No frontier model “draws best” across the board, and the head-to-head visual test that circulated on Hacker News around July 22 does not crown one either. The comparison ran GPT-5.6, Claude, Gemini, and Grok against identical prompts, but the aggregator behind it names only three dedicated text-to-image models in its FAQ (Nano Banana Pro, GPT Image 2, and FLUX 2), none of them Anthropic or xAI products. For teams routing visual generation in production, the useful question was never which chat brand wins. It is which image model wins each subtask, and the per-cell results of this particular test remain unverified as of 2026-07-28.
What did the July 22 drawing test actually compare?
The test compared four chat-model brands on visual output, but the aggregator behind it names only three text-to-image models, and none of those three is an Anthropic or xAI product, which makes a clean four-way ranking impossible.
The reported setup is straightforward: one prompt, four models, ranked outputs, surfaced on Hacker News around July 22. What the headline misses is what each cell in that grid actually measures. TryAI markets itself as a single app for comparing 30-plus frontier models across chat, image, video, and music, and it names GPT-5.5, Claude Opus, Gemini 3 Pro, and Grok as its chat-tier comparison set. Its text-to-image roster, per the site’s FAQ, is a shorter list: Nano Banana Pro, GPT Image 2, and FLUX 2. None of those three is an Anthropic or xAI product, even though both Anthropic and xAI sit in the chat tier.
So when a four-way test reports that “Claude” or “Grok” drew something, the cell is measuring whatever image backend those chat products call, not a same-branded dedicated image model. The named text-to-image roster has no Anthropic or xAI entry. Ranking all four chat brands as if they were equivalent instruments conflates categories that the aggregator’s own lineup keeps separate, and any conclusion drawn from the aggregate inherits that mismatch.
The GPT-5.6 label is not a mystery either. TryAI’s model directory lists GPT-5.6 Sol, GPT-5.6 Terra, and GPT-5.6 Luna as OpenAI’s current flagship family, with GPT-5.5 explicitly marked as “now succeeded by the GPT-5.6 family.” The TryAii Arena ranking is a different artifact and tops out at gpt-5.2, but that is an Arena-rating lag, not evidence the test mislabeled anything.
None of this means the test was worthless. Side-by-side prompt comparisons are exactly the right instinct for visual routing. The problem is the framing: a four-brand aggregate rank implies a single winner, and the underlying lineup shows the brands are not even competing with the same kind of engine.
Which models actually produce the images?
In the aggregator’s lineup, pixels come from the dedicated image models in its text-to-image roster (Nano Banana Pro, GPT Image 2, and FLUX 2), not from the chat frontier models whose names dominate the leaderboards.
What the aggregator does not do is tell you which of those three fits your subtask. TryAI’s model directory names its text-to-image options, but the cached version of that page does not surface per-model positioning copy, specialty labels, or per-subtask scores. A buyer sees three names and no guidance on which one handles photorealism, which handles in-image text, and which follows long prompts.
Image generation is not one capability. Photorealism, multi-reference compositing, instruction fidelity, and typography are genuinely different training targets, and a model optimized for one tends to spend capacity the others need. That is why the “which model draws best” question has no stable answer: the winner flips depending on which of those cells you are testing.
The training-target tradeoff is structural, not a defect the next release patches out. A model tuned to render photorealistic skin and lighting spends its parameter budget and its feedback signal on pixels that read as real at a glance, which is a different objective from rendering a short headline so the letters do not bleed into the background. The same split appears in multi-reference compositing, where the model must hold several input images in working memory and respect each one, and in instruction fidelity over long prompts, where the real test is whether the final clause survives generation. None of these axes tracks the others cleanly, which is why a single slot on an image-quality leaderboard compresses exactly the information a routing decision needs.
Vendor materials reinforce the dedicated-model pattern. Google’s ai.google page, published 2026-07-01, showcases Gemini handling image tasks like transforming a sketch into a realistic image and retouching headshots. OpenAI’s ChatGPT Android app, updated 2026-07-24, advertises an image generator that creates original images from a description or transforms existing ones. In both cases the image capability is a named feature of a product, backed by a dedicated model, not an emergent property of the chat model.
How should you route visual generation by subtask?
