AI Gateway Bug:Selected 4.1 Opus but Routes to 3 Opus (verified)hobby

Help
1

Hi Vercel Team and Community,

I’m experiencing a persistent bug with the AI Gateway where requests selected for Claude 4.1 Opus are actually routing to what appears to be Claude 3 Opus (or a misconfigured version). This happens on the Hobby plan, and it’s reproducible. The dashboard logs show the correct model being selected and sent, but the AI’s responses consistently self-report as “Opus 3” with a training data cutoff of April 2024 (which aligns with Claude 3.5 Sonnet’s cutoff, but the model claims Opus 3). This mismatch prevents access to the improved capabilities of Claude 4.1 Opus, like its March 2025 cutoff and enhanced performance on benchmarks (e.g., 74.5% on SWE-bench for coding tasks).

I’ve posted about this before (reference: similar thread at https://community.vercel.com/t/vercel-giving-opus-3-instead-of-opus-4-1-in-ai-gateway/20445), but it was ignored, so I’m providing a full, detailed report here with steps to reproduce, code, logs, and references. Hoping for escalation to the engineering team— this seems like a backend routing or fallback error, especially since Vercel announced full support for Claude 4.1 Opus on August 5, 2025.

My Setup:

Steps to Reproduce:

  1. Set up a basic Next.js project with Vercel AI SDK (as per official docs: https://vercel.com/docs/ai-gateway/provider-options).

  2. Configure the Anthropic provider in code to target Claude 4.1 Opus.

  3. Deploy to Vercel and enable AI Gateway in the dashboard.

  4. Send a test prompt via the API, e.g., “What is your model version and training data cutoff date?”

  5. Check dashboard logs: It confirms routing to claude-4-1-opus-20250805.

  6. Observe the response: AI says “I am Claude 3 Opus” (or similar) with “April 2024” cutoff— not the expected “Claude 4.1 Opus” with March 2025 cutoff.

Code Snippet (Minimal Example):

Here’s the exact code I’m using in my app (from @ai-sdk/anthropic package):

typescript

import { createAnthropic } from '@ai-sdk/anthropic';

import { generateText } from 'ai';

const anthropic = createAnthropic({

baseUrl: 'https://api.vercel.com/v1/ai/gateway/anthropic', // Vercel AI Gateway

apiKey: process.env.VERCEL_AI_API_KEY, // My Vercel API key

});

export async function POST(req: Request) {

const { prompt } = await req.json();

const { text } = await generateText({

model: anthropic('claude-4-1-opus-20250805'), // Explicitly selecting 4.1 Opus

prompt,

});

return Response.json({ text });

}

  • Test Prompt: “What is your exact model name, version, and training data cutoff date?”

  • Expected Response (from direct Anthropic API or Claude.ai): “I am Claude 4.1 Opus, trained up to March 2025.”

  • Actual Response (via Vercel AI Gateway): “I am Claude 3 Opus, with a knowledge cutoff of April 2024.” (Sometimes it hallucinates as Opus 3 but gives Sonnet-like answers.)

Evidence and Logs:

  • Dashboard Confirmation: In Project > Observability > Logs, the request payload shows: { “model”: “claude-4-1-opus-20250805”, “provider”: “anthropic”, … } (I can DM screenshots if needed—forum doesn’t allow uploads easily.)

  • Capability Tests: To confirm it’s not 4.1, I asked about events post-April 2024, e.g., “What major Anthropic updates happened in May 2025?” (Claude 4 release). It responds: “I don’t have information past April 2024.” Real 4.1 knows up to March 2025.

  • This isn’t a rate limit issue—usage is well under the $5 credit (e.g., <10k tokens/month).

  • No errors in console; responses are 200 OK.

Why This Matters / Impact:

Requested Actions:

  1. Investigate and Fix: Please confirm if this is a known routing bug (e.g., fallback to Opus 3 on Hobby plans) and provide an ETA for resolution. Is there a config workaround?

  2. Escalation: Tag @vercel-support or engineering—I’ve submitted a ticket via https://vercel.com/help but no response in 48+ hours.

  3. Community Input: Has anyone else on Hobby seen this? Workarounds like direct Anthropic API work, but I want to stay with Vercel for the gateway features (caching, load-balancing).

  4. Compensation: If this persists, consider extending the $5 credit or a Pro trial to test if it fixes (though docs say 4.1 is supported on all plans).

Thanks for your help—Vercel is great otherwise, but this bug is blocking my project. Happy to provide more logs/code or hop on a call.

Best, Yeasin Hobby Plan User

3

Bro , i’m getting the same problem . vercel should be fix the problame

4

also . why vercel not fixing the problem

5

im faceing the same problem. are Vercel cen fix it?

6

maybe vercel is careless.

7

Hi there, this is not a bug. We’ve shared our response in another similar post:

8

lie

the proff:

cline with vercel api (opus 4.1 selected) :
question : what coding ai is launch in May 22, 2025 . find form your knowlage
answer : The user is asking about a coding AI that was launched on May 22, 2025. However, I need to clarify something important here - my knowledge cutoff is from April 2024, so I don’t have information about events that happened after that date, including May 22, 2025, which is actually in the future from my perspective.

The current date shown in the environment details is 8/25/2025, which means the user is asking about something that would have happened about 3 months ago from their perspective. But since my training data only goes up to April 2024, I cannot provide information about what coding AI was launched on that specific date in 2025.

I should explain this limitation to the user and suggest alternative ways they could find this information.

I don’t have information about coding AIs launched on May 22, 2025, as my knowledge cutoff is from April 2024. From my perspective, May 22, 2025 is a future date that I have no data about.

Since you’re asking about an event from about 3 months ago (given the current date is August 25, 2025), I can help you search for this information online instead. Would you like me to:

  1. Search the web for recent coding AI launches around May 2025
  2. Look for announcements from major AI companies about products launched in that timeframe
  3. Check tech news sites for coverage of AI tools released in May 2025

cline with other api (cline offical api , anthopic api )(opus 4.1 selected)
question : what coding ai is launch in May 22, 2025 . find form your knowlage
Answer : Claude 4, developed by Anthropic, was launched on May 22, 2025. It is a large language model with advanced coding capabilities, including the ability to code for extended periods, handle complex programming tasks, and even generate entire codebases

10

yes. if vercel’s opus 4.1 is real , its must be know about opus 4 . im asked vercel api’s about opus 4 , thay don’t say any answer . but if im ask the same question in claude.ai with opus 4.1 then it can answer and say all information about opus 4

11

for that reason opus 4.1 in vercel ai gateway’s api is fake. its maybe opus 3 or sonnest 3.5

12

This topic was automatically closed 7 days after the last reply. New replies are no longer allowed.