@petergostev - AI Dynamics

Opus 4.7 Context Window: 400k vs 1m Performance Analysis

By

–

29 April 2026 0h02

Opus 4.7 – 400k vs 1m context – is there a difference?

I've heard @theo talk about the fact that it is unlikely that Anthropic would have offered up a model with 1m context at the same cost, if it wasn't a different (i.e. cheaper to serve) model.

I did a test where I toggled… pic.twitter.com/tQC4Tn17el
— Peter Gostev (@petergostev) 28 avril 2026

Opus 4.7 – 400k vs 1m context – is there a difference? I've heard @theo talk about the fact that it is unlikely that Anthropic would have offered up a model with 1m context at the same cost, if it wasn't a different (i.e. cheaper to serve) model. I did a test where I toggled

→ View original post on X — @petergostev

29 April 2026

Running OpenAI 4o and o4 models simultaneously

By

@petergostev

–

28 April 2026 18h18

I never complain! The peak was when we had 4o and o4 at the same time

→ View original post on X — @petergostev

28 April 2026

Codex-Max Controversy: What Happened and Why It Matters

By

@petergostev

–

28 April 2026 16h23

Are we going to pretend like the while 'Codex-Max' thing never happened?

→ View original post on X — @petergostev

28 April 2026

Improving Image Generation Parallelization and Resolution in Codex

By

@petergostev

–

28 April 2026 9h50

I'd appreciate if you could improve the way images are done in Codex – there seems to be 1) no parallelisation, 2) don't think I can go to 4k, 3) there is something funny sometimes with how Codex sees the generated images, whether they are getting properly passed into the

→ View original post on X — @petergostev

28 April 2026

GPT-5.5 Coding Capabilities Better Than Rankings Suggest

By

@petergostev

–

27 April 2026 23h58

Yeah totally, I agree gpt-5.5 is way better at coding that this ranking suggests, we'll do better & we'll have a better way to measure more broad coding ability too. Appreciate the feedback

→ View original post on X — @petergostev

27 April 2026

Bullshit Benchmark: Evaluating AI Model Reliability

By

@petergostev

–

27 April 2026 20h19

GitHub: https://
github.com/petergpt/bulls
hit-benchmark
… DataViewer: https://
petergpt.github.io/bullshit-bench
mark/viewer/index.v2.html
…

→ View original post on X — @petergostev

27 April 2026

DeepSeek v4 Underperforms on BullshitBench Reasoning Tasks

By

@petergostev

–

27 April 2026 20h19

BullshitBench: sorry to say but DeepSeek v4 did really badly, towards the bottom of the table, whether it is high or low reasoning.

→ View original post on X — @petergostev

27 April 2026

GPT-5.5 Reinforcement Learning Scaling Across Model Sizes

By

@petergostev

–

26 April 2026 15h51

GPT-5.5 by Reasoning Effort: I've asked it in Codex to create a physics-based visualisation of RL cycles for different sized models (70b, 1t, 10t), to demonstrate how the amount of RL you can do differs by model size.

My assessment of each:
– Low: weird slop
– Medium: kinda… pic.twitter.com/6YCNqPyzcR
— Peter Gostev (@petergostev) 26 avril 2026

GPT-5.5 by Reasoning Effort: I've asked it in Codex to create a physics-based visualisation of RL cycles for different sized models (70b, 1t, 10t), to demonstrate how the amount of RL you can do differs by model size. My assessment of each: – Low: weird slop – Medium: kinda

→ View original post on X — @petergostev

26 April 2026

GPT-5.5 Creates 360° Immersive Babylon Gardens World

By

@petergostev

–

25 April 2026 18h35

Creating an immersive Hanging Gardens of Babylon world with 360° GPT-Image-2 & Codex in 1500 images.

I've tasked GPT-5.5 in Codex to construct a whole world that you can walk through 'google street view' style. It took 1,500 of 2:1 images that can be turned into a 360° immersive… pic.twitter.com/ML9d6EgXTz
— Peter Gostev (@petergostev) 25 avril 2026

Creating an immersive Hanging Gardens of Babylon world with 360° GPT-Image-2 & Codex in 1500 images. I've tasked GPT-5.5 in Codex to construct a whole world that you can walk through 'google street view' style. It took 1,500 of 2:1 images that can be turned into a 360° immersive

→ View original post on X — @petergostev

25 April 2026

GPT-5.5 and Pro Models Underperform on BullshitBench

By

@petergostev

–

24 April 2026 22h23

BullshitBench: GPT-5.5 and 5.5-Pro update! They did NOT do well – 5.5 about the same level as GPT-5.4 (around 30-35 rank, 45% pushback). GPT-5.5-Pro did WORSE – only about 35% pushback. I must say the Pro result kind of shocked me. This is actually interesting, what this tells

→ View original post on X — @petergostev

24 April 2026