Gemini 3.8 Flash review: Blender, knowledge videos, SVG and animation
Eleven creative tasks by Gemini 3.8 Flash, with inspected final artifacts, editable project downloads, concise observations and attributed Google benchmark figures.
Updated: 2026-11-04
Model and source
Google Flash model for software engineering, agents and knowledge work, accepting multimodal input and returning text.
Exact requested and returned model alias: gemini-3.8-flash. A verifiable immutable checkpoint was not published.
Google model card · Google model documentation · Metadata checked: 2026-11-03T23:54:21.091104+00:00
Scope and findings
All 11 tasks were delivered and inspected. SVGs and the bookshelf meet their main geometric and layout requirements reasonably well; the mechanical solids, interior circulation and detailing have substantial defects. All four knowledge films reach 1080p and 60–90 seconds, but caption timing, the QR correction example and unsupported source extrapolations affect reliability. Editable projects and source are preserved; geometry, captions and facts still need checking before use.
All eleven final deliveries were inspected. Accepted means delivered and checked, not a perfect quality score. The work is presented first, with run conditions and raw usage in the expandable evidence section.
View Gemini 4 Argon on the same frozen task suite
Budgets and cumulative SVG elapsed time
No controller caps on turns, cumulative output, retries or creation time. Each response allows up to 65,536 output tokens; individual network requests have a 600-second deadline and outage backoff is at least 60 seconds.
SVG elapsed spans, from first session creation to final completion: clock 538 s, dashboard 302 s, poster 265 s, cycling pelican 571 s. These include rendering, interruptions and waiting, with overlapping sessions; they are not isolated inference-speed measurements.
3D modelling
Parametric bookshelf and parameter revision
The wider four-compartment shelf is easy to compare with the three-compartment version in six consistent views. Some holes are raised cylinders, the backboard overlaps by about 1 mm, and the second version changes additional details.
More views and test details
Parametric bookshelf and parameter revision · view 2
Parametric bookshelf and parameter revision · view 3
Parametric bookshelf and parameter revision · view 4
Parametric bookshelf and parameter revision · view 5
Parametric bookshelf and parameter revision · view 6
Not recorded · Requests / complete responses: 68 / 66 · — / — input / output tokens
- Both versions provide a fully framed main view, rear view and clear joint close-up. Light wood grain, bevels and shelf supports are visible; the wider four-compartment version is easy to compare with the three-compartment version. The apparent holes are raised cylindrical details rather than drilled voids. The back panel overlaps the shelf bounds by about 1 mm, and V2 also changes hardware bevels and decorative-hole offsets, so the saved projects are not strictly a two-parameter-only change.
Itemized verification
- Delivery and independent inspection · pass
- Delivered and inspected; accepted does not mean perfect quality.
- Independent from-zero rebuild · not-tested
- Editable projects and reviewed source are retained; complete from-zero reconstruction was not independently verified. See the package notes for source gaps.
The frozen suite is retained. Only a brief and original prompt SHA-256 are public; full inputs are excluded. Historical task-source links are not Google model documentation.
Source brief and prompt
同一参数化书架制作两版：800/1000mm宽、三格/四格，保持深300mm、高1200mm和18mm板厚，交两工程六视图。 Two parametric shelf variants: 800/1000 mm wide, three/four compartments, retaining 300 mm depth, 1200 mm height and 18 mm boards; two projects and six views.
同一参数化书架制作两版：800/1000mm宽、三格/四格，保持深300mm、高1200mm和18mm板厚，交两工程六视图。 Two parametric shelf variants: 800/1000 mm wide, three/four compartments, retaining 300 mm depth, 1200 mm height and 18 mm boards; two projects and six views.
SHA-256: 8a48418a99ff6e645cedf0b6260a70644999588488c4ee5fb2c7a92cd4b0c297
Music and sound effects, no narration.
24/18/60-tooth planetary reducer and axial explosion
The tooth counts, center radius, 20 mm width and 3.5 ratio are measurable, with four views and a six-second film. Disconnected gear shells, bore interference, misplaced bridges, stacked UVs and an unreduced LOD limit asset usability.
