Different jobs need
different evidence.

Current evidence can change the recommendation now.

Technical Coding can receive an * Early evidence — position may change ranking as soon as current external benchmarks and original hands-on comparisons support it. Writing approvals apply only to Fiction Prose and Screenwriting; they do not hold up Coding, Image, Video or Builder reviews.

We identify the model, reasoning effort and coding tools, review the full eligible field, and require recent inspectable comparisons from at least two qualified publishers, including an objective benchmark. Different effort levels and harnesses remain separate observations. Provider claims alone cannot promote a model, and a creator’s articles and videos count as one publisher. Sponsorship and pre-release access are disclosed when known.

The five-factor coding rubric still guides the editorial decision. These factors are judgments informed by evidence, not measured benchmark percentages; we hide the aggregate score on early recommendations to avoid false precision. We explain conflicting findings and what could change the order, then reassess within two days. A benchmark win in code generation does not automatically prove superiority in debugging, refactoring or autonomous work.

Compare the task and the tested version.

We review published blind comparisons and hands-on tests from their original publishers. Every useful observation identifies the task, model version, settings when disclosed, result, and limitations. Sunburst and Flare, base and post-trained models, and 720p and 1080p tests are not interchangeable.

Preference leaderboards show what voters preferred in that test. They do not establish a universal winner for photographic realism, artistic taste, or complex filmmaking. Provider documentation establishes capabilities and access; it does not count as an independent quality comparison. Conflicting findings stay visible, and multiple articles or videos from one publisher count as one source.

New evidence since this ranking contains observations reviewed after the recommendation’s verification date. They may support or challenge it. Adding evidence does not automatically change the order, score, or verification date: those require a documented category reassessment. Rankings with generic or incomplete comparative support remain provisional.

A winner, a runner-up, and a third choice.

BMN publishes an editorial first, second and third place. We make a recommendation from the available evidence, explain why each option earned its place, and state how strongly that order is supported. These are judgments you can inspect, not invented numerical quality scores or claims of universal superiority.

We assess the complete eligible field against the category's primary criteria. When quality is effectively tied, we compare consistency, instruction fidelity, creative control, and breadth, in that order. If these still do not separate the contenders, we make a clearly identified editorial call and explain the deciding tradeoff. Each review states why first beats second, why second beats third, and why the strongest excluded challenger missed the podium.

Quality and editing categories keep price and access separate from the order. Best Overall includes practical workflow, access and value after quality. Capabilities assesses demonstrated usefulness and reliability, not advertised feature counts. Multiple versions from one family can place when their exact configurations earn it.

Video quality uses a consistent comparison.

For text-to-video, we consider visual realization, believable motion and physics, temporal coherence, scene composition and meaningful direction together. Literal camera adherence is important, but does not automatically outweigh the quality of the finished scene. Missing subjects, merged bodies and other artifacts remain genuine failures. Audio-inclusive preference and price do not decide visual quality.

Each adjacent placement explains the scene and attempt coverage and the tested configurations. A handful of 720p generations cannot establish performance across a model’s best available settings. We preserve those limits and never assume a higher-resolution option fixes observed failures.

Optional owner editorial judgment.

Trevor Mayes may contribute a dated filmmaking assessment with his preferred winner, reasoning and viewing basis. When this influences a published order, we identify it publicly and preserve contrary external evidence. It is first-party editorial judgment, not an independent benchmark, blind BMN test or hidden numerical bonus. A decisive owner preference carries limited confidence. Future research must explicitly address it and may reach a different conclusion with stated reasons; it is not a permanent model-specific rule. Trevor can revise or withdraw it for future reviews.

Image-to-video means rendered visual quality.

We compare motion and physics, source identity and composition, temporal stability, detail and artifacts, and instruction fidelity. Single-image animation and multi-reference filmmaking are distinct tests. Price, generation speed, access and audio preference do not decide this quality ranking.

A narrow Elo lead in an audio-inclusive or mixed-resolution pool cannot establish a visual-quality win. We examine original hands-on comparisons alongside those benchmarks, preserve contrary findings, and state differences in resolution, host, input mode and retries. Unmeasured dimensions stay unmeasured. An unresolved comparison can support a limited-confidence editorial choice, not a claim of demonstrated superiority. The same requirements apply to every model.

Photorealism includes ordinary photographs.

A believable phone snapshot can be as successful as a polished studio portrait. We judge the kind of photograph requested: everyday phone snapshots; low-light, flash and motion on phones; professional portraits and editorial; product and commercial photography; documentary, street and events; landscape and architecture. Appropriate noise, flash, softness or imperfect framing is not an AI defect. Unrequested glamour lighting and glossy skin can make a casual photograph less convincing. Impossible anatomy, objects and reflections remain defects.

Each Photorealism review records coverage and gaps across all six photo types. Repeated evidence about professional portraits cannot establish casual-phone realism. Incomplete coverage lowers confidence in the overall order; it does not require an unordered list.

Video editing must preserve what you meant to keep.

We check the requested change, natural skin and faces, unchanged details, motion and consistency across attempts. Tests of newly generated scenes do not prove editing quality. An audio-inclusive vote lead does not establish better visual detail; missing comparisons remain explicit.

Confidence belongs beside the recommendation.

Limited, moderate and strong describe confidence in our editorial ordering, not a calibrated probability. A missing benchmark entry is not a loss. We retain contrary findings, exact versions, resolutions, input modes and original source dates. Two independent comparative publishers must support the selected field, and every selected configuration needs relevant direct comparative evidence. Strong confidence additionally requires two independent direct publishers supporting the winner; incomplete photo-type coverage prevents strong confidence in Photorealism.

An unresolved contender must stay visible.

Eligible contenders awaiting comparative evidence are named beside the recommendations across images, video, technical coding, app builders, General Writing and Business Writing. Being included in the field is different from having a completed comparison. Missing a leaderboard entry does not exclude a model, and it does not establish that the published choices are better. Fiction Prose and Screenwriting retain their separate blind-test and owner-review process.

We track how long each comparison has remained unresolved. After seven days it becomes priority research, even if the category or its sources were checked recently. The next step must seek the missing task-specific evidence through original hands-on comparisons or accessible original articles, transcripts and outputs. Repeating an availability check does not restart that clock. Different settings can inform a clearly limited editorial judgment when their effects are disclosed; we do not require every model to appear in one perfectly identical benchmark.

Escalation does not automatically promote a model or authorize paid testing. Source independence, exact-version relevance and publication safeguards still apply. Until evidence supports a new order, the existing recommendation remains provisional and the unresolved contenders stay visible.

A category review can also publish useful initial findings about a contender without assigning a new rank or numerical score. These findings state task-specific strengths, limitations and the next comparison needed. General Writing and Business Writing can publish supported provisional rankings independently of the creative-writing approval process; they still require evidence for the complete eligible field and the five scoring factors. Technical coding retains its exact-configuration benchmark and corroboration requirements, and builders retain their product-demonstration and consensus requirements.

Hands-on comparisons with affiliate, sponsored, free-access or undisclosed relationships may supplement the review with their limits stated, but never meet the independence minimum. Historical unordered reviews remain available for rollback; a new numbered recommendation requires a documented reassessment. Rereading a test does not make it new.