GPT Image 2 Quality Tier Comparison 2026: High vs Low

Key Takeaways

• GPT Image 2 high scores 4.16 on VibeDex's Sonnet 4.6 blind benchmark — the best of the three quality tiers.
• Medium (4.11) is the cost-effective default: 99% of high's score for roughly a quarter of the price.
• Low (3.95) is still strong — best for drafts, exploration, and high-volume batches rather than final renders.
• Pricing runs $0.014 (low), $0.055 (medium), $0.212 (high) per 1024×1024 image — a 15x spread from cheapest to most expensive.
• High beats medium on just 22 of 50 prompts head-to-head, with an average quality gap of only 0.05 points — so high is worth it for hero assets, not batch work.

What the GPT Image 2 Quality Parameter Controls

GPT Image 2 builds a picture as a sequence of image tokens, and the quality parameter sets how many of those tokens the model spends. More tokens buy finer texture, sharper edges, and more accurate embedded text — at the cost of higher latency and a bigger bill.[1] The parameter moves four things at once: visual detail, text rendering, latency and throughput, and cost.

At 1024×1024 the jump from low to high is roughly 15x the tokens — and GPT Image 2's per-image price tracks that almost exactly ($0.014 → $0.212). The question the token table can't answer is what that 15x spend actually buys. Our blind benchmark can: across the full low-to-high range, the quality score rises only 0.21 points. OpenAI frames the parameter as a token tradeoff; the sections below show what that tradeoff is worth in practice.

GPT Image 2 Benchmark: How We Tested the Quality Tiers

Every score in this GPT Image 2 review comes from VibeDex's independent blind benchmark — not OpenAI's marketing claims, and not our own eyeballing. We ran the same 50-prompt set through all three quality tiers and judged every render in three independent Claude Sonnet 4.6 passes — 150 judgments per tier — across four dimensions: visual fidelity, physics and logic, subject and object integrity, and instruction adherence. The judge never sees which tier produced an image, so tier bias can't creep in. That is what lets us put an actual number on the quality ladder instead of asserting one, and it is why the figures below differ from the "higher is always better" story you will read elsewhere.

The New Tier Ladder

GPT Image 2 exposes three generation quality tiers through providers such as Runware: low, medium, and high.[2] The useful question is not whether high wins. It does. The useful question is how much quality each extra dollar buys.

Scores are applied weighted scores from the current VibeDex Sonnet 4.6 blind benchmark.

GPT Image 2 Pricing by Quality Tier

At 1024×1024, GPT Image 2 costs roughly $0.014 per image on low, $0.055 on medium, and $0.212 on high — a 15x spread from the cheapest tier to the most expensive. Runware's pricing page quotes a flat $0.006, but the real charge depends on the quality tier and the prompt-token length, so high-tier renders cost meaningfully more than that headline figure.[2]

The takeaway: all three tiers are strong, high is meaningfully best, and low is no longer a weak outlier. But the practical gap between high and medium is small enough that medium should be the default unless the image is final, visible, and worth optimizing for that last 0.047 points.

GPT Image 2 Low Quality: When the Cheapest Tier Is Enough

At low quality, GPT Image 2 scores 3.946 on our blind Sonnet 4.6 benchmark — the lowest of the three tiers, but no longer a weak outlier. Each 1024×1024 render costs about $0.014 and spends ~272 image tokens. Low wins outright on 11 of 50 prompts against high, almost always simple scenes where all three tiers land within judging noise. Use it for drafts, creative exploration, and high-volume batches you expect to throw away. Avoid it for small text, close-up faces, or final client work, where the missing detail shows.

GPT Image 2 Medium Quality: The Production Default

Medium is the tier most teams should standardize on. It scores 4.108 — just 0.047 behind high — while costing $0.055 per 1024×1024 image and using ~1,056 image tokens. It beats low on 35 of 50 prompts (+0.162) and loses to high on only 22 of 50. That is roughly 99% of high's quality for about a quarter of the price. Reach for medium on marketing assets, product visuals, social graphics, UI mockups, and any embedded text at normal display sizes.

GPT Image 2 High Quality: For Hero Renders

High is the top-scoring tier at 4.155, but it earns its price only on the hardest prompts. A 1024×1024 render costs about $0.212 and burns ~4,160 image tokens — roughly 15x low and 4x medium. High beats medium on 22 of 50 prompts by an average of just 0.047 points, so the lift is real but small. Spend it where a single final render matters: product hero shots, dense or small text, close-up portraits, difficult material physics, and large-format or print output. For batch work, the extra tokens rarely pay off.

Head-to-Head Results

On the 50-prompt Sonnet set, high beats medium by 0.047 points on average and wins 22 of 50 prompts (medium takes 19, nine tie). Medium beats low by 0.162 points and wins 35 of 50. High beats low by 0.209 points and wins 32 of 50. That is a real ladder, not random noise.

Featured Examples

Two more high-production-value fashion-photography prompts, included for visual range alongside the scored comparison at the top of the page. Both were generated but not formally judged. They are illustrative — shown for output range, separate from the scored head-to-head.

prompt-0178

Generated but not formally judged

“High-fashion cinematic photograph of a model standing in an immense field of lavender in Provence at the moment the sun dips below the horizon, the...”

Show full prompt

Low ($0.014)

Medium ($0.055)

High ($0.212)

prompt-0112

Generated but not formally judged

“Editorial fashion photograph of a model in a flowing crimson silk gown standing at the edge of an infinity pool overlooking Santorini at golden hour,...”