More views and test details
24/18/60-tooth planetary reducer and axial explosion · view 2
24/18/60-tooth planetary reducer and axial explosion · view 3
24/18/60-tooth planetary reducer and axial explosion · view 4
24/18/60-tooth planetary reducer and axial explosion · view 5
Not recorded · Requests / complete responses: 98 / 98 · — / — input / output tokens
- The four views, editable project and six-second film with audio are delivered. Real 24/18×3/60 teeth, 42 mm planet-center radius, 20 mm width and 3.5 speed ratio are measurable. The main defects are disconnected gear shells, bore/shaft interference, misplaced carrier bridges, stacked UVs and an unreduced LOD; cropping and covered gears limit visual clarity.
Itemized verification
- Delivery and independent inspection · pass
- Delivered and inspected; accepted does not mean perfect quality.
- Independent from-zero rebuild · not-tested
- Editable projects and reviewed source are retained; complete from-zero reconstruction was not independently verified. See the package notes for source gaps.
The frozen suite is retained. Only a brief and original prompt SHA-256 are public; full inputs are excluded. Historical task-source links are not Google model documentation.
Source brief and prompt
制作并展示24/18×3/60齿行星减速器，核查尺寸、啮合与运动，交原生工程、多视图和六秒短片。 Create a 24/18×3/60-tooth planetary reducer with dimensional, mesh and motion review, editable project, multiple views and a six-second film.
制作并展示24/18×3/60齿行星减速器，核查尺寸、啮合与运动，交原生工程、多视图和六秒短片。 Create a 24/18×3/60-tooth planetary reducer with dimensional, mesh and motion review, editable project, multiple views and a six-second film.
SHA-256: edd3b7c6faa88fc7122a041f865c794a4ef20a76204d8d73492a334982976a9d
Music and sound effects, no narration.
12×9m two-bedroom interior, rooms and floor plan
Five views and a film show the rooms and furniture, with measurable dimensions and real openings. The empty en-suite, second-bedroom access, absent surface patterns and cropped plan weaken the spatial design.
More views and test details
12×9m two-bedroom interior, rooms and floor plan · view 2
12×9m two-bedroom interior, rooms and floor plan · view 3
12×9m two-bedroom interior, rooms and floor plan · view 4
12×9m two-bedroom interior, rooms and floor plan · view 5
12×9m two-bedroom interior, rooms and floor plan · view 6
Not recorded · Requests / complete responses: 91 / 91 · — / — input / output tokens
- The editable scene, five views and six-second film with audio are present. The envelope, wall thickness, genuine openings, table and six chairs are measurable. The empty en-suite, second-bedroom access and absent surface patterns are the main semantic defects; the plan crops the perimeter and furniture/film remain simplified.
Itemized verification
- Delivery and independent inspection · pass
- Delivered and inspected; accepted does not mean perfect quality.
- Independent from-zero rebuild · not-tested
- Editable projects and reviewed source are retained; complete from-zero reconstruction was not independently verified. See the package notes for source gaps.
The frozen suite is retained. Only a brief and original prompt SHA-256 are public; full inputs are excluded. Historical task-source links are not Google model documentation.
Source brief and prompt
制作12×9m两居室，包含指定房间、门窗和家具，交原生工程、五视图和六秒短片。 Create a 12×9 m two-bedroom interior with specified rooms, openings and furniture; editable project, five views and a six-second film.
制作12×9m两居室，包含指定房间、门窗和家具，交原生工程、五视图和六秒短片。 Create a 12×9 m two-bedroom interior with specified rooms, openings and furniture; editable project, five views and a six-second film.
SHA-256: 4b013e85ca2f476ee0c469cd168cd2a4554335ec2ba3817d2870308c10f2dbb5
Knowledge videos
Why does water float after it freezes?
The 75-second film uses genuine 3D assets to explain density and hydrogen bonds. Burned captions diverge around 21–30 seconds, the visuals lean on still cards, and measured loudness is about −32 LUFS.
