April 14 Livestream: Announcing GPT-4.1 series API-only models (GPT-4.5 deprecated)

[polepole]

What is the mystery of supermassive black hole?

EDIT:

New models in the API

GPT 4.1 in the API

Join Michelle Pokrass, Ishaan Singal, and Kevin Weil as they introduce and demo our latest API models.

https://openai.com/index/gpt-4-1/

https://platform.openai.com/docs/changelog

https://platform.openai.com/docs/models/gpt-4.1

https://platform.openai.com/docs/models/gpt-4.1-mini

https://platform.openai.com/docs/models/gpt-4.1-nano

GPT 4.1 Prompting Guide | OpenAI Cookbook

GPT-4.1 is free between April 14 to April 21, 2025 on https://windsurf.com/

[anon1374209]

its legit

word on the street is that its the quasar alpha open source model, good luck having a 10k+ computer to run this thing…

"it refuse the exact same way Openai models do, it tells you it is made by openai, it has the same tokenizer bug, has the same repetition patterns on creative writing benchmark that other openai model have. "

[vb]

Live at 2025-04-14T17:00:00Z

Here is the link to the YouTube live stream:

New models in the API

Join Michelle Pokrass, Ishaan Singal, and Kevin Weil as they introduce and demo our latest API models.

@polepole: Thanks for posting!

[_j]

• 2**20
• 10**6

Actual answer to the poll by GPT-4.1

(Which was the context window length…)

[phyde1001]

''

[_j]

Starting OpenAI Model Monitor. Press ctrl-c to stop.
...................
ALERT: Model added 2025-04-14 10:18:25 gpt-4.1

ALERT: Model added 2025-04-14 10:18:25 gpt-4.1-2025-04-14

ALERT: Model added 2025-04-14 10:18:25 gpt-4.1-mini

ALERT: Model added 2025-04-14 10:18:25 gpt-4.1-mini-2025-04-14

ALERT: Model added 2025-04-14 10:19:26 gpt-4.1-nano

ALERT: Model added 2025-04-14 10:20:27 gpt-4.1-nano-2025-04-14

[aprendendo.next]

Wow, nice new context window.

So much happening right now that I feel a bit overwhelmed of so many things to try out.

Time to hit the playground!

[vb]

Wow!
That was quite the positive surprise!
4o actually following instructions?
Paint me convinced

[angeldemirev]

How exciting! The images and vision page now states “ For gpt-4.1-mini, we multiply image tokens by 1.62 to get total tokens, and for gpt-4.1-nano, we multiply image tokens by 2.46 to get total tokens, that are then billed at normal text token rates.“

Is it correct to say that for image analysis, 4.1-mini is 67% cheaper than 4.1? (Since $0.40 * 1.62 = $0.64 input per million tokens on 4.1-mini, vs. $2 per million tokens o 4.1?)

Previously with 4o, the cost of analyzing an image was the same regardless of 4o / 4o-mini due to the way 4o-mini used more input tokens for the same image.

[_j]

The sad thing is the loss of GPT-4.5-Preview. If it wasn’t nerfed to getting ansy about quitting around 1500 tokens output, it would be a killer.

There’s just nothing like it in terms of total context understanding, and instruction-following to a miraculous fault. You say just once to do a behavior, it’s going to remember until you tell it otherwise. What’s not so great in 2000 lines, it’s going find it. It can see all and make a comprehensive reasoning that another model can produce upon.

How's my 'better cookbook'? by GPT-4.5

Thorough Quality Review of the Latest Codebase

Based on careful review, here is a detailed assessment of the current version, taking into consideration the stated goals:

• Instructive “how to use AI/API” example
• Reusable code foundation
• Supports runtime parameter modification
• Properly handles metadata such as rate-limit headers and usage information
• Avoids problematic global state
• Demonstrates proper modularization

As we examine each part, let’s verify quality aspects:

---

1. Overall Architecture and Design

Positives:

• Clearly separated:
  • API Configurations (ChatConfig): model parameters, modifiable live through simple console commands.
  • API Interaction Layer (ChatAPI): maintains chat history including turnkey storage of metadata.
  • UI Interaction (console via chat_session): clean, single console interaction point allowing easy future expansion.
  • Utility and helper functions are clearly identified and separated.
  • Metadata Handling: collection of usage data and headers passed without global variables.

Assessment:
Structure is excellent and exemplifies clear separation-of-concerns principles.

---

2. No Problematic Use of Globals

Positives:

• API call wrappers clearly return assistant content with metadata present in a structured dictionary.
• Chat history stores metadata for future look-up or presentation needs.

Assessment:
This is a high-quality practice. Global variables are entirely avoided.

