BenchOpen NavoThree rounds. One rooftop shot.
October 13, 2026 · 8 results
Same sheets and prompt · two results per model.
Within each test: highest complete AI rating, then lower historical cost. Partial panels stay unranked by rating. Lowest cost reverses that priority. Unknown values follow known ones. Ranges use their upper estimate. Human votes stay separate.
Video uses USD per delivered minute. Images use USD per generation until complete spend and human usability checks establish cost per usable image. Failed submissions and retries stay in the test history. Fees and taxes are unconfirmed.
Standard prompt
Both exact sheets were submitted, body first and face second.
Exact approved synthetic sheet submitted to every attempt.
Exact approved synthetic sheet submitted to every attempt.
Standard prompt
Both exact sheets were submitted, body first and face second.
Exact approved synthetic sheet submitted to every attempt.
Exact approved synthetic sheet submitted to every attempt.
Standard prompt
Both exact sheets were submitted, body first and face second.
Exact approved synthetic sheet submitted to every attempt.
Exact approved synthetic sheet submitted to every attempt.
Standard prompt
Both exact sheets were submitted, body first and face second.
Exact approved synthetic sheet submitted to every attempt.
Exact approved synthetic sheet submitted to every attempt.
Standard prompt
Both exact sheets were submitted, body first and face second.
Exact approved synthetic sheet submitted to every attempt.
Exact approved synthetic sheet submitted to every attempt.
Standard prompt
Both exact sheets were submitted, body first and face second.
Exact approved synthetic sheet submitted to every attempt.
Exact approved synthetic sheet submitted to every attempt.
Standard prompt
Both exact sheets were submitted, body first and face second.
Exact approved synthetic sheet submitted to every attempt.
Exact approved synthetic sheet submitted to every attempt.
Standard prompt
Both exact sheets were submitted, body first and face second.
Exact approved synthetic sheet submitted to every attempt.
Exact approved synthetic sheet submitted to every attempt.
Four models. Same Iris sheets and prompt.
Native sizes and baseline dates differ. Two attempts per route; no winner or usable rate yet.
Two attempts per route. Exact common prompt and reference bytes are unchanged, body first and face second. Google’s two standard-prompt October 12 outputs are reused without cherry-picking its guide arm. The other six jobs were interleaved OpenAI, FLUX, Qwen twice. No paid request was retried, repaired, upscaled or prompt-optimized. An absent model field in the first OpenRouter response paused dispatch; the saved response, charge and image were reconciled offline before continuing only unsubmitted jobs. OpenAI is pinned to its OpenAI provider; its response model field is unavailable, so identity is based on the requested endpoint contract. Model labels are visible before community voting. Prompt SHA-256: 8d53af2812ab95e42ffc50c65be0c4957600ecc360e1d9c5ea82763cedb956b5
New provider-reported OpenRouter API usage subtotal: $0.313310. FLUX/Qwen tariff estimate subtotal: $0.246180864; actual fal charges are unknown. Reused Google historical upper bounds total $0.339621 and incur no new generation charge here. Invoice, fees, taxes and refunds are unknown.
Native settings and sizes differ; Google is a historical baseline. Two attempts per route are descriptive, not a resolution-controlled ranking or universal winner. Human acceptance, visual scores and usable rates are pending. No video motion or lip-sync claim.
October 12, 2026 · 4 results
Same model, sheets and settings. Standard prompt versus Google’s prompt guide.
Within each test: highest complete AI rating, then lower historical cost. Partial panels stay unranked by rating. Lowest cost reverses that priority. Unknown values follow known ones. Ranges use their upper estimate. Human votes stay separate.
Video uses USD per delivered minute. Images use USD per generation until complete spend and human usability checks establish cost per usable image. Failed submissions and retries stay in the test history. Fees and taxes are unconfirmed.
Standard prompt
Both exact sheets below were submitted in the same order for all four images. Compare face and body shape separately.
Exact synthetic sheet submitted to every attempt. No display edit changes the model input.
Exact synthetic sheet submitted to every attempt. No display edit changes the model input.
Standard prompt
Both exact sheets below were submitted in the same order for all four images. Compare face and body shape separately.
Exact synthetic sheet submitted to every attempt. No display edit changes the model input.
Exact synthetic sheet submitted to every attempt. No display edit changes the model input.
Google guide
Both exact sheets below were submitted in the same order for all four images. Compare face and body shape separately.
Exact synthetic sheet submitted to every attempt. No display edit changes the model input.
Exact synthetic sheet submitted to every attempt. No display edit changes the model input.
Google guide
Both exact sheets below were submitted in the same order for all four images. Compare face and body shape separately.
Exact synthetic sheet submitted to every attempt. No display edit changes the model input.
Exact synthetic sheet submitted to every attempt. No display edit changes the model input.
Same Iris sheets. Two prompts, one model.
One Google model only. The separate Iris rooftop model comparison uses this standard prompt.
One frozen scene intent, one model, two identical reference images and identical native settings. Two common-template attempts and two official-guide attempts ran in the frozen interleaved order. No retries, prompt optimization, repairs or native-output changes. The initial executor preflight was unsubmitted; an incomplete usage breakdown paused the run after the first delivery. Both holds remain in the private history. Incomplete billing detail is conservatively bounded at the highest output-token rate; actual bills remain unknown. Native dimensions and aspect ratios are recorded per attempt. Model and prompt labels are visible before voting; these community votes are unblinded. Community votes and verified rubric scores remain separate.
Compare each result with both submitted sheets. Check head scale, face, body shape, framing and contact with the roof. Combined conservative API usage upper bound: $0.6848610; this is not a confirmed bill.
Owner acceptance, visual rubric scores, usable rate, actual charges, fees, taxes and refunds are unknown. Two attempts per prompt are descriptive. This does not establish a model winner, a prompt-guide benefit, video motion or lip sync.
October 10, 2026 · 1 result
An earlier scene with a different reference pack and prompt. Context for the later tests, not a matched comparison.
Exact sheet submitted to this attempt. Owner-approved body direction, 5′2″ / 130 lb.
Exact sheet submitted to this attempt. Owner-approved face and expression sheet.
One model. Different scenes and prompts.
Google Nano Banana Pro only. Changing scenes and prompt history do not measure model differences.
Judge face, head scale and body shape together. Historical shot directions are retained. Six former Flex attempts use Standard on the same Google model, explicitly labeled. These changing prompt arms are not a controlled model ranking.
Human acceptance, usable rate and invoice costs are unknown. Gravis’s body sheet has a three-quarter panel rather than a true left profile, and his face pack is not fully validated. Fifteen historical attempts remain unsubmitted: six need the exact original movie frame; nine need route checks, including five dependent scene inputs. Earlier votes are not pooled with this cohort.
We first revised the references, then compared two prompts, then tried four models with the standard prompt.
11 distinct images appear 13 times across these rounds. The two standard-prompt Google images are reused in the model comparison, with no new generation charge.
Original votes and AI scores remain attached to their recorded round and attempt. They are not pooled into a project score. Changed references and native settings limit comparisons; no final image or model winner has been selected.
Costs are recorded per result. Total settled project cost, fees, taxes and refunds are not confirmed.