Hi there, I could use some help.

Hi there, I could use some help.
A few days ago, my product, which is built on Cloudflare Workers and integrates with Workers AI (GPT-OSS-120b) through bindings, broke because the interface changed (

prompt_tokens

prompt_tokens

prompt_tokens

prompt_tokens and

completion_tokens

completion_tokens

completion_tokens

completion_tokens were replaced with

input_tokens

input_tokens

input_tokens

input_tokens and

output_tokens

output_tokens

output_tokens

output_tokens).
And today, I ran into another issue: the

instructions

instructions

instructions

instructions parameter stopped working when the input was in an OpenAI-like format (like

[{role: "assistant", content: "foobar" }, ...]

[{role: "assistant", content: "foobar" }, ...]

[{role: "assistant", content: "foobar" }, ...]

[{role: "assistant", content: "foobar" }, ...]).

Where can I check for these kinds of interface changes in advance so I can update my product before it breaks?

samjs•11/17/25, 4:41 PM

Hey @boke0 -- sorry about that, that was our bad. We've been updating our gpt-oss implementations to try and bring it to be consistent with the openai responses API. But we overlooked people who might be depending on the current behaviour.

With respect to the latter issue. It sounds like accepting that may have been a bug in the first place? From the OpenAI docs,

instructions

instructions

should be a string.

samjs•11/17/25, 4:41 PM

Either way though, we should have done a better job communicating those changes.

boke0OP•11/17/25, 4:53 PM

Thanks for the response!
It’s helpful to know that I can follow the OpenAI specifications when implementing features with GPT-OSS.
I’d also really appreciate it if you could consider providing announcements or communication around updates in the future.

I’ll continue to rely on the great service that Workers AI provides!

samjs•11/17/25, 4:55 PM

Thanks for your understanding!

Celestial Rose•11/20/25, 11:37 AM

Well instructions seemed to have stopped working for me as well, despite being a string ! I looked for

instructions

instructions

on the discord server, knew id find something. this stopped working on monday for me too. any change to the API ? are changes deployed without new workerd/wrangler versions/changes ?

Celestial Rose•11/20/25, 11:38 AM

Here's what worked perfectly until now

  const aiResponse = await env.AI.run("@cf/openai/gpt-oss-120b", {
                            instructions,
                            input: markdownContent,
                        });

  const aiResponse = await env.AI.run("@cf/openai/gpt-oss-120b", {
                            instructions,
                            input: markdownContent,
                        });

Celestial Rose•11/20/25, 11:38 AM

both are strings

Celestial Rose•11/20/25, 1:13 PM

nice QoL would be to have changelogs and dates for documentation updates .. https://developers.cloudflare.com/workers-ai/models/gpt-oss-120b/ I have no clue when this got updated

(these are only propositions btw)

Cloudflare Docs

gpt-oss-120b

OpenAI’s open-weight models designed for powerful reasoning, agentic tasks, and versatile developer use cases – gpt-oss-120b is for production, general purpose, high reasoning use-cases.

Celestial Rose•11/20/25, 1:28 PM

btw @samjs the actual Responses API requires the content key to be an object with type "input_text" and text in the "text" key. Currently this doesn't work using Workers AI, and content only works when its a string. So we're in the middle of compliance !

boke0OP•11/21/25, 8:55 AM

It sounds like accepting that may have been a bug in the first place? From the OpenAI docs, instructions should be a string.

Oh, I forgot to respond to this issue earlier. Previously, I wrote the code like this:

const aiResponse = await env.AI.run("@cf/openai/gpt-oss-120b", {
  instructions,
  input: [{ role: 'user', content: message }],
});

const aiResponse = await env.AI.run("@cf/openai/gpt-oss-120b", {
  instructions,
  input: [{ role: 'user', content: message }],
});

I resolved the issue by rewriting it as follows:

const aiResponse = await env.AI.run("@cf/openai/gpt-oss-120b", {
  input: [
    { role: 'developer', content: instructions },
    { role: 'user', content: message }
  ],
});

const aiResponse = await env.AI.run("@cf/openai/gpt-oss-120b", {
  input: [
    { role: 'developer', content: instructions },
    { role: 'user', content: message }
  ],
});

However, this change resulted in increased token usage. It used to consume around 7–8 Newrons, but now it consumes 20–23. As @Celestial Rose mentioned, it might be because instructions isn’t actually supported, and switching to a system message could be causing the additional token overhead.

Celestial Rose•11/21/25, 8:58 AM

The think is, the Responses API (as per OpenAI’s doc) doesn’t allow for the content key to be a string ; it should be an object where type is always input_text, and text key with the content value

Celestial Rose•11/21/25, 8:59 AM

Is the token usage increase justified ?

Hi there, I could use some help.

Similar Threads

Similar Threads

Similar Threads