OpenAI shipped GPT Image 2 on April 21, 2026 with native typography across 50+ languages, web-search-grounded thinking mode, and multi-reference image editing (OpenAI, 2026). As of May 5, both text-to-image and edit flows are live on Teno. For ad creative teams, that's the first time text-heavy, multilingual layouts feel like a first-class generation job instead of a Photoshop cleanup step.
Key Takeaways
- Both text-to-image and edit flows ship today, May 5, 2026, on every paid plan and on the free tier at low quality.
- Multilingual typography is the headline capability. The model renders accurate text across 50+ scripts including Arabic, Japanese, Korean, and Devanagari (Microsoft, Azure AI Foundry Blog, April 2026).
- Three quality tiers (low / medium / high) separate prompt exploration, reviewable drafts, and final renders without switching models.
- Use it alongside Nano Banana Pro, not instead of it. OpenAI's release owns typography and precise edits; Nano Banana Pro still owns recurring character and product-hero consistency.
This is the launch post and, until we publish a longer hands-on guide, the canonical Teno write-up on GPT Image 2: what it does, where it fits, and when to pick it over the rest of the image stack.
What is GPT Image 2, and what's actually new?
OpenAI's April 21, 2026 release is a next-generation image model with native typography accuracy across 50+ languages, thinking mode with web search, up to eight coherent variants per prompt, and image-to-image editing with flexible multi-reference inputs (OpenAI, 2026). On Teno, the ad-team subset is exposed today: prompt, image size, quality, number of images, output format, and (on edit) multi-reference image URLs.
Four capabilities matter for paid creative:
- Multilingual typography that survives small sizes. Most 2024–2025 image models drew English text passably and mangled CJK or Arabic. This generation renders accurate copy across Arabic, Japanese, Korean, Chinese, Hindi, Bengali, Cyrillic, Devanagari, and Latin scripts (Microsoft, Azure AI Foundry Blog, April 2026). For cross-border DTC teams shipping packaging mockups or feed-card copy in 5+ markets, that removes a Photoshop step that used to gate every AI-generated background.
- Thinking mode with web search. The model can fetch real-time references before drawing, which matters when a prompt depends on a current product, logo, or seasonal context. Latency is the trade-off, so we'd reserve it for hero renders rather than batch tests.
- Up to 8 variants per prompt with consistency. A single generation returns a coherent set, not eight independent rolls. Useful for grid layouts, A/B variant batches, or storyboard frames in one shot.
- Precise instruction following on multi-object scenes. Spatial constraints ("logo top-right, product center, hand entering from frame-left") hold more reliably than on prior OpenAI image models, which is why the Teno team description calls out fine typography and precise image edits.
How should teams set GPT Image 2 quality on Teno?
At the default landscape_4_3 size, high quality costs 3 credits per image on Teno's May 2026 model catalog, while low and medium presets cost 1 credit each for the same canvas. Treat the quality dial as a workflow control first: low for prompt exploration, medium for reviewable internal drafts, and high for final creative where typography, product detail, and layout precision need to survive export.
The practical quality map:
- Low: fastest way to test prompt direction, rough layout, language variants, and copy placement.
- Medium: best checkpoint for creative review when the idea is close but not ready for paid placement.
- High: final render setting for packaging, multilingual social cards, hero product shots, and any asset with readable text on the canvas.
Live credit-to-USD conversion follows your plan; see pricing for the exact rate. The operating rule is simpler: don't judge a prompt from one high-quality roll. Explore at low or medium, then re-render the winner at high.
That low-to-high pattern mirrors how experienced video teams already test Seedance 2.0 or Veo prompts: separate creative exploration from final export so the team makes better taste calls before polishing the asset.
When should you reach for GPT Image 2 vs. Nano Banana Pro or Nano Banana 2?
OpenAI documents typography across 50+ languages on this model (OpenAI, 2026), so for most paid-creative teams it's now the default pick for anything with text on the canvas: packaging mockups, multilingual social cards, OOH-style ad layouts, and price/CTA overlays. Nano Banana Pro remains the right pick for recurring-character or recurring-product hero shots across a campaign. Nano Banana 2 fits high-volume background and lifestyle work. None of the three replaces the others; the stack splits cleanly by job.
A quick cross-model map:
| Model | Best at | Honest trade-off |
|---|---|---|
| Nano Banana 2 | Volume lifestyle creative; flexible background generation | Typography on the image is the weakest of the three |
| Nano Banana Pro | Commercial-grade character and product consistency across a campaign | Less compelling when typography is the main deliverable |
| GPT Image 2 | Typography across 50+ languages; precise multi-object scenes; multi-reference edits | Recurring character likeness drifts more than Nano Banana Pro |
The decision rule we'd give a paid creative team:
- Anything with text on the canvas → GPT Image 2. Packaging, price tags, CTAs, multilingual feed cards, OOH-style mockups.
- Recurring brand character or hero product across many frames → Nano Banana Pro. Founder cam, mascot, signature SKU.
- High-volume backgrounds and lifestyle scenes → Nano Banana 2. Strong choice when typography isn't the deliverable.
- Start-and-end frames feeding Seedance 2.0 image-to-video → any of the three, picked by which capability matters most for that specific transition.