Route by subtask, not by brand. The aggregator names three text-to-image models (Nano Banana Pro, GPT Image 2, and FLUX 2) and gives no per-subtask guidance for any of them. The only way to assign each to the right cell is to test all three against your own workload.
The working matrix, built from what the aggregator actually publishes:
| Subtask cell | What to test |
|---|---|
| Photorealistic generation and photo editing | Run Nano Banana Pro, GPT Image 2, and FLUX 2 on your photo prompts; the aggregator publishes no per-subtask scores |
| In-image text, typography, and multi-reference compositing | Same three models on text-heavy and reference-heavy prompts; these axes are not rated by the roster |
| Prompt-adherence-critical generation | Same three models on long, constrained prompts; instruction fidelity is not documented |
| High-volume, cost-sensitive generation | Compare the cheapest tier that clears your quality bar against the frontier options |
Two mechanisms sit underneath this table. The first is the cost gradient. TryAI explicitly invites users to “compare a $30/M frontier model against a $0.27/M open model.”1 That is roughly a 110x price spread, and for visual subtasks where the cheap model clears your quality bar, paying frontier rates is pure waste. The only way to know which cells those are is to test them; the price tag is not a quality proxy in either direction.
The second mechanism is churn. Model directories rotate, and the cached pages already show one example: GPT-5.5 marked as superseded by the GPT-5.6 family. A routing layer that hardcodes model names instead of subtask assignments will break on that schedule. Route by capability cell, resolve the cell to a current model at the edge, and keep the mapping table somewhere you can edit without a deploy.
Why don’t chat leaderboard scores predict drawing ability?
Chatbot Arena ratings measure conversational preference, not image quality, so a model’s position on the chat leaderboard carries no information about its drawing cell.
The numbers make this concrete. On TryAii’s Arena-sorted list, gemini-3-pro-preview sits at 1501, claude-opus-4-5 at 1493, gpt-5.2 at 1490, grok-4-latest at 1480, and grok-4-1-fast-reasoning-latest at 1455.2 If chat rank predicted visual rank, you would expect Gemini to win every drawing cell and the two Grok variants to lose all of them. There is no evidence in any fetched source that this ordering holds for images, and the structural reason it would not is already visible: the chat models and the image models are different artifacts. Gemini’s chat rating is a property of gemini-3-pro-preview, a conversational model; the image models live in a separate roster. Conflating the two is a category error with a leaderboard attached.
It is telling that TryAii, the same site publishing those Arena rankings, builds its entire product on the opposite premise. It positions itself as a routing engine where you describe a task and the system picks the model that fits, arguing that “generic picks miss the per-prompt best.” Its own marketing claims that three fixed-tier picks cover the easy cases but miss the long tail, and that it routes every prompt across all 30 mapped models sorted by Arena rating. Even granting that the Arena sort is doing the work there, the architectural claim is the right one: per-prompt routing beats fixed-tier picking because capability is not uniform across a model’s surface.
The aggregate-score problem generalizes beyond Arena. Any single-number multimodal ranking averages across cells, photorealistic, schematic, in-image text, instruction following, and the average hides exactly the splits that decide routing. A model can win the aggregate while losing the one cell your pipeline lives in. Vendors have little incentive to surface those splits; a single rising number is a better marketing asset than a matrix with some weak cells. That gap in how multimodal capability gets reported is the second-order story here, and it transfers the benchmarking burden onto the buyer.
The same incentive shapes how image models get benchmarked publicly. A vendor shipping a new text-to-image model has every reason to publish the cell it wins and leave the cells it loses unmeasured, and the aggregator’s FAQ, which lists three image names and no per-subtask data, fits that pattern. The buyer inherits the measurement work as a result. The only matrix that matters for routing is the one your own prompts produce, and the aggregate leaderboard above is a marketing surface where a per-cell matrix is an engineering artifact. The two rarely agree on which model to pick.
What does a production routing decision look like?
The practical decision procedure is: enumerate the visual subtasks your pipeline actually performs, map each to candidate image models, run a small cell-level test on your own prompts, pin versions, and re-test on a schedule driven by vendor lineup changes.
Concretely, that looks like this. First, write down your cells. A documentation pipeline might have three: schematic diagrams, hero images, and screenshots with synthetic UI text. A marketing pipeline has different ones. The four-way drawing test’s cells (photorealistic, schematic, instruction-following) are a decent generic starting grid, but yours should come from your prompt distribution, not from someone else’s test design.