Show full prompt

Low ($0.014)

Medium ($0.055)

High ($0.212)

See the Tiers Side-by-Side

We generated all three tiers for every prompt in our 50-prompt benchmark and judged them blind in three independent Sonnet 4.6 passes. Below are 20 representative prompts, grouped by which tier won — the winning render is ringed in accent. Hover any image for the judge's rationale, or click for the lightbox.

High tier wins (10)

Prompts where the premium tier earned its price — fine commercial detail, complex compositions, multi-element scene logic, and material physics where the extra spend shows up in the score.

prompt-0053 · spread 0.24 · High tier wins

“Extreme macro close-up of a human eye filling the entire frame, crystalline iris patterns in blue-green, individual eyelashes in sharp focus in...”

Show full prompt

?

Low ($0.014)

4.37

Imaging Quality 4.5/5: Exceptional sharpness throughout -- iris crystalline patterns render with fine radial detail, individual eyelashes are crisp in the foreground plane, and a clear window-and-trees reflection is visible in the pupil. Noise is negligible. | Aesthetics 4.5/5: The blue-green iris color is vivid and beautiful, skin tones around the eye are natural and warm, and tonal gradation from the sclera to the dark pupil is well-executed. | Composition 4.5/5: Eye fills the frame as instructed, eyelashes in the foreground establish depth, and the centered composition places the pupil reflection at a visually engaging position.

?

Medium ($0.055)

4.48

Imaging Quality 5/5: Excellent macro fidelity: iris fibers and crystalline radial patterns are rendered with high microstructural detail; individual eyelashes are tack-sharp with believable shallow depth; skin pores around the eye visible. No noise or AI smearing. | Aesthetics 4/5: Striking blue-teal/amber iris palette reads cinematic; tear film highlights add polish. Slightly clinical look and the very symmetric iris pattern feels a touch stylized rather than naturalistic. | Composition 4/5: Eye fills the frame as instructed with pupil near center; balanced. Slight crop on lash tips and a fairly flat horizontal orientation keep it from being exceptional.

?

High ($0.212)

4.61

Imaging Quality 5/5: Exceptional clinical sharpness throughout — iris fiber patterns, individual eyelashes, and skin pores rendered with zero visible noise or softness. Technically flawless macro execution. | Aesthetics 4.5/5: Beautiful blue-green crystalline iris with rich tonal gradients; the composition and color palette are striking and visually compelling. | Composition 4.5/5: Frame-filling iris placement is precise; eyelashes sharp in foreground create depth, and the window pupil reflection adds a focal anchor.

prompt-0058 · spread 0.39 · High tier wins

“A woman wearing a bright blue blazer and white pants standing in front of a bright yellow door, holding a red handbag in her left hand and a green...”

Show full prompt

?

Low ($0.014)

4.02

Semantic Accuracy 5/5: All five specified color-item pairings are correct: bright blue blazer, white pants, yellow door, red handbag, and green umbrella; from her perspective the red handbag is in her left hand and green umbrella in her right hand, matching the prompt. | Spatial Framing 4.5/5: Full figure is well-framed with yellow door prominent in background and city street visible; both hands and held items clearly visible.

?

Medium ($0.055)

3.97

Semantic Accuracy 3/5: Blue blazer, white pants, yellow door, red handbag, and green umbrella are all present with correct colors, but the hand assignment is reversed: the red handbag is in her right hand and the green umbrella in her left, contra the prompt's specified left/right assignment. | Spatial Framing 5/5: Full-body framing is clean with the yellow door prominently behind her and a city street visible, matching the specified background context.

?

High ($0.212)

4.36

Semantic Accuracy 5/5: All specified elements present: bright blue blazer, white pants, yellow door, red handbag in left hand, green umbrella in right hand, city street background — all confirmed correct. | Spatial Framing 4.5/5: Woman is well-centered before the yellow door, city street recedes naturally behind her, and accessories are clearly visible within the frame.

prompt-0073 · spread 1.12 · High tier wins

“Ultra high resolution product photograph of a luxury watch, every microscopic detail of the dial visible, sapphire crystal catching light, brushed...”

Show full prompt

?

Low ($0.014)

3.36

Aesthetics 3.5/5: Clean three-quarter presentation with a professional neutral-warm gradient background; the subdued elegance suits a luxury timepiece, though the image lacks the polish of a true editorial shot. | Imaging Quality 3/5: Dial text is legible and the sapphire crystal has a catch-light, but sharpness is clearly limited by the low-output quality setting -- Zaratsu polishing is suggested rather than crisp, leather strap grain is soft, and the image falls well short of zero-noise magazine standard. | Composition 3/5: Standard product photography angle with the watch slightly small in frame; margins are tight and the composition is functional but unremarkable.

?

Medium ($0.055)

3.81

Aesthetics 4/5: Classic Grand-Seiko-style dial reads as a real luxury watch: deep blue sunburst dial, hour markers gleam, crown and case have polished/brushed contrast. Lighting is product-photo clean. | Imaging Quality 3/5: Output resolution is modest; while the dial and hands are crisp at this size, the 'every microscopic detail' / '8K' brief is not really delivered — leather grain and brushed steel anisotropy are suggested rather than micro-resolved. | Composition 4/5: Three-quarter watch angle with crown to the right and strap trailing off is textbook magazine product framing; clean negative space behind the case.

?

