Where to start with published results
Meta’s announcement is the primary source for its own Wonder and Full evaluations. It discusses memory, sustained task execution and coding, including FocusBench and SWE-Marathon. These are provider-reported evaluations, not hands-on results produced by this site.
Read Meta’s original release and evaluation context ↗
For an independent evaluation, prefer the evaluator’s own model page and methodology over a copied screenshot. We have not verified a current Artificial Analysis score or a universal aggregate rank for this guide. A blank field should remain unverified, rather than being replaced with an inferred score.
A checklist for comparable scores
| Record | Why it matters |
|---|---|
| Exact model variant | Regular Wonder and Full are different entries |
| Reasoning effort | A max-effort run may have a different budget |
| Harness and tool access | The surrounding agent can change task outcomes |
| Benchmark version and date | Task sets and implementations evolve |
| Attempts and selection rule | Best-of-many and first-attempt results differ |
| Latency and total spend | Success alone does not describe operating cost |
Do not compare an Elo rating directly with an accuracy percentage or treat different benchmark revisions as one scale. A model that performs well on a coding suite is not thereby proven best for legal, creative or private-data tasks.
Design a small evaluation for your workload
- Choose representative tasks before looking at the model outputs.
- Write a rubric that checks correctness, required sources and unacceptable errors.
- Freeze input material, tool permissions and stopping conditions.
- Run a small controlled set and retain unsuccessful attempts.
- Review outputs with the same criteria and report the sample size.
- Calculate spend per accepted result, including retries and tool calls.
For example, suppose a test budget spends $12 and six tasks meet your prewritten acceptance criteria. Cost per accepted task is $2. These are hypothetical inputs illustrating a formula; they are not Wonder measurements. Use your own measured costs and accepted outcomes.
Reviews and Reddit discussions
User discussions can reveal workflow friction, but they do not substitute for a controlled evaluation. Separate a user’s actual task, model variant and date from a general impression. Do not promote an anonymous score, an unattributed benchmark chart or an affiliate review into an official result.
When a search specifically asks for Reddit, the discussion platform is the intended destination. This page does not invent quotes or pretend to be a Reddit thread. It explains what information to look for before using any review in a purchase decision.
Use Full versus max effort to identify the variant, then the same-workload pricing comparison to build a budget. Actual performance still requires testing the tasks you plan to run.
Sources & evidence
Editorial check: 17 October 2026. Product facts come from the following primary sources; planning checklists and examples are our explanations. No paid inference, model benchmark or regional checkout was run for this guide.