GenAI Chronicle
2010
2011
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2025

August

6 events

Editors' summary

GPT-5 reaches everyone, Microsoft gets a model of its own

On 7 August, OpenAI released GPT-5 to every ChatGPT user, free tier included, routing each question between a fast model and one that thinks. Taking away the ability to pick older models drew enough complaint that OpenAI brought 4o back.

On the 5th it published gpt-oss with open weights — the first time it had released weights since GPT-2. Anthropic shipped Claude Opus 4.1 the same day.

On the 28th, Microsoft released MAI-1 and MAI-Voice-1, its first in-house foundation models, read as adding a pillar of its own beside the OpenAI technology it had leaned on.

On the 26th, a Google image model went round the world under the name "nano banana."

This block is written by the editors. It is kept separate from the sourced record below.

Record6 events
  1. 5
    ModelAnthropic

    Claude Opus 4.1 released

    Anthropic released Claude Opus 4.1, scoring 74.5% on SWE-bench Verified across its 500 problems and citing gains in research, data analysis, tracking detail, and agentic search. Customers reported roughly a one-standard-deviation improvement over Opus 4 on junior developer benchmarks. It shipped to paid Claude users and Claude Code, the Anthropic API, Amazon Bedrock, and Google Cloud's Vertex AI at the same price as Opus 4. OpenAI released its open-weight gpt-oss models the same day, making the date emblematic of the release cadence that summer.

  2. 5
    ModelOpenAI

    OpenAI releases open-weight models gpt-oss

    OpenAI released gpt-oss-120b and gpt-oss-20b, open-weight reasoning models, under the Apache 2.0 license. They were the company's first open-weight language models since GPT-2 in 2019. gpt-oss-120b runs on a single 80GB GPU and gpt-oss-20b on edge devices with 16GB of memory, with the model weights downloadable from Hugging Face. Both are Mixture-of-Experts (MoE) models: gpt-oss-120b holds 116.8 billion parameters in total but runs only 5.1 billion of them for any given token, which is what lets it fit on one GPU. The release drew attention as a shift for OpenAI, which had long pursued a closed-model strategy.

  3. 7
    ModelOpenAI / Sam Altman

    GPT-5 launched to all ChatGPT users including free tier

    OpenAI launched its flagship model GPT-5. It was designed as a unified system combining a fast model, a deeper reasoning model (GPT-5 thinking), and a real-time router that chooses between them per query. OpenAI claimed state-of-the-art performance in coding, math, and health, along with reduced hallucinations. GPT-5 rolled out to all ChatGPT users including the free tier, marking the first major generational upgrade in roughly two and a half years since GPT-4.

  4. 8
    ProductOpenAI / Sam Altman

    OpenAI puts GPT-4o back in the picker after a backlash

    When GPT-5 launched, OpenAI removed GPT-4o, GPT-4.1, GPT-4.5, GPT-4.1-mini, o3, o3-pro, o4-mini, and o4-mini-high from ChatGPT's model picker with no notice, having designed GPT-5 to replace them all. Users revolted. The router that assigns queries to models broke for hours on launch day, making GPT-5 look worse than it was. Most of the anger was about personality rather than capability: "not all of us need a pro coding model," ran one widely shared complaint, "some of us rely on 4o for creative collaboration, emotional nuance, roleplay." The next day Sam Altman said the company would "bring it back for plus users, and will watch usage to determine how long to support it"; on the 12th he wrote that "4o is back in the model picker for all paid users by default. If we ever do deprecate it, we will give plenty of notice." A "show additional models" toggle brought back o3, 4.1, and GPT-5 Thinking mini as well. Older models in the API were never affected.

  5. 26
    ModelGoogle

    Google's "nano banana" image model becomes a global craze

    Google released Gemini 2.5 Flash Image, the model known during development by the codename "nano banana." Beyond text-to-image generation, it drew praise for targeted edits to existing images, blending several images into one, and maintaining the same person or character across generations — all through natural-language instructions. The response was unusual: between August 26 and September 9 the Gemini app added more than 23 million new users and people generated over 500 million images, taking Gemini to the top of both the App Store and Google Play. Trends such as 3D figurine-style self-portraits and vintage saree photos filled social feeds, particularly in India. It was a case of an appealing experience rather than raw capability driving adoption, and it briefly moved the center of image generation from OpenAI to Google.

  6. 28
    ModelMicrosoft / Mustafa Suleyman

    Microsoft ships its first in-house foundation models

    Microsoft released two models developed in-house by the Microsoft AI division under Mustafa Suleyman. MAI-Voice-1 was an efficiency-focused speech generation model that ran on a single GPU and was already powering features such as Copilot Daily and podcast-style explanations. MAI-1-preview was a text foundation model aimed at instruction following and consumer Copilot conversations, trained on roughly 15,000 NVIDIA H100 GPUs. Suleyman said both were designed around efficiency and cost-effectiveness. Microsoft framed the release not as an immediate replacement for OpenAI but as the opening move in a long-term plan to orchestrate specialized systems by use case. For a company that had depended on OpenAI's models since 2019, it put Microsoft among those building foundation models of their own.

↑ 2025