High ($0.212)

4.48

Aesthetics 4.5/5: Elegant, clean product shot with dark neutral background; lighting highlights the watch dial and steel case effectively with a refined, magazine-worthy feel. | Imaging Quality 4.5/5: Low noise, sharp throughout from dial text to leather strap grain; sapphire crystal light catchment is clearly rendered; minor lighting falloff on the left case edge. | Composition 4.0/5: Effective three-quarter angle on a dark surface; slightly left-weighted but creates a natural product presentation with shadow definition.

prompt-0094 · spread 0.89 · High tier wins

“Photorealistic 3D render of a modern cantilevered house extending over a cliff edge, the building's structural concrete beams clearly visible in the...”

Show full prompt

?

Low ($0.014)

3.40

Static Physics 2.5/5: The structure uses visible diagonal timber/concrete posts supporting the extended section from below rather than a true cantilever. The prompt describes load-distributing beams anchored to the hillside, but the rendered solution shows conventional vertical prop supports, undermining the core structural physics. | Material Physics 4/5: Concrete, glass, and steel railing materials render with convincing physical properties including surface roughness on concrete, specular glints on steel cable railings, and plausible glass reflections under golden-hour lighting.

?

Medium ($0.055)

3.65

Static Physics 2.5/5: House does cantilever over the cliff with a visible diagonal strut underneath, but the structural concrete beams aren't clearly articulated to show how load distributes back to the anchored portion. The cantilever is relatively short and doesn't read as a daring engineering feat. No rooftop infinity pool with overspill is visible — the roof appears to be a planted terrace with glass railing. Anchoring into hillside is implied but not clearly resolved. | Material Physics 3.5/5: Concrete reads as concrete with reasonable texture, glass walls show plausible reflections of sky/sun, vegetation on green roof looks natural. Steel cable railings are present though tension isn't clearly conveyed. Late afternoon golden hour lighting and long shadows on the cliff face are plausible. Materials feel physically grounded if not Octane-perfect.

?

High ($0.212)

4.29

Static Physics 4.0/5: Cantilevered concrete beams extending over the cliff are clearly visible and structurally plausible, with appropriate mass and a credible anchor into the hillside. Infinity pool on the rooftop and steel cable railings under visible tension are correctly rendered. | Material Physics 4.0/5: Concrete, glass, and steel cable materials all render with physically convincing properties — glass panels reflect correctly, concrete surfaces show appropriate texture, and the green roof vegetation is natural. Octane-style global illumination is convincing.

prompt-0097 · spread 0.52 · High tier wins

“Balanced rock cairn of seven smooth river stones stacked from largest to smallest on a beach at low tide, each stone's center of gravity precisely...”

Show full prompt

?

Low ($0.014)

4.23

Static Physics 4/5: Seven stones are stacked convincingly from largest at base to smallest at top with centers of gravity appearing well-aligned; the stack reads as physically plausible. The base stone shows subtle sand compression beneath it as specified. | Material Physics 4/5: Stones have realistic smooth river-polished texture with natural grey-brown color variation and a slight wet sheen; the wet sand around the base transitions naturally from saturated dark to drier surrounding sand.

?

Medium ($0.055)

3.77

Static Physics 4.0/5: Stack contains seven stones tapering from a large base to a small top, with each stone reasonably centered over the one below. Contact points read as plausible and the overall cairn looks balanceable. The top stone is small though perhaps slightly larger than a thumb. No stone appears to defy gravity or float. | Material Physics 4.0/5: Wet sand around the base shows a slight depression/ripple consistent with the stack's weight and recent water. Stones have a damp, smooth river-stone appearance. Lighting on the wet sand and surface reflectivity are convincing. Soft golden-hour sand wetness reads correctly.

?

High ($0.212)

4.29

Static Physics 4.5/5: Seven stones are stacked largest-to-smallest with centers of gravity convincingly aligned through the column, producing a stable-looking structure. The top stone is appropriately tiny and the wet sand beneath the base stone shows subtle compression. | Material Physics 4.0/5: Smooth river stone surfaces have a correct matte finish with a subtle wet sheen on some stones, and the beach sand texture reads as authentic damp low-tide sand.

prompt-0102 · spread 0.68 · High tier wins

“High-end commercial photography of a luxury perfume bottle on a black marble surface, the heavy cut crystal bottle refracting an intricate rainbow...”

Show full prompt

?

Low ($0.014)

3.77

Static Physics 3/5: The bottle sits stably and surface reflections are coherent, but the requested perfume drop forming a meniscus beside the bottle is missing, leaving one key static-physics detail unrendered. | Material Physics 3.5/5: The rainbow caustic pattern cast onto the surface below is vivid and convincing, demonstrating accurate light refraction through cut crystal facets. However, the amber liquid inside shows no visible depth-based gradient and the single perfume drop with meniscus specified in the prompt is absent.

?

Medium ($0.055)

4.00

Material Physics 3.5/5: Cut crystal bottle shows good faceted refraction. A rainbow/caustic pattern is visible on the marble below the bottle, which is the standout element. Liquid shows a subtle vertical gradient from lighter top to deeper amber at the bottom. However, the atomizer's mirror-polished metal collar is absent — the cap reads as more crystal glass rather than metal with a clean rectangular softbox highlight. No droplet of perfume with a meniscus is visible beside the bottle. | Static Physics 3.0/5: Bottle sits stably on the marble with appropriate weight presence. No droplet/meniscus is depicted, so the surface-tension aspect isn't demonstrated. Reflection footprint on marble is plausible.

?

