Grok 3 and Grok 3 mini API released; Epoch flags o1 Apr 2025

X.ai released the Grok 3 and Grok 3 mini API on April 11, 2025, and Epoch AI independently confirmed Grok 3 as an o1 level model in a now deleted tweet. Sama drummed up some hype for today's Memory update in ChatGPT, but with very little technical detail, and there is evidence that o3 and o4-mini are coming soon, alongside press leaks of 4o's upgrade to GPT4.1.

Grok 3's public launch and immediate reactions

The Grok 3 API rollout included both grok-3 and grok-3-mini endpoints. Reporting noted pricing details were published for grok-3 and grok-3-mini, and grok-3-mini exposes two modes, described as low reasoning effort and high reasoning effort. Epoch AI's independent confirmation that Grok 3 is an o1 level model appeared in a tweet that has since been deleted, creating a short-lived validation signal for the release. The primary source also notes Grok 3 was last covered in February, placing this launch in an ongoing product timeline.

What else moved in the community during April 9–10, 2025

Community monitoring covered 7 subreddits, 433 Twitters and 30 Discords (230 channels, and 6,924 messages), producing an estimated reading-time savings of 601 minutes at 200 words per minute. On Twitter, discussion ranged from hardware (TPUv7 and Google Ironwood comparisons) to model releases. Meta's Llama 4 release generated mixed reactions, with some users expressing excitement and others reporting underwhelming performance, especially in coding. Anthropic released Claude 3.5 Opus, and ByteDance announced Seed-Thinking-v1.5 with 20B activated and 200B total parameters.

Open research and smaller teams also made noise. UC Berkeley open-sourced a 14B model that, after applying reinforcement learning to DeepSeek-R1-Distilled-Qwen-14B, was reported to rival OpenAI o3-mini and o1 on coding tasks; that finetuning reportedly ran on 32 H100 for 2.5 weeks at a cost of approximately $26,880. Sakana AI highlighted a gold medal win at the AI Mathematical Olympiad using SFT and RL on DeepSeek variants. Moonshot AI, Alibaba, and others released multimodal and generative models on Hugging Face during the same window.

The Discord snapshot showed high activity in developer and research communities. Notable servers included LMArena, Unsloth AI, OpenRouter, aider, Manus.im, Perplexity AI, and LM Studio, among many others. Several channels posted hundreds of messages over the two-day span, reflecting active troubleshooting, announcements, and product discussion across infrastructure, models, and tooling topics.

Why it matters

A single API release and a brief confirmation tweet can shift community attention quickly, even when technical detail is sparse. The Grok 3 launch and the deleted Epoch confirmation show how vendor releases and ephemeral endorsements shape early perceptions. At the same time, open-source finetuning work that reports concrete compute and cost figures highlights an alternative path: smaller teams can use RL to close gaps with larger commercial models on specific tasks. Monitoring across subreddits, Twitter, and dozens of Discord channels continues to reveal where engineers and researchers direct effort and skepticism.

What to watch

Look for clearer technical detail on ChatGPT's Memory update and for formal announcements or artifacts confirming o3 and o4-mini availability, and for any follow-up from Epoch AI about its deleted tweet. Community benchmarks that compare Grok 3 and Grok 3-mini modes, and independent replication of the UC Berkeley RL finetuning results, would be the next concrete signals to confirm how these releases perform in practice.