GPT Image 2
Image Available Full comparison ↗ openai/gpt-image-2 · by OpenAI
· Autoregressive
Pricing — 1 offering(s)
Low · 1024×1024
- $0.006 / image Current 2026-04-21 → present
Medium · 1024×1024
- $0.053 / image Current 2026-04-21 → present
High · 1024×1024
- $0.211 / image Current 2026-04-21 → present
Showing the active price and any recorded history. Full pricing history is available via the paid API — see API docs.
See how GPT Image 2 fits into a cost-aware routing setup
See how →Capability profile
How GPT Image 2 rates across core capability dimensions, with the task-level evidence behind each rating.
OpenAI's own launch framing emphasizes layout reasoning and output verification specifically to improve prompt fidelity over gpt-image-1.
OpenAI's own launch material specifically claims near-perfect text rendering across Japanese, Korean, Chinese, Hindi, and Bengali, in addition to Latin scripts.
4K output per OpenAI's own launch material — an improvement over gpt-image-1's "moderate" rating (that entry topped out at 1024/1536 square/portrait/landscape, not a high-res print model).
OpenAI's own framing calls it "fast," an explicit improvement claim over gpt-image-1 (rated "weak" on this dimension) — but the added reasoning/web-search stage for complex prompts plausibly reintroduces some latency versus a pure single-pass generation, so not rated "strong" without a direct side-by-side latency figure to cite.
Benchmarks
Operator guidance
OpenAI's previous-generation image model. OpenAI's image guide now lists it among its earlier GPT Image models and recommends GPT Image 2.5 Sunburst or Flare for new integrations. Keep it for existing pipelines tuned to its output; route new text-heavy, multilingual or high-resolution work to GPT Image 2.5. Reasoning ("Thinking mode") is limited to paid plans in the consumer ChatGPT product, so confirm API-side availability for your integration.
Use cases
- Complex instructions with in-image text across multiple scripts/languages
- Agentic/tool-assisted generation — web search before generating, multi-image batching
- 4K-resolution output for print or high-detail use cases
- Editing and reference-guided generation
Provenance notice
- Outputs carry C2PA content-credential metadata, same as gpt-image-1
- Subject to OpenAI safety filtering
Limitations
- Reasoning/'Thinking mode' capability gating on the API side (vs. the consumer subscription tiers) not independently confirmed this pass
- Architecture undisclosed; closed weights, same as gpt-image-1
- Capability ratings are qualitative except where a specific benchmark is cited