High ($0.212)

4.45

Material Physics 4.5/5: Rainbow caustic pattern cast on the black marble is vivid and physically plausible for heavy cut crystal; the liquid inside shows a credible pale-gold-to-amber depth gradient; the gold collar has a clean rectangular softbox highlight indicating mirror polish; and the perfume drop beside the bottle displays a clear meniscus with visible surface tension. | Static Physics 4.0/5: Bottle sits stably on the flat marble surface, the drop forms a correctly hemispherical bead, and all structural components of the atomizer appear sound and correctly proportioned.

prompt-0103 · spread 0.39 · High tier wins

“Freshly poured craft beer in a tulip glass showing distinct layers of carbonation, tiny bubbles nucleating at the base and growing larger as they rise...”

Show full prompt

?

Low ($0.014)

4.07

Static Physics 4/5: The tulip glass sits stably on the bar surface with no structural anomalies; the liquid level and foam distribution obey gravity correctly. | Material Physics 4.5/5: Carbonation bubble streams rising from the base are clearly rendered with convincing nucleation and growth, the thick creamy head shows realistic micro-foam layering, condensation droplets on the chilled glass exterior are present, and the amber liquid glows with warm translucency from pub backlighting exactly as described.

?

Medium ($0.055)

4.28

Material Physics 3.5/5: Condensation droplets on the outside of the chilled glass are well-rendered and convincing — clearly cold-glass material physics is captured. The head shows varying bubble sizes with a creamy upper region. Amber liquid has good warm translucency catching the pub light. However, the interior carbonation looks more like an even bubble texture throughout than organized rising streams nucleating at specific points on the base and growing as they ascend. | Static Physics 3.5/5: Tulip glass stands stably with appropriate weight on the bar surface. Foam head sits proudly above the rim with a domed shape consistent with surface tension and recent pour. Liquid level reads correctly.

?

High ($0.212)

4.46

Material Physics 4.5/5: Organized bubble streams nucleating at the base and growing larger toward the surface are clearly rendered; the creamy head shows convincing micro-foam at the top transitioning to larger bubbles below; condensation droplets on the exterior glass are highly realistic; and the amber liquid glows with convincing translucency under warm pub light. | Static Physics 4.0/5: Glass sits stably on the pub table, the head is maintained at a realistic height and density, and no structural inconsistencies are present.

prompt-0119 · spread 0.22 · High tier wins

“Concept art turnaround sheet of a post-apocalyptic wanderer showing front, side, and three-quarter views, consistent proportions across all views with...”

Show full prompt

?

Low ($0.014)

3.80

Subject Human 3.5/5: Consistent proportions across front, side, and three-quarter views are achieved; prosthetic left leg below the knee is clearly visible with mechanical joint detail, but asymmetric right-cheek scarring and bandaged fingertips are not clearly discernible at this rendering quality. | Scene Logic 3.5/5: Turnaround sheet format is correctly structured with labeled views, material palette, and accessory callouts; the layout reads as a professional concept art sheet. | Object Integrity 4.0/5: Layered weathered clothing shows correct overlap and material weight; the prosthetic leg is convincingly detailed with mechanical joint articulation visible.

?

Medium ($0.055)

3.87

Scene Logic 4/5: Turnaround sheet well-structured with front, side, and three-quarter views clearly laid out alongside annotations and silhouette guides; standard concept art presentation correctly followed. | Subject Human 4/5: Proportions consistent across all three views; asymmetric facial scarring visible on the right cheek, prosthetic left leg with mechanical knee joint clearly rendered, layered clothing shows believable overlap and weight. | Object Integrity 4/5: Mechanical prosthetic leg detailed with visible joint articulation; layered post-apocalyptic garments including long coat, strapped boots, and wrapped hands are structurally coherent.

?

High ($0.212)

4.02

Scene Logic 3.5/5: Turnaround sheet is well-organized with front, side, and three-quarter views clearly labeled, plus detail callout insets for face, hands, and mechanical leg — the format reads as a professional character reference sheet. | Subject Human 4.0/5: Character proportions are consistent across all three views, the weathered face with asymmetric scarring is visible in detail insets, and the overall figure anatomy is plausible. | Object Integrity 4.0/5: Prosthetic left leg with visible mechanical joint detail is present, bandaged hands with exposed fingertips are shown in callouts, and layered clothing overlap is convincingly rendered.

prompt-0133 · spread 0.79 · High tier wins

“Busy Saturday morning brunch restaurant scene, a waiter carrying a tray of three plated dishes navigating between occupied tables, the tray balanced...”

Show full prompt

?

Low ($0.014)

3.45

Biomechanics 3/5: The waiter carries the tray single-handed with reasonably natural wrist posture and visible finger spread beneath, but the tray is at chest/elbow height rather than the shoulder height specified; the posture is functional but does not match the specific biomechanical description. | Static Physics 3.5/5: Three plated dishes are arranged on the tray in a balanced configuration and the tray appears level; the load distribution looks plausible for the carrying position shown. | Material Physics 3.5/5: The three breakfast dishes (eggs benedict with hollandaise, a pancake stack with fruit, and avocado toast) are rendered with recognizable textures and appear as distinct, realistic food items.

?

