A model can be cheap on the first turn and unexpectedly expensive by the twentieth. Each result pulled into a tool loop can expand the next request. A model that finishes an early step correctly might also be the wrong choice for reviewing the exceptional cases that arrive afterward. To decide where a smaller model belongs, trace what the workflow actually sends and what the next step receives.
Meta’s November 9 Muse Ember announcement gives teams a timely reason to do that analysis. Meta positions Ember for recurring agent tasks and says it is the first model trained using Muse Wonder Full. The release describes availability in Muse Code and on the Meta Model API as muse-ember-1.0. The public API documentation says it is Standard mode only and defaults to maximum reasoning effort. These are listed product conditions, not evidence of particular account entitlements, or a release of Ember’s weights. Model documentation
A larger request can change the rate for the whole call
At up to 200,000 input tokens, the published prices per million are $0.10 for input, $0.004 for cached input, and $0.50 for output. Above that threshold, the rates are $0.30, $0.012, and $1.50 respectively. All input, including cached tokens, contributes to determining the band. On a continued conversation, retained input sent through previous_response_id contributes too. The higher band applies to the entire request rather than only the excess. Meta API pricing
For an arithmetic example, assume no cache or separately charged tools, with 5,000 output tokens in both requests. At 195,000 input tokens, the token charge is $0.022. At 205,000 input tokens, it is $0.069. The extra 10,000 tokens are not the main reason for the difference: the whole request moves to another rate. These values are calculated examples, not charges observed in an account.
In a real run, cached input can reduce a bill, but it does not keep a long request in the lower band. A record of the visible user message is also insufficient to predict the bill. The system prompt, tool results, and retained state may be a large part of the input. Compare actual usage by step, recording the share of rounds in each band and the effective cost of each branch, retry, and escalation. Separate the latency effect of a longer context from its billing effect.
Strict completion and partial reward describe different outcomes
Meta’s release shows Ember at 37.8% on AutomationBench-AA, versus 13.9% for Haiku 5.5 and 20.2% for GPT-6 Luna, all at maximum reasoning. The methodology identifies the metric as strict pass@1 on simulated business workflows: all objectives must be met with no guardrail violation. The same launch comparison puts Ember behind Haiku on FrontierCode 2.0, a different assessment concerned with whether engineering changes are mergeable. Launch results, evaluation methodology
The methodology also distinguishes internal common-loop comparisons from external results that retain different coding products or harnesses. Its OSWorld 2.1 method gives partial credit as its main score and reports strict completion separately. Infrastructure or grading errors can be excluded under the general methodology, with deviations documented by benchmark. A published result is therefore a reason to select a pilot workload, not an estimate of your own complete work or failure rate. Evaluation methodology
Use two comparisons if the decision warrants both. First compare models in the same loop to learn what changes when the model changes. Then compare the full configurations you could actually deploy, with their intended prompts, tools, retries, and recovery. The first helps isolate a cause; the second establishes what a buying or deployment decision would change. Treat maximum effort as one setting. Even when every vendor calls it maximum, the configurations can consume different token budgets and take different amounts of time.
Make the handoff small enough to audit
Consider a synthetic vendor-record maintenance run. Several attachments need to be read, candidate fields need to be reconciled, and a human may eventually approve an update. One approach adds the complete history to every extraction call. Another creates a bounded handoff with the relevant page, field definition, supporting evidence, and a place to mark uncertainty. The latter could use less input; it could also miss a qualification on an omitted page. Test that loss explicitly before shrinking context to keep the rate down.
For the pilot I would include ordinary records, documents with prior revisions, and deliberately conflicting but harmless fixture data. Have the evaluator check each field against a frozen reference and verify that uncertain cases reach the right reviewer. Keep approval attached to the eventual action and check it at the application boundary. That prevents an attractive extraction score from obscuring automated writes on drafts that have not been verified.
For borderline long tasks, compare three designs: preserve the context and pay the higher rate, provide smaller evidence packages and test for information loss, or route the job to a different model with an appropriate evidence budget. A cheaper route still needs to carry the constraints required to do the work. Track the time spent reconstructing context in every condition. If reviewers have to reopen every source that the extractor omitted, an apparent token saving can become human work.
Run an evaluation on the complete workflow
An unrun comparison I would propose uses the incumbent workflow as a control, then tests Ember in the same loop and Ember only for bounded subtasks. Freeze inputs, tool behavior, retry policies, and acceptance criteria. Vary reasoning effort within the configured model and measure effects on particular task groups. Define how to count tool errors and missing grades before running the study.
| Observation | Practical question |
|---|---|
| Correctly completed workflows and total assigned workflows | Did the system deliver more useful work, including cases it refused or left unfinished? |
| Input distribution by step, cache share, and pricing band | Which requests make the model cheap or expensive on this workload? |
| Escalation, retries, and human correction time | Did the new route remove work or transfer it? |
| Last-step latency and unresolved exceptions | Does the successful average hide a slower or riskier tail? |
For a decision, divide the full operating cost by the number of independently accepted outcomes. Keep the failure rate and unresolved external effects beside that denominator. Start with a candidate task family whose evidence is compact and whose output can be checked cheaply. Ember’s prices and reported evaluations make it worth testing there; a broader rollout should follow the workflow evidence.