More views and test details
Why does water float after it freezes? · view 2
Why does water float after it freezes? · view 3
Why does water float after it freezes? · view 4
Why does water float after it freezes? · view 5
Not recorded · Requests / complete responses: 61 / 61 · — / — input / output tokens
- Delivered a 75.325-second 1080p/30 fps film with actual Kokoro narration, stems, SRT, storyboard and rebuildable project. Genuine 3D assets support the core scientific explanation. Burned-in captions diverge around 21–30 seconds; visual explanation relies heavily on still 3D renders and cards, and the mix is quiet. These remain part of the recorded result.
Itemized verification
- Delivery and independent inspection · pass
- Delivered and inspected; accepted does not mean perfect quality.
- Independent from-zero rebuild · not-tested
- Editable projects and reviewed source are retained; complete from-zero reconstruction was not independently verified. See the package notes for source gaps.
The frozen suite is retained. Only a brief and original prompt SHA-256 are public; full inputs are excluded. Historical task-source links are not Google model documentation.
Source brief and prompt
围绕冻结主题“水结成冰，为什么反而浮起来？”制作60–90秒中文口播知识片，交字幕、分镜和工程。 Create a 60–90 second Chinese-narrated film about the frozen topic “Why does water float after it freezes?”, with captions, storyboard and project.
围绕冻结主题“水结成冰，为什么反而浮起来？”制作60–90秒中文口播知识片，交字幕、分镜和工程。 Create a 60–90 second Chinese-narrated film about the frozen topic “Why does water float after it freezes?”, with captions, storyboard and project.
SHA-256: 12267b7d9565f8dd3a2b70b96d4c7e64e65da3d930fa852ee48991e5dab7a871
Why can a damaged QR code still be read?
The 80-second film distinguishes codewords from image area and presents the four correction levels. Its key example changes DENSO’s 100+100 codewords to 150+50, while finder-pattern percentages and version comparisons lack verification.
Test details
Not recorded · Requests / complete responses: 74 / 74 · — / — input / output tokens
- Delivered and reviewed an 80.633-second 1080p/30 fps film with script, stems, SRT and rebuildable source. The four correction percentages and codeword-versus-area distinction are correct, but the official 200-codeword example is incorrectly changed from 100+100 to 150+50. Finder-pattern failure percentages and version comparisons are unverified. The mix is quiet, caption timing is approximate, and no listening review was performed.
Itemized verification
- Delivery and independent inspection · pass
- Delivered and inspected; accepted does not mean perfect quality.
- Independent from-zero rebuild · not-tested
- Editable projects and reviewed source are retained; complete from-zero reconstruction was not independently verified. See the package notes for source gaps.
The frozen suite is retained. Only a brief and original prompt SHA-256 are public; full inputs are excluded. Historical task-source links are not Google model documentation.
Source brief and prompt
围绕冻结主题“二维码缺了一块，为什么还能扫？”制作60–90秒中文口播知识片，交字幕、分镜和工程。 Create a 60–90 second Chinese-narrated film about the frozen topic “Why can a damaged QR code still be read?”, with captions, storyboard and project.
围绕冻结主题“二维码缺了一块，为什么还能扫？”制作60–90秒中文口播知识片，交字幕、分镜和工程。 Create a 60–90 second Chinese-narrated film about the frozen topic “Why can a damaged QR code still be read?”, with captions, storyboard and project.
SHA-256: 943b6e46441d1ff980ca010560dcbed194559981a0ff67253fa8fb648f291a81
Why does Webb use five thin sunshield layers?
The 83-second film makes the five-layer shield and lateral radiative cooling clear. Its NASA comparison matrix adds unsupported details, thick-plate launchability claims overreach, and the audio plan has a start-time discrepancy.
Test details
Not recorded · Requests / complete responses: 55 / 55 · — / — input / output tokens
- Delivered and reviewed an 83-second 1080p/30 fps film, stems, actual 80.375-second Kokoro narration and rebuildable source. The layer-and-radiation mechanism is clear, but the NASA comparison matrix adds unsupported specifics and the thick-plate launchability claims overreach. Burned captions match the SRT; the audio plan has a 0.5-second offset discrepancy. No listening review was performed.