Medium ($0.055)

3.49

Biomechanics 3.0/5: Waiter holds the tray with a single hand from below, but the tray is at chest/lower-rib height rather than the prompt's specified shoulder height. The supporting hand and wrist angle are partially visible and look plausible, though the spread-fingers-beneath posture isn't clearly shown. Overall body posture reads as natural standing-and-carrying. | Static Physics 3.5/5: Tray balance with three plates is plausible — the tray is held flat, plates are centered/spread. No visible tipping or sliding. Plates rest on the tray surface coherently. | Material Physics 3.5/5: Food materials (eggs, pancakes, avocado toast) read with reasonable texture. Tablecloths, wood furniture, and ceramics all look physically grounded.

?

High ($0.212)

4.24

Biomechanics 3.5/5: The waiter's posture navigating between tables is natural, but the tray is held in front at roughly chest/waist height using what appears to be both hands rather than the specified one-hand at shoulder height with correct wrist angle and spread fingers — the key biomechanical requirement is only partially met. | Static Physics 4.0/5: Three distinctly plated dishes — eggs benedict, pancake stack, and avocado toast — are all visible and appear stably balanced on the tray with plausible weight distribution. | Material Physics 3.5/5: Food textures are generally convincing — hollandaise on eggs benedict, fluffy pancake stack, avocado toast — but surface detail and material specificity are somewhat generic.

prompt-0147 · spread 0.47 · High tier wins

“Anime scene of a ramen shop at night, the noren curtain reading らーめん一番 in white characters on navy fabric, a glowing sign above reading ICHIBAN RAMEN...”

Show full prompt

?

Low ($0.014)

3.78

Spatial Framing 4/5: Exterior storefront view is well composed with the sign, noren, and lit interior visible in a single coherent frame. | Text Rendering 4/5: The glowing sign above correctly renders both Japanese and English ICHIBAN RAMEN. Noren curtain also shows the correct Japanese characters in white on navy fabric as specified. | Semantic Accuracy 3.5/5: Ramen shop at night with noren, glowing sign, and warm light spilling out are all present, but the visual style leans photorealistic rather than anime as the prompt specifies, only partially achieving the requested aesthetic.

?

Medium ($0.055)

4.12

Text Rendering 2/5: Japanese characters approximate らーめん一番 but with visible glyph distortion — strokes are off and ichiban (一番) reads more like 一桑 in places. ICHIBAN RAMEN romaji is largely legible. Multiple banners repeat the same string instead of placing one on the noren and one on the sign above. | Spatial Framing 3/5: Storefront framing is workable with curtains at entrance and signage above the door, but the noren curtain is not clearly distinguished from the signboards — both render as solid lit signs rather than a fabric noren over the doorway. | Semantic Accuracy 3/5: Night ramen shop with warm light spilling out is captured. However the noren is specified as white characters on navy fabric — image shows beige/cream backgrounds with dark text. Two distinct text elements (noren + glowing sign) are not cleanly separated.

?

High ($0.212)

4.25

Text Rendering 4.0/5: Top sign shows '一番らーめん' and 'ICHIBAN RAMEN' in correct bilingual pairing; noren curtain displays Japanese characters. Text is readable and stylistically integrated into the anime aesthetic, though character ordering on the noren does not precisely match the specified 'らーめん一番' sequence. | Semantic Accuracy 4.0/5: Night setting, warm interior light spilling out, noren curtain, glowing bilingual sign, and anime visual style are all present and well-executed. | Spatial Framing 4.0/5: Storefront framed cleanly from street level; sign above and noren below are placed as specified; warm interior visible through windows.

Medium tier wins (6)

Prompts where medium hit the sweet spot. Often mid-complexity scenes and product photography where high adds polish but not score, and low loses on small details.

prompt-0006 · spread 0.32 · Medium tier wins

“Chef's hands chopping vegetables with a kitchen knife, proper grip technique, cutting board”

?

Low ($0.014)

3.91

Biomechanics 3.0/5: The guiding hand is not in a proper professional claw grip — fingers are not fully curled back as required for safe technique. Knife-hand grip is acceptable but lacks the precision of a pinch grip. | Material Physics 4.0/5: Vegetable textures, blade reflections, and cutting board grain are all rendered authentically with appropriate material properties. | Static Physics 3.5/5: Knife blade contact with the cutting board is geometrically plausible and the vegetables are positioned stably on the board.

?

Medium ($0.055)

4.23

Biomechanics 4.5/5: Guiding hand shows correct claw-grip with curled fingertips, knife hand has solid pinch grip on blade — highly convincing professional chopping technique. | Material Physics 4/5: Freshly cut vegetable surfaces (peppers, spring onion) show correct texture and reflectance; cutting board wood grain is realistic. | Static Physics 4/5: Knife angle and contact point with the board are physically plausible; board and food items are stably placed.

?

High ($0.212)

4.11

Biomechanics 4.5/5: The knife hand shows a professional pinch grip on the blade spine and the guiding hand uses a proper claw grip protecting the fingertips — both are correct culinary technique. Posture and arm position are appropriate for active chopping. | Material Physics 4.5/5: Cut vegetables display realistic textures and freshly sliced surfaces; red pepper strips, green onion rounds, and red onion are all rendered with plausible material properties. | Static Physics 4.0/5: The cutting board is flat and stable on the counter, knife is in a natural mid-stroke position, and all objects are properly grounded with no spatial anomalies.

prompt-0008 · spread 0.46 · Medium tier wins

“Person typing on laptop keyboard, all ten fingers positioned naturally over keys, office desk”

?

