Muse Ember and Claude Haiku 5.5 Both Cost $0.10 Per Million Tokens. On Terminal-Bench 4.0, Ember Scores Higher at $0.22 a Test to Haiku’s $3.12.

3 small models, 1 price

The cheap tier of AI models has converged on one number. Claude Haiku 5.5, released October 7, lists at $0.10 / $0.50 per million input and output tokens. GPT-6 Luna, released September 22, lists at $0.10 / $0.50. Muse Ember, which Artificial Analysis dates to November 9, lists at $0.10 / $0.50.

If you pick a model off a pricing page, these 3 look interchangeable. On a real agentic coding benchmark, the bill for the same task spans 14x.

Prices from Meta, Anthropic and OpenAI. Terminal-Bench 4.0 from Vals.ai (Mini-SWE-agent harness, updated November 9, 2026). Index numbers from Artificial Analysis. Cost per solved test is cost per test divided by accuracy.

Where the $2.90 goes: tokens

When the per-token price is identical, the only thing left to explain the bill is how many tokens each model spends. Artificial Analysis publishes that number. To run its Intelligence Index, Muse Ember produced 115M output tokens. Claude Haiku 5.5 produced 298M, 2.6x as many, and both landed on the same score of 36.

That is why Ember’s cost per index task is $0.14 and Haiku’s is $0.41: the same score for about a third of the money. GPT-6 Luna is just as frugal as Ember, 110M tokens and $0.15 a task, but it scores 28.

Terminal-Bench 4.0 widens the gap. Its tasks run long, 26 minutes per test for Ember and 52 for Haiku 5.5 on Vals.ai’s board, and long agent loops pile up context. Haiku 5.5 bills prompts over 100,000 tokens at $0.50 / $2.50, 5x its base rate. Ember’s higher tier starts at 200K and is 3x its base. Vals.ai doesn’t break the cost down by tier, so treat the price cliff as a likely contributor, not a measured one.

The cheap end of Terminal-Bench 4.0

Ember isn’t only cheaper than Haiku. On cost per solved test, it is the cheapest model on the whole 63-model leaderboard.

Vals.ai Terminal-Bench 4.0, all runs in the Mini-SWE-agent harness. Cost per solved test is our division of the 2 published columns.

Ember ties Tencent’s Hy4 on accuracy at 17% of its cost per test. It beats GPT-5.6 Sol, 39.39% to 37.88%, at $0.22 a test to $7.98. And it finishes in 26 minutes, the fastest of any model in the table above.

It is still a small model. Claude Opus 5.5 and Kimi K3.1 Pro solve roughly 25 more tasks per 100. If a task has to be solved on the first try, the top of the board still wins.

Meta’s own table, and where Ember loses

Meta’s model page compares Ember (max effort) with Haiku 5.5 and GPT-6 Luna, both at max. Most rows favor Ember. 2 matter for coders.

From dev.meta.ai/models/muse-ember, as published by Meta.

Haiku 5.5 writes more mergeable code. On FrontierCode 2.0-AA, which judges real engineering tasks on whether the change could be merged, Haiku leads 34.9 to 27.2. Meta put that row in its own table.

The headline Terminal-Bench gap is smaller on neutral ground. Meta reports 40.9 to 39.2 using the “best published harness” for each model. On Vals.ai, with one harness for everyone, it’s 39.39% to 36.36%. Ember still wins, but read vendor tables as the ceiling, not the typical result.

Meta’s docs also say plainly what Ember is not for: hard single-shot coding problems, knowledge-heavy questions and visual coding. Meta recommends pairing it with Muse Spark or Muse Wonder, with the bigger model planning and Ember doing the volume.

How to get it

Ember is on Meta’s Model API as muse-ember-1.0, with Responses, Chat Completions and Messages endpoints, and on Google Cloud Marketplace. Reasoning effort runs from minimal to max and defaults to max, the setting Meta benchmarked.

It also comes with Muse Code, Meta’s subscription, from $5 a month for 10 to 50 prompts every 5 hours, up to $200 for 80x that. Meta’s docs say Ember is included in every plan, though the $5 plan’s card on the model page lists only Muse Spark, Muse Wonder and Muse Image. Check before you subscribe for Ember alone.

What this means if you code with AI

Stop comparing models by token price. 3 models share a $0.10 / $0.50 list price and cost $0.22, $0.35 and $3.12 per Terminal-Bench test. The price that matters is cost per solved task, and only benchmarks that publish cost per test let you compute it.

Use Ember-class models for the volume, not the hard part. In an AI coding agent, most calls are small: read a file, run a test, grep, summarize a log. That’s the work Meta built Ember for. Keep a frontier model for planning and the tricky diff.

Watch the long-context line. Haiku 5.5 switches to 5x rates at 100K tokens, Ember to 3x at 200K, GPT-6 Luna to 2x input and 1.5x output past 272K. Agent loops cross those lines without telling you. If your tool shows context size, keep an eye on it.

Run your own 10 tasks. Ember wins on Terminal-Bench and loses on mergeable code. Which result matches your work depends on your repo, so test both on real tickets before you switch a whole team.

Full disclosure: we sell an AI coding agent. On MegaBrain you can bring your own API key or a local model on every plan, Free included, with no limits from us. Qwen, Kimi and GLM are included on the paid plans, and Claude, GPT and Gemini come from Pro. See the plans. MegaBrain is not affiliated with Meta, Anthropic, OpenAI or Vals.ai.

Fun facts

Ember is the fastest of the 3 on Artificial Analysis: 440 output tokens per second, to Haiku 5.5’s 254 and GPT-6 Luna’s 102.

Ember can write up to 524,288 tokens in one response, about 4x GPT-6 Luna’s 128,000 cap. Its cached input price, $0.004 per million tokens, is 60% below the other 2 models’ $0.01.

On Terminal-Bench 4.0, Ember ties Hy4 at 39.39%. Hy4 Preview, on the same board, scored 8.08%.

Sources

Pages opened November 10, 2026. Every number above comes from these pages, or from dividing 2 of their numbers where we say so.

• Meta — Muse Ember model page (pricing tiers, channels, benchmark table, Muse Code plans)
• Meta — Muse Ember API docs (model ID, context window, max output, reasoning effort, recommended and not-recommended uses)
• Vals.ai — Terminal-Bench 4.0 leaderboard (updated November 9, 2026)
• Artificial Analysis — Muse Ember (Max) model page
• Artificial Analysis — Claude Haiku 5.5 model page
• Artificial Analysis — GPT-6 Luna model page
• Anthropic — Claude API pricing (Claude Haiku 5.5 rates under and over 100,000 tokens)
• OpenAI — GPT-6 Luna model docs (pricing, long-context tier, limits)