Itemized verification
- Delivery and independent inspection · pass
- Delivered and inspected; accepted does not mean perfect quality.
- Independent from-zero rebuild · not-tested
- Editable projects and reviewed source are retained; complete from-zero reconstruction was not independently verified. See the package notes for source gaps.
The frozen suite is retained. Only a brief and original prompt SHA-256 are public; full inputs are excluded. Historical task-source links are not Google model documentation.
Source brief and prompt
围绕冻结主题“韦布为什么用五层薄膜，而不是一块厚板？”制作60–90秒中文口播知识片，交字幕、分镜和工程。 Create a 60–90 second Chinese-narrated film about the frozen topic “Why does Webb use five thin sunshield layers?”, with captions, storyboard and project.
围绕冻结主题“韦布为什么用五层薄膜，而不是一块厚板？”制作60–90秒中文口播知识片，交字幕、分镜和工程。 Create a 60–90 second Chinese-narrated film about the frozen topic “Why does Webb use five thin sunshield layers?”, with captions, storyboard and project.
SHA-256: d81a72e59baa97ffeb2adf3c9798985302a538809de3cb1b57ccd880baf328cb
Why does noise cancellation work without music?
The 84.5-second film explains ANC without music through moving waveforms and meters, with residual-sound and idealization labels. Text is dense, dB/energy/health claims exceed the frozen Bose source, and loudness is below the plan.
Test details
Not recorded · Requests / complete responses: 75 / 75 · — / — input / output tokens
- The 84.5-second film explains ANC without music using six graphic layouts, moving waveforms/meters and 21 captions, with idealization and residual-sound notes. Full decode and segment timing checks pass. Dense text and dB/energy/health assertions beyond the frozen Bose source are the main limitations. Loudness is below its stated plan; actual listening and pronunciation remain unverified.
Itemized verification
- Delivery and independent inspection · pass
- Delivered and inspected; accepted does not mean perfect quality.
- Independent from-zero rebuild · not-tested
- Editable projects and reviewed source are retained; complete from-zero reconstruction was not independently verified. See the package notes for source gaps.
The frozen suite is retained. Only a brief and original prompt SHA-256 are public; full inputs are excluded. Historical task-source links are not Google model documentation.
Source brief and prompt
围绕冻结主题“降噪耳机没放音乐，为什么也能安静？”制作60–90秒中文口播知识片，交字幕、分镜和工程。 Create a 60–90 second Chinese-narrated film about the frozen topic “Why does noise cancellation work without music?”, with captions, storyboard and project.
围绕冻结主题“降噪耳机没放音乐，为什么也能安静？”制作60–90秒中文口播知识片，交字幕、分镜和工程。 Create a 60–90 second Chinese-narrated film about the frozen topic “Why does noise cancellation work without music?”, with captions, storyboard and project.
SHA-256: b905c7799bf1a66b13ec6948a8272b3ac837556f4179182bab3a9c5cbdf2f360
SVG and animation
An analog clock at 9:41
All twelve numerals, the 9:41 hand positions and the 09:41 badge are correct. The metal rim and light dial are clear, with readable text and intact margins.
Test details
Not recorded · Requests / complete responses: 19 / 18 · — / — input / output tokens
- The metal rim and light dial are clear. All twelve numerals and the 09:41 badge are readable; the shorter hour hand sits between 9 and 10 and the longer minute hand points at minute 41. No clipping or text overlap observed.
Itemized verification
- Delivery and independent inspection · pass
- Delivered and inspected; accepted does not mean perfect quality.
- Independent from-zero rebuild · not-tested
- Editable projects and reviewed source are retained; complete from-zero reconstruction was not independently verified. See the package notes for source gaps.
The frozen suite is retained. Only a brief and original prompt SHA-256 are public; full inputs are excluded. Historical task-source links are not Google model documentation.
