.

When is Claude Haiku 5.5 the better choice?

Haiku 5.5 is the better choice when you already work in Claude Code with Claude models: the haiku alias resolves to Haiku 5.5 on the Anthropic API in Claude Code v2.1.293 or later, with no key or endpoint to add. It also led Muse Ember on FrontierCode 2.0-AA in Meta's own results, 34.9% to 27.2%. GPT-6 Luna is the built-in small model in Codex in the same way.

Can Muse Ember replace Claude Opus 5.5 for coding?

No. Frontier models are much better at real, multi-step coding work, and none of the launch results compares Muse Ember with one. Meta's docs give multi-file changes and a coding agent's main loop to Muse Wonder, and Anthropic's guide says to start most work on Claude Opus 5.5. Use Muse Ember under a larger model, for the bounded steps that model hands out.

Does Meta train on what you send to Muse Ember?

On Meta Model API, Muse Ember runs only in Standard mode, and Meta's pricing page says content sent in Standard mode is not used to train Meta models, as of November 10, 2026. Muse Code subscriptions start in Contributor mode, which lets Meta train on your content in exchange for more usage until you switch it off in settings.

Sources

• Introducing Our Smallest, Most Efficient Model: Muse Ember, Meta AI Research blog, November 9, 2026
• Start building with Muse Ember, Meta developer blog
• Muse Ember, Meta Model API docs
• Pricing and rate limits, Meta Model API docs
• Choose a model, Meta Model API docs
• Use Model API with coding agents, Meta Model API docs
• Reasoning, Meta Model API docs
• Prompt caching and token counting, Meta Model API docs
• Muse Code and Muse Code subscriptions, Meta docs
• Meta: Muse Ember, OpenRouter's listing
• Muse Ember, Claude Haiku 5.5 and GPT-6 Luna, Artificial Analysis
• Pricing, Claude Platform docs, Anthropic
• Claude Haiku 5.5, Claude Platform docs, Anthropic
• Which Claude model should you build with?, Anthropic, November 9, 2026
• LLM gateways, Claude Code docs
• API pricing, OpenAI
• GPT-6 Luna, OpenAI API docs
• Models in Codex, OpenAI
• Advanced configuration, Codex docs, OpenAI
• AutoRouter and AutoRouter privacy