AI Flash Report

NEW AI MODEL RELEASES - May 28, 2026

There are no major frontier‑lab model releases in the last 24–48 hours from OpenAI, Anthropic, Google, Meta, Mistral, xAI, Cohere, Stability, Microsoft, or Hugging Face based on public trackers and news as of early UTC. In that case, here are the most recent notable model launches from roughly the past month, which shape the current landscape.

---

•GPT‑5.5 Instant by OpenAI — A lightweight, lower‑latency variant of GPT‑5.5 aimed at cheaper, high‑throughput workloads, listed as a recent proprietary release on model‑tracking sites.[4]
- Capabilities: General‑purpose chat and coding with GPT‑5.5‑class behavior, optimized for speed and cost (positioned similarly to prior “Turbo” lines, but at the 5.5 generation).[4]
- Key specs: Exact parameter count and context window not yet publicly detailed; expected to share the large‑context regime of GPT‑5.5 while trading some peak capability for latency and price (inferred from prior OpenAI “Instant/Turbo” lines).[4]
- Availability: API via OpenAI; proprietary, no open weights.[4]

•Mistral Medium 3.5 by Mistral AI — An open‑source mid‑sized model that refreshes Mistral’s “Medium” tier, appearing as a new open‑weight release.[4]
- Capabilities: Strong general‑purpose reasoning and coding; positioned between smaller “7B‑class” models and large frontier models, with a focus on efficiency and self‑hosting.[4]
- Key specs: Exact parameter count not listed on the tracker; typical Mistral “Medium” configurations are in the tens of billions of parameters with multi‑hundred‑thousand token context (specs partly inferred from earlier Mistral families).
- Availability: Open source / open weights, suitable for on‑prem and cloud deployment.[4]

•Grok 4.3 by xAI — The latest proprietary Grok iteration, marked as a new release about three weeks ago.[4]
- Capabilities: General chat, reasoning, and coding, following Grok 4.20 and 4.2 as xAI’s flagship line; targeted at X/Twitter integration and agentic use cases.[1][4][5]
- Key specs: Specific parameter counts and context are not disclosed; Grok 4.x is positioned as a frontier‑class multimodal LLM.[1][5]
- Availability: Closed / API access via xAI and X; no open weights.[4][5]

•GPT‑5.5 by OpenAI — Still the newest major frontier model from OpenAI, launched a little over a month ago and underpinning many current products.[1][2][4]
- Capabilities: OpenAI calls it its “smartest and most intuitive” model, with significant gains in agentic coding, software control, and knowledge work compared with GPT‑5.4.[1][2]
- Key specs: Very large‑scale model (parameters undisclosed) with a long context window; GPT‑5.5 Pro adds heavier test‑time compute for high‑precision domains like law and medicine.[1][2]
- Availability: API and ChatGPT; proprietary. GPT‑5.5 Pro available for enterprise and specialized workloads.[1][2]

•Gemini 3.1 Ultra by Google — Google’s current flagship Gemini model, released in preview in April and still among the most recent frontier‑level launches.[1]
- Capabilities: Multimodal across text, image, audio, and video, with a focus on “agentic” workflows and deep tool use.[1][3]
- Key specs: Up to 2‑million token context window, one of the largest publicly announced contexts among major labs.[1]
- Availability: Preview / API via Google Cloud and Workspace integrations; closed weights.[1][3]

•Gemma 4 by Google — New open model family announced in April as part of Google’s broader AI update.[3]
- Capabilities: General‑purpose open‑weight LLM, intended for developers who want self‑hosted or fine‑tunable models, complementing proprietary Gemini.[3]
- Key specs: Multiple sizes; optimized for efficient deployment rather than absolute SOTA; detailed parameter counts and context sizes provided in Google’s technical docs.[3]
- Availability: Open weights for research and commercial use under Google’s Gemma license.[3]

•DeepSeek V4 Pro / V4‑Pro‑Max by DeepSeek — A frontier‑class low‑cost model line that has drawn attention for its aggressive pricing and strong evaluations.[2][4]
- Capabilities: General‑purpose reasoning and coding with competitive benchmark scores; CAISI at NIST published a formal evaluation emphasizing its performance and reliability.[2]
- Key specs: V4 Pro and V4‑Pro‑Max variants, with very low hallucination and extremely low per‑token prices (V4 Pro priced around $0.0036 per million tokens, undercutting GPT‑5.5 by ~97%).[2]
- Availability: Open‑weight and API variants; positioned as an alternative to US frontier labs, especially for cost‑sensitive users.[2][4]

---

What This Means:
Model launches in the last week are incremental variants (e.g., GPT‑5.5 Instant, Mistral Medium 3.5, Grok 4.3) rather than entirely new generations, suggesting a phase of optimization (latency, cost, open‑weights) on top of already‑released frontier models like GPT‑5.5 and Gemini 3.1 Ultra.[1][3][4] At the same time, aggressive pricing from players like DeepSeek and continued open‑weight releases from Mistral and Google’s Gemma line indicate a two‑track race: proprietary frontier systems for maximum capability, and rapidly improving open models plus ultra‑cheap APIs competing hard on cost and flexibility.[2][3][4]

---

Sources / URLs

[1] New AI Model Releases News – May 2026 (overview of GPT‑5.5, Gemini 3.1 Ultra, Grok 4.20, Mistral 3, etc.)
[2] “AI by AI Weekly Top 5: April 27 – May 3, 2026” – Champaign Magazine (GPT‑5.5 launch, GPT‑5.5 Pro, Codex 2.0, DeepSeek