Low ($0.014)

3.54

Subject Human 3.0/5: Both hands are positioned naturally on the keyboard but only hands are shown with no face; confirming all 10 fingers with correct count is difficult as several fingers are partially obscured by angle and overlap. | Object Integrity 4.0/5: Laptop body, keyboard, trackpad, and desk objects (mug, notebook, plant) are all rendered realistically with correct proportions and no structural anomalies. | Scene Logic 4.0/5: Office desk scene is coherent and well-arranged — laptop, motivational mug, notebook, pen, and plant are naturally distributed on a wooden desk.

?

Medium ($0.055)

3.97

Subject Human 4/5: Both hands visible with natural finger positioning; ten fingers plausibly represented with correct wrist angles and joint articulation, no obvious deformities at this view angle. | Object Integrity 4/5: Laptop keyboard accurately rendered with clearly defined keys and correct form factor; desk accessories (mug, pen holder, notepad, charts) are structurally coherent and recognisable. | Scene Logic 4/5: Office desk environment contextually consistent with uniform lighting, naturalistic object placement, and a convincing work-environment composition.

?

High ($0.212)

3.51

Subject Human 3.5/5: Both hands present and anatomically plausible with natural proportions; fingers are distributed across keys in a credible typing pose, but the 'all ten fingers' requirement is only partially delivered — some fingers are partially obscured and the hand position is casual rather than a textbook home-row grip. | Object Integrity 4.0/5: Laptop body and keyboard look convincing and well-rendered with identifiable key layout and screen hinge; no visible deformations or implausible elements. | Scene Logic 4.0/5: Clean office desk with books, mug, plant, notebook, and pen arranged in a spatially coherent and realistic manner.

prompt-0016 · spread 0.39 · Medium tier wins

“Water being poured from a glass pitcher into a tumbler, liquid mid-splash, light refracting through”

?

Low ($0.014)

4.30

Material Physics 4.0/5: Water stream, splash crown, and light refraction through the glass are convincingly rendered with caustic-like patterns on the surface; fluid dynamics look physically reasonable. | Static Physics 4.0/5: Pitcher and tumbler are stably positioned with a plausible pouring angle; the stream trajectory follows a believable arc from spout to glass.

?

Medium ($0.055)

4.69

Material Physics 4.5/5: Water stream, mid-splash crown with bubble formation, and light refraction through glass and water are all convincingly rendered with realistic fluid dynamics. | Static Physics 4.5/5: Pitcher tilt and stream trajectory follow natural gravity; tumbler placement and fill level are physically consistent.

?

High ($0.212)

4.38

Material Physics 4.5/5: The water stream shows appropriate transparency, surface tension, and light refraction through both the glass pitcher and tumbler is convincingly rendered with caustic patterns on the marble surface. The mid-pour splash in the glass is believable. | Static Physics 4.5/5: The pour angle, gravity-driven water column thickness, and glass stability on the surface are all physically plausible; the stream tapers naturally from pitcher mouth to glass.

prompt-0056 · spread 0.22 · Medium tier wins

“Exactly three red apples on white surface”

?

Low ($0.014)

4.62

Semantic Accuracy 5/5: Exactly three apples are present, all clearly red, on a clean white surface — precise count and color match the prompt exactly. | Spatial Framing 4.5/5: Apples are cleanly arranged in a slightly staggered trio formation with good separation; composition is simple and effective with all three fully in frame.

?

Medium ($0.055)

4.84

Semantic Accuracy 5/5: Exactly three apples are present, all visibly red, on a clean white surface — every constraint in the prompt is satisfied. | Spatial Framing 5/5: Three apples are evenly spaced and centered in the frame with a clean white background and balanced negative space.

?

High ($0.212)

4.72

Semantic Accuracy 5/5: Exactly three red apples present on a clean white surface — count and color specification met precisely. | Spatial Framing 4.5/5: Apples evenly spaced in a clean horizontal row with generous white space, well-balanced composition.

prompt-0142 · spread 0.42 · Medium tier wins

“Product photography of a premium coffee bag standing upright on a marble countertop, the bag made of matte black kraft paper with a clear window...”

Show full prompt

?

Low ($0.014)

4.48

Spatial Framing 4.5/5: Clean product photography framing with the bag centered and filling the frame well; soft shadow beneath grounds the object naturally on the marble surface. | Text Rendering 4.5/5: SUMMIT ROASTERS in gold embossed serif at top, ETHIOPIAN YIRGACHEFFE below, SINGLE ORIGIN LIGHT ROAST 340G in small caps, and ORGANIC CERTIFIED circular stamp with leaf icon all render accurately and legibly. Side text ROASTED IN PORTLAND OR is partially visible; typography hierarchy progresses correctly in size. | Semantic Accuracy 4.5/5: Matte black kraft paper bag upright on marble countertop with clear window showing whole beans, gold foil embossed label, and barcode on side all present as specified. White background with soft gradient shadow beneath the bag is correctly rendered.

?

Medium ($0.055)

4.82

Text Rendering 3/5: SUMMIT ROASTERS and ETHIOPIAN YIRGACHEFFE are legible at top tier. Third line (SINGLE ORIGIN / LIGHT ROAST / 340G) is partially garbled. ORGANIC CERTIFIED stamp is rendered but ring text is distorted. No visible barcode or ROASTED IN PORTLAND OR side text. Hierarchy approximately correct but several required text elements missing or unreadable. | Spatial Framing 3/5: Bag stands upright and centered with macro-like fine detail; soft gradient shadow beneath is present. However the surface is not clearly marble — it reads as a dark wood/textured surface, missing a key spatial/setting requirement. | Semantic Accuracy 2/5: Matte kraft paper texture passable, beans visible through a window OK, but no clear window with whole beans (the window is large and dominant), no side barcode, no Portland origin text, no marble countertop. Several specified label elements absent.