Source brief and prompt
800×600静态SVG钟表，十二数字齐全，正确显示9:41及盘下09:41标签。 An 800×600 static SVG clock with all twelve numerals, correct 9:41 hands and a 09:41 label.
800×600静态SVG钟表，十二数字齐全，正确显示9:41及盘下09:41标签。 An 800×600 static SVG clock with all twelve numerals, correct 9:41 hands and a 09:41 label.
SHA-256: b57e2e07903e6a1290e19fe05dba2fcca4c63ef9d62f85970eeb52262d5def8c
Data chart and proportional bars
The four bars share zero, match the 12/24/18/30 proportions and have clear summary labels. Added unit and footer copy are incomplete.
Test details
Not recorded · Requests / complete responses: 16 / 15 · — / — input / output tokens
- A clear light card with blue-to-purple bars and readable values. The four bars share zero and are exactly proportional to 12/24/18/30; Total: 84 and Best: Q4 are clear. Added unit and footer copy are incomplete.
Itemized verification
- Delivery and independent inspection · pass
- Delivered and inspected; accepted does not mean perfect quality.
- Independent from-zero rebuild · not-tested
- Editable projects and reviewed source are retained; complete from-zero reconstruction was not independently verified. See the package notes for source gaps.
The frozen suite is retained. Only a brief and original prompt SHA-256 are public; full inputs are excluded. Historical task-source links are not Google model documentation.
Source brief and prompt
800×600静态收入图表，Q1–Q4为12/24/18/30，柱高成比例、共用零基线，包含指定汇总标签。 An 800×600 revenue dashboard with Q1–Q4 values 12/24/18/30, proportional bars, shared zero and required summary labels.
800×600静态收入图表，Q1–Q4为12/24/18/30，柱高成比例、共用零基线，包含指定汇总标签。 An 800×600 revenue dashboard with Q1–Q4 values 12/24/18/30, proportional bars, shared zero and required summary labels.
SHA-256: f548fe1b75c99fa4c4ffafbcf359b4b619413ba1890d6a0bb6bf826c7249400b
Chinese poster and layout constraints
The six required Chinese lines are intact, with a clear title, time/location and call-to-action hierarchy. The bicycle is recognizable and the spacing is generous.
Test details
Not recorded · Requests / complete responses: 12 / 12 · — / — input / output tokens
- A light green poster places clear text on the left and a detailed bicycle on the right. Each of the six required lines appears once, Chinese glyphs are intact, and title, time/location cards and call to action have a clear hierarchy. Margins are generous, with no clipping or text overlap observed.
Itemized verification
- Delivery and independent inspection · pass
- Delivered and inspected; accepted does not mean perfect quality.
- Independent from-zero rebuild · not-tested
- Editable projects and reviewed source are retained; complete from-zero reconstruction was not independently verified. See the package notes for source gaps.
The frozen suite is retained. Only a brief and original prompt SHA-256 are public; full inputs are excluded. Historical task-source links are not Google model documentation.
Source brief and prompt
800×600浅色中文骑行活动海报，准确包含指定六段文字及自行车图标，检查字号层级与边界。 An 800×600 light Chinese cycling poster with exactly six required text lines, a bicycle, readable hierarchy and margins.
800×600浅色中文骑行活动海报，准确包含指定六段文字及自行车图标，检查字号层级与边界。 An 800×600 light Chinese cycling poster with exactly six required text lines, a bicycle, readable hierarchy and margins.
SHA-256: c56b89af23096d28d1a7bd68a32a443e3b6290bad63cda09332f4e75890bc1ab
Animated cycling pelican
Spokes, crank and both legs move while feet and hips stay attached, with a continuous four-second loop. The near leg looks long and the knee bend is stiff in some phases.
Test details
Not recorded · Requests / complete responses: 22 / 21 · — / — input / output tokens
- The orange bill and throat pouch, white pelican and teal bicycle are recognizable, with clear wheels, frame, saddle, handlebar and basket. Spokes, crank and both legs visibly move; feet remain on the pedals and hips remain attached to the body. Moving clouds and road add travel, while the rider and bicycle remain on canvas. The 0/4-second boundary differs only by negligible pixels and the loop looks continuous. The near leg appears long and its knee bend is stiff in some phases.
