DeepSeek V4 is now available on Vercel AI Gateway.

There are 2 model variants: DeepSeek V4 Pro and DeepSeek V4 Flash. A 1M token context window is the default across both models.

DeepSeek V4 Pro focuses on agentic coding, formal mathematical reasoning, and long-horizon workflows. It handles feature development, bug fixing, and refactoring across stacks, with tool use that works across harnesses like MCP workflows and agent frameworks. It also writes clear, well-structured long-form documents.

DeepSeek V4 Flash performs close to V4 Pro on reasoning and holds up on simpler agent tasks, with a smaller parameter size for faster responses and lower API cost. It's a good fit for high-volume workloads and latency-sensitive use cases.

To use DeepSeek V4, set model to deepseek/deepseek-v4-pro or deepseek/deepseek-v4-flash in the AI SDK.

1
import { streamText } from 'ai';
2

3
const result = streamText({
4
  model: 'deepseek/deepseek-v4-pro', // or 'deepseek/deepseek-v4-flash'
5
  prompt:
6
    `Audit this repository for unsafe concurrent access patterns,
7
     propose a refactor that introduces proper synchronization,
8
     and open the changes as a PR with a migration plan.`,
9
});

AI Gateway provides a unified API for calling models, tracking usage and cost, and configuring retries, failover, and performance optimizations for higher-than-provider uptime. It includes built-in custom reporting, observability, Bring Your Own Key support, and intelligent provider routing with automatic retries.

Learn more about AI Gateway, view the AI Gateway model leaderboard or try it in our model playground.

GPT Image 2 is now available on Vercel AI Gateway.

OpenAI's newest image model supports detailed instruction following, accurate placement and relationships between objects, and rendering of dense text across multiple aspect ratios.

The model can render fine-grained elements including small text, iconography, UI elements, dense compositions, and subtle stylistic constraints, at up to 2K resolution. Non-English text is also supported and reads coherently.

GPT Image 2 can produce photos, cinematic stills, pixel art, manga, and other distinct visual styles, with consistency in texture, lighting, composition, and detail. This suits workflows like game prototyping, storyboarding, marketing creative, and medium-specific asset generation.

To use GPT Image 2, set model to openai/gpt-image-2 in the AI SDK, or try it directly in our model playground.

1
import { generateImage } from 'ai';
2

3
const result = await generateImage({
4
  model: 'openai/gpt-image-2',
5
  prompt: 'Poster of Vercel AI products, Bauhaus style.',
6
});

AI Gateway provides a unified API for calling models, tracking usage and cost, and configuring retries, failover, and performance optimizations for higher-than-provider uptime. It includes built-in custom reporting, observability, Bring Your Own Key support, and intelligent provider routing with automatic retries.

Learn more about AI Gateway, view the AI Gateway model leaderboard or try it in our model playground.

Kimi K2.6 from Moonshot AI is now available on Vercel AI Gateway.

The model focuses on long-horizon coding tasks, with generalization across languages such as Rust, Go, and Python and across front-end, devops, and performance optimization work. K2.6 can turn simple prompts into complete front-end interfaces with structured layouts.

For autonomous, proactive agents that run continuously across multiple applications, K2.6 improves on API interpretation, long-running stability, and safety awareness during extended research tasks.

To use Kimi K2.6, set model to moonshotai/kimi-k2.6 in the AI SDK.

1
import { streamText } from 'ai';
2

3
const result = streamText({
4
  model: 'moonshotai/kimi-k2.6',
5
  prompt:
6
    `Build a landing page for a developer tools product,
7
     including a hero section with scroll-triggered
8
     animations and generated imagery.`,
9
});

AI Gateway provides a unified API for calling models, tracking usage and cost, and configuring retries, failover, and performance optimizations for higher-than-provider uptime. It includes built-in custom reporting, observability, Bring Your Own Key support, and intelligent provider routing with automatic retries.

Learn more about AI Gateway, view the AI Gateway model leaderboard or try it in our model playground.

Retention policies no longer delete the latest preview deployment for branches with open or unmerged pull requests. Previously, deployments for active branches could be removed if they exceeded the configured retention window.

This means you can safely use shorter retention windows without risking losing active preview deployments. This change applies to all plans.

Your 10 most recent production deployments and any aliased deployments continue to be preserved regardless of retention settings.

Learn more about Deployment Retention.

Vercel Flags is now generally available.

Vercel Flags is a feature flag provider built into the Vercel platform. Create and manage feature flags with targeting rules, user segments, and environment controls directly in the Vercel Dashboard.

The Flags SDK provides a framework-native way to define and use these flags within Next.js and SvelteKit applications, integrating directly with your existing codebase:

1
import { vercelAdapter } from "@flags-sdk/vercel"
2
import { flag } from "flags/next"
3

4
export const showNewFeature = flag({
5
  key: "show-new-feature",
6
  adapter: vercelAdapter()
7
})

Once you define a flag, you can use them within your application in a few lines of code:

1
import { showNewFeature } from "~/flags"
2

3
export default async function Page() {  
4
  const isEnabled = await showNewFeature()
5

6
  return isEnabled ? <NewDashboard /> : <OldDashboard />
7
}

For teams using other frameworks or custom backends, the Vercel Flags adapter supports the OpenFeature standard, allowing you to plug Vercel Flags into their provider agnostic SDK.

Try it out or learn more about Vercel Flags.

Claude Opus 4.7 from Anthropic is now available on Vercel AI Gateway.