?

High ($0.212)

4.40

Text Rendering 4.5/5: All text elements rendered correctly — SUMMIT ROASTERS in gold serif, ETHIOPIAN YIRGACHEFFE in white sans-serif, SINGLE ORIGIN bullet LIGHT ROAST bullet 340G in small caps, circular certification stamp with leaf icon. Typography hierarchy is clearly legible with each line at progressively smaller sizes. | Spatial Framing 4.5/5: Bag is well-centered upright on marble countertop, clean white background with soft gradient shadow beneath exactly as specified; 100mm macro-style clarity throughout. | Semantic Accuracy 4.5/5: Matte black kraft paper bag with clear window showing whole beans, barcode and side text visible, gold foil embossed effect rendered convincingly — all major descriptors accounted for.

prompt-0184 · spread 0.37 · Medium tier wins

“Magazine-quality food photograph of a freshly sliced medium-rare wagyu ribeye steak on a heated cast iron plate, the cross-section showing...”

Show full prompt

?

Low ($0.014)

4.38

Aesthetics 4.5/5: Excellent food photography aesthetic with warm directional light, appetising marbling cross-section, melting herb butter, and thyme garnish creating a rich, magazine-ready presentation. | Imaging Quality 4/5: Marbling web and pink-to-brown gradient are convincingly rendered; grill marks, pooling juices, and compound butter melt transition are all present with solid detail, though ultra-fine texture at the linen level is not achieved. | Composition 4/5: Steak fills the frame with the cast iron partially visible; butter pat and thyme sprig create natural visual interest and the overall framing achieves a magazine-quality feel.

?

Medium ($0.055)

4.61

Aesthetics 5/5: Magazine-quality food photography with rich dark tones, an intensely appetizing cross-section revealing vivid wagyu marbling, melting herb compound butter, and a well-placed thyme sprig — highly polished and appealing. | Imaging Quality 5/5: Extraordinary marbling detail in the cross-section, individual grill marks with caramelized micro-blisters, pooling juices with visible viscosity, realistic solid-to-liquid butter gradient, cast iron patina detail, and individual thyme leaf pairs all rendered with exceptional technical precision. | Composition 4/5: Angled sliced sections create strong diagonal movement, the butter and thyme draw focus effectively, and the cast iron plate frames the subject well; slightly tight cropping reduces sense of the cutting board as a stage.

?

High ($0.212)

4.24

Aesthetics 4.5/5: Warm directional lighting, rich deep colors, and genuinely appetizing rendering achieve magazine-quality food photography. The melting herb butter and thyme garnish are beautifully styled. | Imaging Quality 4.5/5: Exceptional sharpness — marbling is visible in the cross-section, Maillard crust with caramelization detail is convincing, individual thyme leaf pairs are distinct, and the cast iron seasoned surface texture is evident. Juices show realistic viscosity. | Composition 4.0/5: Well-arranged with the sliced steak revealing the cross-section, melting butter centered on top, and thyme garnish framing the subject naturally on the cast iron plate.

Low tier wins (1)

Prompts where the cheapest tier scored highest. Usually a sign that all three tiers handled the prompt comfortably and judging noise dominated the small gap.

prompt-0063 · spread 0.76 · Low tier wins

“Cyberpunk Tokyo alley at night, neon signs reflecting on wet pavement, cinematic color grading”

?

Low ($0.014)

4.25

Aesthetics 4.5/5: Strong cyberpunk color palette of blue, pink, and purple neon against deep blacks, with vivid wet-pavement reflections creating layered depth; cinematic mood is convincingly atmospheric. | Imaging Quality 4/5: Neon signs and foreground pavement are sharp; background depth softens naturally and shadow areas show minor compression artifacts, but overall rendering quality is solid. | Composition 4/5: Central leading line with lone figure under umbrella, neon signage framing both sides of the alley -- a classically strong composition that draws the eye.

?

Medium ($0.055)

3.49

Aesthetics 4/5: Mood nails the brief: red/magenta neon dominance, wet asphalt reflections, lone hooded figure, cinematic color grade. Strong vibe even at low res. | Imaging Quality 2/5: Output is rendered at very low effective resolution; signage glyphs and architectural detail are mushy/illegible, edges are soft, neons lack crisp bloom structure. Significantly undercuts a 'cinematic' brief. | Composition 3/5: Centered one-point-perspective alley with figure silhouetted works as a familiar cyberpunk frame, but it is generic and the symmetrical signage on both sides feels uncomposed.

?

High ($0.212)

4.17

Aesthetics 4.5/5: Strong moody cyberpunk atmosphere with vivid neon saturation, excellent color temperature contrast between signs and dark ambient; genuinely cinematic color grading. | Imaging Quality 4.0/5: Some compression banding in shadow areas and mild softness in the mid/far background, though foreground pavement reflections are sharp and convincing. | Composition 4.0/5: Classic symmetrical alley composition with lone figure at vanishing point; effective but conventional; neon signs frame the scene well vertically.

Effectively tied (3)

Prompts where two or three tiers landed within 0.05 points of each other. The tier choice barely moved the score — pick on cost.

prompt-0004 · spread 0.37 · Effectively tied

“A professional classical guitarist's left hand pressing an F barre chord on the fretboard, all fingers arched naturally with proper technique, visible...”

Show full prompt

?