For where this fits in the broader paid-creative stack, see our AI ad creative platform guide.
What ad-creative jobs is GPT Image 2 actually best for?
The four jobs where this model noticeably outperforms its peers on Teno are multilingual social cards, packaging and label mockups, multi-object shot composition, and reference-based brand edits. All four reduce to the same underlying capability: accurate small-text rendering plus tight instruction following on spatial constraints, which is the gap OpenAI optimized this generation for (OpenAI, 2026).
A few worked examples:
- Multilingual social card (DTC, 4 markets in one batch). Prompt: a single product hero, four language variants of the same headline (EN, DE, JP, AR) requested in one generation. Typography holds correctly across all four scripts; most prior models would corrupt the Arabic and produce near-glyphs in Japanese.
- Packaging mockup with regulatory copy. Skincare or supplement category, where the back-of-pack copy is small, dense, and language-specific. Generate the front and back panels in one prompt; ship to legal review with real readable text instead of placeholder Latin lorem.
- Multi-object scene with strict layout. "Hero product center, brand logo top-right corner, accent prop at frame-left, all on a soft gradient background, soft daylight." Spatial constraints hold; earlier image models tended to migrate the logo into the center.
- Reference-based edits. Send 2–3 reference images plus a prompt to the edit endpoint and the model composes a new scene that respects the references. The cleanest use case is updating an existing hero image with new seasonal copy without regenerating the entire layout.
Our finding: Across the typography-heavy ad concepts we ran in early May 2026, GPT Image 2 usually reached ad-grade quality on the first or second roll for multilingual prompts, where the previous-generation models we'd tested often needed 4–6 retries plus manual text overlay.
That pass was a launch-readiness check, not a formal benchmark: we focused on multilingual feed cards, packaging-style layouts, and reference-based edits where readable on-canvas copy mattered more than photorealism. The pattern was consistent enough to change our default recommendation for text-heavy ad assets from "generate background, then add text manually" to "try GPT Image 2 first, then reserve manual cleanup for edge cases."
For start- and end-frame anchors that feed Teno's video models, pick the image flow that exposes the model you need: Product Photo for its supported stack (including Nano Banana Pro on product-led workflows), and Asset Generator when those frames should be built with GPT Image 2. It is only available there today.
Frequently asked questions
When did OpenAI release GPT Image 2, and when did it arrive on Teno?
OpenAI released GPT Image 2 on April 21, 2026, rolling out to the API, Codex, and ChatGPT for all users. Teno wired both the text-to-image and edit endpoints into Asset Generator on May 5, 2026, available across every paid plan and on the free tier at the low quality preset.
Should I use GPT Image 2 instead of Nano Banana Pro for product photos?
Different jobs. GPT Image 2 wins on typography-heavy creative: packaging, multilingual social cards, instruction-rich prompts, and precise multi-reference edits. Nano Banana Pro wins on recurring character or product-hero consistency across a campaign. The choice is capability fit, not price.
Does GPT Image 2 support Chinese, Japanese, Korean, and Arabic text on images?
Yes. OpenAI documents native typography across 50+ languages, including Arabic, Japanese, Korean, Chinese, Hindi, and Bengali, plus Cyrillic, Devanagari, and Latin scripts. For cross-border ad teams, this collapses one Photoshop or Figma step previously needed to overlay localized copy on AI-generated backgrounds.
Which quality preset should I use for GPT Image 2 on Teno?
Use low for first-pass prompt exploration, medium for reviewable drafts, and high for final ad assets where typography, product detail, and composition need to hold up. That keeps quality decisions tied to the stage of the creative process rather than treating credits as the main selection factor.
Can I edit existing images with GPT Image 2 on Teno?
Yes. The image-to-image edit endpoint is live on Teno from day one, with multi-reference image input, the same six canonical sizes, and the same low/medium/high quality dial. Default size on edit is auto, which lets the model preserve the source aspect ratio rather than reformat your asset.
How to use GPT Image 2 on Teno today
Open Asset Generator, pick GPT Image 2 from the model dropdown, choose between text-to-image and edit, set the size, set the quality, and write the prompt. For edit, attach 1–3 reference images alongside the prompt. Default quality is high; default size on text-to-image is landscape_4_3 and on edit is auto, which preserves the source aspect ratio.
If you're testing prompts at scale, set the quality dropdown to low for the first few rolls, then re-render the winner at high. The free trial is open: start a Teno account and test where the model is strongest, text-heavy product visuals, localized social cards, and precise reference-based edits.
For the studio-level view of how GPT Image 2 fits alongside our other 13 image and video models, including the decision matrix and credit cost table for the full 14-model catalog, see the Asset Generator launch post.
Sources
- OpenAI, Introducing ChatGPT Images 2.0, April 21, 2026, retrieved 2026-05-23.
- OpenAI, GPT Image 2 Model (API documentation), retrieved 2026-05-23.
- Microsoft, Introducing OpenAI's GPT-image-2 in Microsoft Foundry, Azure AI Foundry Blog, retrieved 2026-05-23.
- Teno model catalog, May 2026 (credit calibration, quality-tier multipliers, and canonical sizes). Internal documentation; live credit-to-USD conversion reflected on /pricing.
