Muse Ember's price sheet is the easiest table on this site to misread: its rates are identical to Claude Haiku 5.5 and GPT-6 Luna, yet the saving is real; and the headline $0.10 holds only below 200K input tokens — above it, the whole request bills at 3×. This guide works through both tiers, the whole-request math and three worked examples.
| Item (per 1M tokens) | ≤ 200K input | > 200K input |
|---|---|---|
| Input | $0.10 | $0.30 |
| Cached input | $0.004 | $0.012 |
| Output | $0.50 | $1.50 |
The word that matters is "whole request": for a request over 200,000 input tokens, input, cached input and output all bill at 3× — applied to the whole request, not just the part above 200K. That is why tech-insider ran dedicated cost math on the cliff on Nov 10–11, and it is the most distinctive line in Ember's pricing story. Standard mode only; no Contributor mode.
Take a task with a fixed 50K-token output and vary only the input:
| Task | Input | Output | Input cost | Output cost | Total |
|---|---|---|---|---|---|
| Short-context task | 100K | 50K | $0.010 | $0.025 | $0.035 |
| Exactly 200K | 200K | 50K | $0.020 | $0.025 | $0.045 |
| Just over 200K | 250K | 50K | $0.075 | $0.075 | $0.150 |
Input rises from 200K to 250K (+25%), but the bill rises from $0.045 to $0.150 (+233%) — the output rate triples along with it. Long-document parsing and over-long-context agent work hit this line most easily: 200K tokens is roughly 150,000 English words or about 300,000 Chinese characters, so one long report plus conversation history can cross it. Ways around it: split the request, shrink the context, or restructure the over-long part for higher cache hits.
Ember, Haiku 5.5 and GPT-6 Luna share the same list price ($0.10/$0.50); the saving is in usage, not rates. On the AA Index, Ember used ~115M output tokens against Haiku's ~298M — Ember does the same task set with fewer tokens, so cost per task is $0.14 vs $0.41. Cached input is also ~60% cheaper ($0.004 vs Haiku's ~$0.01, per gptclub). Read the other way: if your task makes Ember output as much as Haiku, the saving disappears. "A third of the cost" is a per-task claim, not a per-token claim — don't write it as "Ember's rates are a third of Haiku's", which is wrong.
Alexandr Wang's claim: Ember delivers nearly Muse Spark 1.3 (max) intelligence (AA 37 vs 36) at a 17× cheaper price per AA — Spark $2.39 per task vs Ember $0.14. Simon Willison's hands-on test calls Ember 40× cheaper than Wonder. The two multiples use different bases (per-task cost vs an unstated basis), so don't mix them. Note: Meta's chart notes Ember's costs were estimated at proposed pricing, and the AA cost comparison is Meta's reading of AA results, not third-party audited unit economics.
No source in this round — the Meta blog, the API docs, TechCrunch, or a dozen analyst pieces — mentions a free tier. Don't budget on free credits, and don't rule one out in future; check Meta's official pages. Google Cloud and Oracle Cloud pricing detail for Ember was not itemised in this round — check both providers' official pages; Vercel AI Gateway pricing is per the gateway's page.
With input at or below 200K tokens: $0.10 in / $0.004 cached in / $0.50 out per 1M tokens. Above 200K: the whole request bills at 3× ($0.30 / $0.012 / $1.50). Prices per Meta's official pages and the checkout page.
Per whole request: a request over 200,000 input tokens bills input, cached input and output all at 3× — not just the excess. Example: 200K input + 50K output totals $0.045; 250K input + 50K output totals $0.150.
Per task, yes ($0.14 vs $0.41); per token, no — the list prices are identical ($0.10/$0.50). The saving comes from Ember using fewer output tokens on the same test set (~115M vs ~298M). If your task makes Ember output as much as Haiku, the saving disappears.
No source mentions one. The Meta blog, API docs and a dozen analyst pieces in this round say nothing about free credits; check Meta's official pages.
Reference: Meta research blog — two-tier pricing and the headline claim · Artificial Analysis — cost per task $0.14 / $0.41 / $2.39 · tech-insider.org (cost math on the 200K cliff) · gptclub.ai (~60% cheaper cache) · agentfred.ai (output-token usage ~115M vs ~298M) (rewritten and organized by ModelBrief; worked examples computed from official list prices; facts per Meta's pricing page and the AA evaluation)