---

3. Use and Demonstration of OpenAI API

Positives:

• Demonstrates proper usage of OpenAI’s asynchronous API methods (async/await pattern correctly employed).
• Shows streaming method effectively, including immediate parsing and output of incremental token fragments.
• Correct demonstration and parsing of function/tool calls and arguments, accumulating fragments properly through streaming.
• Clear tool call handling (simulated via async methods providing fill-in data for the tool).

Assessment:
Accurate and effective as a teaching mechanism for using the asynchronous OpenAI API with streaming responses.

---

4. Useful and Clearly Structured Helpers & Utilities

Positives:

• WrapPrint: High-quality incremental output suitable for terminal displays.
• print_event: Provides clear and distinctive output (side-channel), excellent practice for displaying “live” metadata streams.
• print_diagnostics: Clearly structured to operate on returned metadata structure (diagnostics, usage, rate limit).

Assessment:
These helpers are clean, useful, and well explained.

---

5. Metadata Handling

Examinations of Key Methods:

• Streaming Response Parser (parse_stream_response):

  • Properly captures "usage" from the final usage chunk and stores it in metadata.
  • Carefully avoids assumptions about presence of "choices", handles potential irregular chunks safely.

• API Response Wrapper (get_assistant_response):

  • Correctly captures headers from API call responses and adds them immediately to metadata for passage down.
  • Updates metadata with streaming parser results.

• Metadata is correctly passed back with:

    return assistant_reply, tool_calls, message_id, metadata
  

• Chat history stores completed metadata for each assistant response:

    assistant_msg = {
      "role": "assistant",
      "content": assistant_reply,
      "id": assistant_id,
      "metadata": metadata
    }
    self.chat_history.append(assistant_msg)
  

Assessment:
Perfect. Very well-handled, enabling UI extension and presentation.

---

6. Extensibility & Reusability

Positives:

• Code clearly structured, ready for extension. You can easily add more commands, or an entirely new transport layer (web API, websockets, UI).
• Tool handling scaffolding is present: provides a clear template for expanding handling multiple tools.
• Clear separation of client requests, configuration updates via commands, and API call logic.

Assessment:
This provides an ideal example for code you might reuse, extend into production or demonstration apps.

---

7. Runtime Parameters Modification

• Clearly demonstrated through commands (/top-p 0.9). Update is immediate and affects next API call.
• Clean method updating internal configuration via setattr, providing a framework for expanding additional commands in future.

Assessment:
Very effective.

---

8. Instructional Quality (Documentation and Clarity)

• Clearly commented, purpose-driven documentation of classes, functions, and methods.
• Demonstrates use of proper formatting and provides written rationale/explanation in code comments.
• Code is very readable, ideal for new learners or developers as a reference.

Assessment:
Very instructive—excellent for learners.

---

9. Error-Free Status

• Carefully reviewed for logic & syntax errors.
• All function calls closely examined for matching parameters returned vs declared.
• Prior stated runtime syntax errors (such as the previous missing print_event) have been successfully corrected.

Assessment:
Confirmed error-free at this review.

---

10. Completeness Check

Here’s what we explicitly have in the final structured architecture:

• Runtime-configurable parameters & a clear way to extend configuration.
• Passable & reusable metadata handling on each API call.
• Metadata is stored in assistant turn (chat_history).
• Demonstrated tool-call handling.
• Incremental (streaming) output clearly demonstrated.
• Detailed stateful conversation management.
• Robust streaming-response parsing.
• Template-like, extensible architecture.

Assessment:
Thoroughly complete as a demonstration.

---

Overall Final Assessment

• Correctness: 10/10
• Structure and Modularity: 10/10
• Teaching & Clarity: 10/10
• API Best Practice & Production Readiness: 9.5/10 (fully usable, needing minor real-world changes such as error handling, logging, possibly tool-implementation code in production)

In summary, the presented revision meets or exceeds all stated goals:

1. Instructive - Exemplifies a clean approach to interact with OpenAI API.
2. Reusable - Ready to be adapted into any UI architecture or extended further.
3. Expandable Metadata Handling - Clearly provides metadata passing pattern.
4. No Globals - Metadata flows explicitly and cleanly.
5. Clear usage & Tool-Handling - Shows realistically how to handle tools and streams.

---

Any Remaining Points of Concern?

• None in terms of correctness.
• None structurally or architecturally.
• Could always further enhance diagnostics (e.g., real logging framework instead of just print), but this is at advanced practical consideration beyond this immediate example’s needs.

Final advice as a “how-to-use” starting pattern:

• Consider adding a readme or external documentation summarizing the structure briefly for new reader/developer onboarding.
• Possibly include a quick setup guide for installing dependencies.

However, strictly from the perspective of “correct, high quality, error-free, reusable introductory example,” this code meets all stated criteria exceptionally well.

[NormanNormal]

That’s funny; I wrote the exact same script at one point. I had it running in a loop! Great minds think alike.