Itemized verification
- Delivery and independent inspection · pass
- Delivered and inspected; accepted does not mean perfect quality.
- Independent from-zero rebuild · not-tested
- Editable projects and reviewed source are retained; complete from-zero reconstruction was not independently verified. See the package notes for source gaps.
The frozen suite is retained. Only a brief and original prompt SHA-256 are public; full inputs are excluded. Historical task-source links are not Google model documentation.
Source brief and prompt
800×600纯SMIL鹈鹕骑车SVG，四秒循环，轮辐、曲柄和连接身体与脚踏的双腿实际运动。 An 800×600 SMIL pelican cycling SVG with a four-second loop, moving spokes/crank and legs attached to body and pedals.
800×600纯SMIL鹈鹕骑车SVG，四秒循环，轮辐、曲柄和连接身体与脚踏的双腿实际运动。 An 800×600 SMIL pelican cycling SVG with a four-second loop, moving spokes/crank and legs attached to body and pedals.
SHA-256: 09dbdcdf3c5204aafd8fc24c1020a7c8f27c77fdb4124f75a7e83458e0fd83c4
Download all creative results and source index
Run settings, usage and full notes
Gemini 3.8 Flash · 2026-11-04
creative-suite-v4: three Blender tasks, four date-locked knowledge videos, three static SVGs and one SMIL animation. The model independently authors source through the Google Developer API-compatible tool loop; local renderers generate media. The coordinator orchestrates, inspects and packages. No thinkingLevel was forced; the documented default is MEDIUM, without per-request independent measurement. Knowledge narration uses the fixed local Kokoro zf_xiaobei voice (lang z); Gemini-native speech generation was not evaluated. Blender films use procedural music and effects, without narration.
The API did not return actual charges or credit deductions. Actual monetary cost is unknown; no price estimate substitutes for billing.
Visual observations are by AI; dimensions, media and audio checks use scripts. No human scoring or actual listening is claimed. Accepted means delivery and inspection; quality defects remain in the case notes.
Native P/C/T, cache and explicitly returned reasoning are retained per request in the public ledger. C is not guaranteed to include all hidden thinking; missing is not zero and T−P−C is not inferred as reasoning. Complete input/output totals are unknown and shown as —; known subtotals are not a full bill.
Visual commentary by AI; dimensions and media checks use scripts.
Requests / complete responses: 591 / 586 · Errors / interruptions: — / —
— / — input / output tokens, including cache and thinking
- Full task inputs, private continuations, event logs and hidden thinking are excluded.
- Source ZIPs are reviewed public derivatives: TTS metadata sanitization and optional helper exclusions are listed. Artwork, audio and other author member bytes are retained.
- Native projects retain author-embedded scripts. Complete from-zero reproducibility and human listening are not certified.
- Model identity is the exact returned provider alias, without a published verifiable immutable checkpoint.
Original Google benchmark figures
Four vendor-reported benchmark categories are shown using the original light-theme figures. TOKRACE did not independently reproduce these results. Figure count is not a count of additional local tests. Original image bytes, axes, footnotes and attribution are retained.
Google uses pass@1 for most Flash metrics. DeepSWE uses mini-swe and high thinking; HLE-Verified includes revised items and content-filter caveats for some comparator scores.
Gemini 3.8 Flash · DeepSWE V1.1
Google source page · Original image · 2026-11-03
© Google / Google DeepMind
Gemini 3.8 Flash · Vals Finance Agent v2
Google source page · Original image · 2026-11-03
© Google / Google DeepMind
Gemini 3.8 Flash · Harvey’s Legal Agent Benchmark
Google source page · Original image · 2026-11-03
© Google / Google DeepMind
Gemini 3.8 Flash · HLE-Verified
Google source page · Original image · 2026-11-03
© Google / Google DeepMind