Second, assign candidates per cell. The aggregator names three text-to-image models (Nano Banana Pro, GPT Image 2, and FLUX 2) and publishes no per-subtask scores, so every assignment starts as a hypothesis. Test all three against each cell rather than assuming a specialty. Include at least one cheap tier per cell, because the frontier-versus-open price spread only matters if you actually measure where it stops mattering.
Third, run the test yourself. A few dozen representative prompts per cell, scored against criteria you write down in advance, is enough to catch a model that is badly wrong for your workload. This is the step the Hacker News test was trying to crowdsource, and the instinct is correct; the flaw was doing it with someone else’s prompts and reading the aggregate instead of the cells.
Scoring criteria matter more than sample size. A rubric written in advance, checking whether a schematic has all labeled parts, whether a hero image matches the brand palette, whether synthetic UI text renders legibly at the target resolution, forces a decision you can defend. Reading outputs after the fact invites you to rationalize whatever the model happened to ship. A modest prompt set scored against a written rubric beats a large one scored by gut. Pin the model version alongside each result, because a routing decision made against FLUX 2 in July is not evidence about whatever FLUX ships in October; the version string is part of the measurement, not metadata attached to it.
Fourth, treat the mapping as configuration with an expiry date. The cached directory already shows GPT-5.5 marked as superseded by the GPT-5.6 family, and every vendor on these lists rotates tiers. Quarterly re-validation is the minimum cadence that keeps a visual pipeline honest.
What remains unverified about the four-way test?
The per-cell results of the July 22 test, which model won the photorealistic, schematic, and instruction-following cells, could not be verified against any fetched source as of 2026-07-28, and neither could the Hacker News point count. The GPT-5.6 label is catalogued in the aggregator’s chat tier; the specific GPT-5.6 tier used in each cell is not.
That limitation bounds everything above. The routing framework, the dedicated-versus-chat-model distinction, and the critique of aggregate multimodal scores all stand on the aggregator documentation cited throughout. But any claim of the form “GPT-5.6 won photorealistic” or “Claude won schematics” originating from this specific test should be treated as unverified until the per-cell data surfaces.
The verdict, then, is conditional but not weak. Route visual generation by subtask, test the aggregator’s three named image models (Nano Banana Pro, GPT Image 2, FLUX 2) against each cell on your own prompts, and treat chat-leaderboard ratings as irrelevant to drawing. The aggregator publishes no per-subtask scores, so the routing table is yours to build. Expect it to need revision the next time a vendor refreshes a tier.
The drawing test asked which frontier model draws best. The more useful answer its own lineup supplies is that frontier chat brands do not draw at all; image models do, each on a different cell, and the teams that route accordingly stopped caring about the aggregate rank some time ago.
Frequently Asked Questions
Which specific GPT-5.6 tier did the July 22 test use?
The aggregator’s directory lists GPT-5.6 Sol, Terra, and Luna as the current flagship family, but the test documentation does not specify which tier was routed for each image cell. Without that version pin, the results cannot be reproduced or compared against other runs using different tiers in the same family.
How does the $30M frontier cost compare to open weights for image generation?
The aggregator explicitly invites users to compare a $30M frontier model against a $0.27M open model, creating an approximately 110x price spread. This spread means that for subtasks where the open model clears the quality bar, paying frontier rates is pure waste, but the only way to know which cells those are is to test them against your own prompts.
Why do Chatbot Arena ratings fail to predict image quality?
Arena ratings measure conversational preference, not pixel fidelity, so a model’s position on the chat leaderboard carries no information about its drawing cell. For example, gemini-3-pro-preview ranks 1501 on TryAii while claude-opus-4-5 sits at 1493, yet these are properties of conversational models that are structurally distinct from the dedicated image models in the aggregator’s text-to-image roster.
What happens when a vendor retires a model tier?
Model directories rotate on a schedule that breaks pipelines hardcoding specific model IDs. The cached directory already shows GPT-5.5 marked as superseded by the GPT-5.6 family, and the original Nano Banana is scheduled to retire on October 2, 2026. Routing by capability cell and resolving the cell to a current model at the edge prevents these breaks.
