Choosing a model should start with what you need from your agent. It should understand you, answer useful questions, plan work, operate your tools and bring the job to completion.
AmolfiBench brings model choice into view through Work, Gen and Voice. Work considers the primary model you would leave your Amo on. Gen and Voice consider the quality of their outputs. Letter grades describe the standard of the work and the guidance it needs.
A request ends with useful work.
Choose a kind of work. The brief, expected output, and review change together.
Turn a client brief and a folder of source material into a six-slide pitch.
A presentation someone can actually send, with claims grounded in the source files.
Story, factual accuracy, visual hierarchy, and whether revisions follow the brief.
Each task pairs a brief with a deliverable and review criteria.
Start with what the person needs
A useful result respects the request, the source material and the boundaries of the task. Speed alone cannot compensate for a missing deliverable or a decision that should have stayed with the person.
The brief. The result. The judgment.
Quality is more than a plausible answer
Consider the finished artifact and the path taken to produce it. Some qualities are directly checkable: a formula can be recomputed and a file can be opened. Others need judgment against the brief. An elegant response that leaves the task undone is still incomplete work.
Facts, formulas and source material stay correct. Images preserve identity; video preserves continuity.
The requested work is complete, relevant and suited to its audience.
Clear writing, deliberate design, legible images and convincing motion.
The result opens, works and is practical to use or edit.
Think of a professor grading the work.
An excellent assignment fulfills the brief with nothing meaningful to correct. An exceptional one makes a capable professor ask, “How did you do that?” The same distinction applies to a finished document, an image, a video or a voice interaction.
- SBeyond expectations
- Goes beyond the brief with skill or insight a capable evaluator could not have produced. The surprise is in the substance, not just the presentation.
- AExcellent
- Fulfills the brief with judgment, accuracy and polish. No meaningful complaints or required corrections.
- BStrong with guidance
- Produces good, useful work with clear direction. Needs focused feedback or editing to reach an excellent result.
- CUneven
- Has useful strengths and obvious gaps. Needs substantial guidance and revision; choose its tasks carefully.
- DLimited
- Can serve a narrow purpose, but important gaps and major rework make it a weak general choice.
- FNot recommended
- Fails the essential brief or produces unusable work. There is no compelling reason to choose it for this category over better alternatives.
The result and the help it needed
Good judgment, fidelity to the brief, craft and dependable delivery all matter. A polished result with important errors still needs correction. A model that succeeds only after repeated rescue earns a different grade from one that gets there with clear initial direction.
S is exceptional, not simply first.
S means the work demonstrates skill or insight beyond what a capable evaluator could have produced. It can be earned today, but a category can have several S models or none. It does not mean perfection on every possible task or establish superintelligence.
Grades have an edition.
The current rubric is September 2026. Grades reflect the tasks and expectations of that edition. As the work changes, a new edition can ask harder questions without rewriting an earlier judgment. Plus and minus show a model’s placement within its tier; they are not hidden success percentages.
A grade describes the category and current model shown. It is a considered judgment, not a guarantee that every result will meet the same standard.
Quality, with its cost in view
Choose Work, Gen or Voice. Each point represents a current model, with a letter grade and estimated cost. Superseded versions leave the comparison when their replacement is available. Selecting a model in the table shows or hides its point. The 2D view pairs grade with cost. The rotatable 3D view adds speed where comparable measurements are available.
Work considers understanding and context, judgment and planning, delegation, verification and recovery, direct computer use, and follow-through. A primary agent can answer you directly, operate a computer, or brief a specialist. It remains responsible for checking the result and completing the request. More workers or more reasoning do not automatically mean better work.
These are dated, provisional editorial judgments informed by published exact-model task evidence. Existing letter tiers remain cautious starting placements under the primary-agent definition. Professional artifacts, research, workflows and coding inform suitability; they do not directly establish reliable worker supervision or an ongoing conversation. No matched primary-Amo evaluation is represented. The evidence and gaps are available with each model.
Reasoning follows the model and the work.
We use an explicit reference configuration for each model where comparable costs exist. It is an editorial starting point or a coverage-limited published setting, not a claim about the setting people use most. Other settings remain supporting evidence. We do not invent percentages for Low, Medium, High or Max, and the same label does not imply equal capability or computation across models.
A complete Auto comparison would use the same task mix, tools and specialist pool, with a documented routing policy calibrated for each model. It would judge the final outcome and include the lead, router, workers and retries in spending and completion time. That system-level comparison has not been measured here. A model’s ability to adapt its thinking inside a fixed setting is also distinct from a router changing the setting.
Work cost compares published evaluator consumption at the named configurations on professional-deliverable, workflow and coding tasks. Each workload is normalized to GPT-6.1 Sol Medium before combining the ratios; 1× is that reference. The cost-reference mix weights professional deliverables 55%, workflows 20% and coding 25%; those are not primary-agent grading weights. It excludes a matched computer-use family, routing, workers, infrastructure and human review. The provisional broad suitability grade is not a measured quality score at the cost-reference setting.
Medium is the routine reference for Astra, Sol, Opus and Sonnet; Fable uses High following provider guidance. Luna uses High to match its narrow office evidence. Grok High and Muse Extra High are coverage-limited references, rather than inferred optimal routine settings. Other rows identify their published configuration. Missing costs stay missing. The 3D speed reference uses matching output-throughput observations; it does not measure agent-session completion time. Gen and Voice retain their image, short-clip and generated-minute references.
Explore AmolfiBench