Opus 4.7 is optimized for long-running, asynchronous agents and handles complex, multi-step tasks with reliable agentic execution. The model shows gains on knowledge-worker tasks, particularly where it needs to visually verify its own outputs.

Opus 4.7 is also stronger at programmatic tool-calling with image-processing libraries to analyze charts and figures, including pixel-level data transcription. It has high-resolution image support, which is useful for computer use, screenshot understanding, and document analysis workflows. Opus 4.7 now has improved memory, with agents that maintain structured memory store across turns seeing more reliable recall and fewer dropped facts without additional prompting.

To use Claude Opus 4.7 set model to anthropic/claude-opus-4.7 in the AI SDK. You can also try a new effort level: 'xhigh'.

1
import { streamText } from 'ai';
2

3
const result = streamText({
4
  model: 'anthropic/claude-opus-4.7',
5
  prompt: 'Explain the halting problem in one paragraph.',
6
  providerOptions: {
7
    anthropic: {
8
      thinking: { type: 'adaptive' },
9
      effort: 'xhigh',
10
    },
11
  },
12
});

Opus 4.7 also introduces the task budgets feature. Task budgets let you set a total token budget for an agentic turn via taskBudget. The model sees a countdown of remaining tokens, which it uses to prioritize work, plan ahead, and wind down gracefully as the budget is consumed. Thinking content is also now omitted by default for Opus 4.7. To receive thinking content, set display to 'summarized':

1
import { streamText } from 'ai';
2

3
const result = streamText({
4
  model: 'anthropic/claude-opus-4.7',
5
  prompt: 'Research how this codebase handles authentication and suggest improvements.',
6
  providerOptions: {
7
    anthropic: {
8
      thinking: { type: 'adaptive', display: 'summarized' },
9
      effort: 'high',
10
      taskBudget: { type: 'tokens', total: 50000 },
11
    },
12
  },
13
});

AI Gateway provides a unified API for calling models, tracking usage and cost, and configuring retries, failover, and performance optimizations for higher-than-provider uptime. It includes built-in custom reporting, observability, Bring Your Own Key support, and intelligent provider routing with automatic retries.

Learn more about AI Gateway, view the AI Gateway model leaderboard or try the model in our model playground.

You can now access Bytedance's latest state-of-the-art video generation model, Seedance 2.0, via AI Gateway with no other provider accounts required.

Seedance 2.0 is available on AI Gateway in two variants: Standard and Fast. Both share the same capabilities. Standard produces the highest quality output, while Fast prioritizes generation speed and lower cost.

Seedance 2.0 is strong at maintaining motion stability and fine detail across frames, producing consistent output even in complex scenes with facial expressions and physical interactions. The model also generates synchronized audio natively, with support for speech in multiple languages and dialects.

Beyond text-to-video and image-to-video, Seedance 2.0 adds multimodal reference-to-video, letting you combine image, video, and audio inputs as reference material in a single generation. It also supports video editing and video extension, along with professional camera movements, multi-shot composition, and in-video text rendering.

To use this model, set model to bytedance/seedance-2.0 or bytedance/seedance-2.0-fast in the AI SDK or try it out in the AI Gateway Playground.

Text to Video

Generate video from a text prompt. Describe the scene, camera movement, and audio for the model to produce.

1
import { experimental_generateVideo as generateVideo } from 'ai';
2

3
const { videos } = await generateVideo({
4
  model: 'bytedance/seedance-2.0',
5
  prompt:
6
   `Black triangle sticker peels off laptop and zips across the office. It smashes
7
    through the window and into the San Francisco sky.`,
8
  aspectRatio: '16:9',
9
  resolution: '720p',
10
  duration: 5,
11
});

Image to Video

Generate video from a starting image. The model animates the image based on the text prompt while preserving the visual content of the source frame.

1
import { experimental_generateVideo as generateVideo } from 'ai';
2

3
const { videos } = await generateVideo({
4
  model: 'bytedance/seedance-2.0',
5
  prompt: {
6
    image: catImageUrl,
7
    text: 'The cat is celebrating a birthday with another cat.',
8
  },
9
  duration: 10,
10
  providerOptions: {
11
    bytedance: { generateAudio: true },
12
  },
13
});

Reference to Video

Generate video using image, video, or audio references as source material. You can combine multiple reference types in a single generation to control visual style, motion, and sound.

1
import { experimental_generateVideo as generateVideo } from 'ai';
2

3
const { videos } = await generateVideo({
4
  model: 'bytedance/seedance-2.0',
5
  prompt: 'Replace the cat in [Video 1] with the lion from [Image 1].',
6
  duration: 10,
7
  providerOptions: {
8
    bytedance: {
9
      referenceImages: [Image 1],
10
      referenceVideos: [Video 1],
11
      generateAudio: true,
12
    },
13
  },
14
});

AI Gateway does not charge any markup on video generation: Seedance 2.0 and 2.0 Fast are at the same price as going direct to the Bytedance provider.

Learn more about AI Gateway, view the AI Gateway model leaderboard or try it in our model playground.

Vercel is reducing the price of Turbo build machines by 16%. All builds are now priced at $0.0035 per CPU per minute. With this new model:

• Turbo machines, with 30 CPUs, are now $0.105 per minute (previously $0.126)

• Enhanced machines, with 8 CPUs, continue to be priced at $0.028 per minute

• Standard machines, with 4 CPUs, continue to be priced at $0.014 per minute

This change will begin rolling out on April 27, and will appear on invoices for the current billing cycle as "Build CPU Minutes".

Learn more about build machine pricing or monitor your Builds usage from Project Usage.