GPT-Image 2.5 brings faster API edits
Need a better product shot without rebuilding the whole scene? Or an app idea while your hands are busy? Today’s updates put more creative control, faster models, and lower-risk shipping paths in your toolbox.
In today’s edition
- OpenAI added GPT-Image-2.5 API models with lower latency and tighter editing control.
- Astra is now available across paid ChatGPT tiers, Codex, and ChatGPT Work.
- Mercury 2.5 reaches general availability with parallel decoding and sharply reduced token prices.
- Lovable drafts let teams test parallel project versions before applying changes.
- Bolt now lets you describe an app aloud, then plans, builds, and publishes it.
OpenAI ships two image models for builders
OpenAI has introduced GPT-Image-2.5 Flare and Sunburst in the API, giving builders sharper outputs, stronger style adherence, and more control over edits.
- Flare delivers higher-quality images at 50% lower latency than GPT-Image-2.
- Sunburst is positioned for extra precision across detailed edits.
- The models improve natural lighting, textures, complex layouts, and transparent backgrounds.
- Use cases include image generators, photo editors, image-to-video flows, and design platforms.
What it means Start with Flare for high-volume creative features, then test Sunburst where customers will pay for precise, production-ready edits that preserve the rest of an image.
Source: OpenAIDevs
Astra reaches paid ChatGPT and Codex users
Astra is fully rolled out to Plus, Pro, Business, and Enterprise users in Codex and ChatGPT Work.
- The rollout covers Plus, Pro, Business, and Enterprise plans.
- It is available in both Codex and ChatGPT Work.
- OpenAI points users to a live demonstration for ideas on how to use it.
What it means If your team already has an eligible plan, put Astra into a real build or operations workflow this week and compare its output against your current default.
Source: OpenAI
Mercury 2.5 cuts cost for long outputs
Mercury 2.5 is now generally available on OpenRouter, using diffusion-style parallel token decoding instead of generating one token at a time.
- OpenRouter calls it the fastest model on its Fastest Models leaderboard that does not require specialized hardware.
- The model costs $0.04 per million input tokens and $0.15 per million output tokens.
- OpenRouter says the general-availability input price is 5x below preview pricing and 6x below Mercury 2 on input.
- It has a 260K context window, according to OpenRouter.
What it means Benchmark Mercury 2.5 on latency-sensitive, long-output tasks such as agents and coding flows, where parallel decoding and lower token costs could reshape your unit economics.
Source: OpenRouter
Lovable adds drafts for parallel app ideas
Lovable now lets you create parallel versions of a project, so teams can explore changes without committing them to the main work.
- Drafts create separate versions of a project.
- Teams can explore ideas and tinker freely in parallel.
- Changes are applied only when you are happy with them.
What it means Use drafts for risky feature tests, client revisions, and alternative onboarding ideas so your team can compare outcomes before changing the version users rely on.
Source: Lovable
Bolt lets you build apps by voice
Bolt now accepts spoken app descriptions: hold the mic, describe the name, colors, and layout, and Bolt plans, builds, and publishes the app.
- You can describe an app out loud rather than type the brief.
- Bolt takes inputs covering the app name, colors, and layout.
- The workflow moves from planning to building and publishing.
What it means Try voice prompts when you are sketching a rough concept or collecting stakeholder feedback, then turn the spoken brief into a shareable first version faster.
Source: boltdotnew