Low ($0.014)

3.33

Biomechanics 3.0/5: The hand is pressing what appears to be a partial barre or open chord grip rather than a clear F barre — the index finger doesn't visibly bar across all strings with proper technique. Remaining fingers are reasonably arched but the defining barre form is ambiguous. | Static Physics 3.5/5: Guitar neck, fretboard, and hand contact points are spatially plausible; strings sit correctly on the frets and the instrument geometry is consistent. | Material Physics 4.0/5: Wood grain, string texture, and skin detail including visible fingertip calluses are convincingly rendered with appropriate surface properties.

?

Medium ($0.055)

3.70

Biomechanics 3.5/5: Index finger barre visible across fret, other fingers arched naturally with plausible pressure, but full F barre chord configuration is ambiguous — ring and pinky placement not definitively correct. | Static Physics 4/5: Guitar neck geometry, fret spacing, and string tension all appear physically coherent and accurately rendered. | Material Physics 4/5: Skin texture with faint calluses, nylon string surface, and warm wood grain all rendered with appropriate material properties.

?

High ($0.212)

3.68

Biomechanics 2.5/5: The hand is not forming a proper F barre chord — the index finger is not laid flat across all strings in the characteristic barre position, making this a different chord shape. Calluses are not distinctly visible and the defining technique element of the prompt is absent. | Static Physics 4.0/5: Fretboard geometry, string spacing, and fret layout are all physically accurate. The hand's contact with the neck is spatially plausible. | Material Physics 4.0/5: String material, wood grain of the fretboard, and the metallic sheen of the frets all render realistically with appropriate surface properties.

prompt-0043 · spread 0.23 · Effectively tied

“Birthday cake with HAPPY BIRTHDAY written in blue frosting, lit candles on top”

?

Low ($0.014)

4.80

Text Rendering 5/5: HAPPY BIRTHDAY is rendered in clean, legible blue frosting lettering across two lines on the cake top with no misspellings or distortion. | Spatial Framing 5/5: Candles are correctly positioned on top of the cake and the cake is well-centered in frame with appropriate context. | Semantic Accuracy 5/5: All prompt elements are present: birthday cake, blue frosting text, and multiple lit candles on top — fully faithful to the brief.

?

Medium ($0.055)

4.75

Text Rendering 5/5: HAPPY BIRTHDAY rendered in clean, clearly legible blue 3D frosting letters split across two lines; spelling is perfect and color matches the prompt exactly. | Spatial Framing 5/5: Cake is well-centered with appropriate headroom for the lit candles, and the composition cleanly presents the subject. | Semantic Accuracy 5/5: All prompt elements present: birthday cake, blue frosting text, and lit candles on top.

?

High ($0.212)

4.57

Text Rendering 5/5: HAPPY BIRTHDAY rendered in large, perfectly legible blue frosted letters with correct capitalization and clean letterforms. No misspellings or distortions. | Spatial Framing 4.5/5: Cake is well-centered in frame, all lit candles clearly visible on top, and gift boxes appear naturally in background without cluttering. | Semantic Accuracy 4.5/5: Cake present with blue frosting text and multiple lit candles as specified; sprinkle border and gift prop add appropriate context without contradicting the prompt.

prompt-0045 · spread 0.19 · Effectively tied

“Street sign at intersection showing BROADWAY and 42ND ST, New York City buildings in background”

?

Low ($0.014)

4.83

Text Rendering 5/5: Both BROADWAY and 42ND ST are rendered in perfectly legible white text on authentic NYC green street sign format with no errors or glyph distortion. | Spatial Framing 5/5: Signs are stacked at an intersection pole in the foreground with NYC skyscrapers including a Chrysler-like spire filling the background correctly. | Semantic Accuracy 5/5: NYC street corner setting with correct sign design and appropriate urban canyon backdrop; Statue of Liberty motif on signs is an accurate embellishment.

?

Medium ($0.055)

4.79

Text Rendering 5/5: BROADWAY and 42ND ST are fully legible on authentic NYC-style green street signs with correct typography and no spelling errors. | Spatial Framing 5/5: Signs occupy the foreground clearly while the Chrysler Building and NYC skyline are visible in depth, creating strong compositional context. | Semantic Accuracy 5/5: Correct intersection signage, recognizable Manhattan skyline, and urban street furniture all match the prompt precisely.

?

High ($0.212)

4.64

Text Rendering 5/5: Both signs render BROADWAY and 42ND ST in perfectly legible NYC-style green-and-white format with correct typography and no character errors. | Spatial Framing 5/5: Signs are prominently positioned at the intersection pole, NYC skyline including the Chrysler Building fills the background convincingly. | Semantic Accuracy 5/5: Correct intersection signage, authentic NYC architectural context including the iconic Chrysler Building, and street-level perspective as expected.

Cost Decision

High costs about 3.9x more than medium for a 0.047-point lift. Medium costs about 3.9x more than low for a 0.162-point lift. That makes medium the cleanest default: it buys most of the available quality without pushing into premium pricing.

Which GPT Image 2 Quality Should You Use?

A simple rule covers most cases. Default to medium — it clears the bar for almost all production work at a quarter of high's price. Drop to low when you are iterating, exploring directions, or generating throwaway volume; on GPT Image 2 the low tier is genuinely usable, not a degraded fallback. Step up to high only when you hit a specific constraint — small or dense text, a close-up face, fine material detail, or a final render that has to print large. The 3.9x price jump from medium to high buys just 0.047 points on average, so make it a deliberate choice, not a default.

Recommendation