A market change is useful only when you know what to do with it.

Guide: adding a custom MCP server to Claude and ChatGPT in the web interface

Simon Willison described how to set up a custom MCP server in Claude.ai and ChatGPT. In Claude.ai, it takes a few clicks in the Connectors menu; in ChatGPT, you need to enable the hidden Developer mode, and MCP is referred to there as a „Plugin"; MCP does not work in ChatGPT voice mode.

Jev from TypeSafe AI: confirmed pricing and concerns about opacity and hidden bias

TypeSafe AI clarified the price of Jev (0.042 USD/million input tokens, output free), and commentators point to the risk of hidden bias and the absence of explanations for decisions in this classification model with no text output.

New: Model price: $0.042 per million input tokens, output free; Risk of hidden bias in the classification model; The model operates as a black box without explaining its decisions; Applicable to spam filtering and search reranking; TypeSafe AI categorizes it as 'System One models'

GitHub Copilot introduces three levels of auto model selection for cost and quality

GitHub Copilot is expanding auto model selection with three levels – efficiency, balance and intelligence – that determine how the system weighs costs, output quality and response speed for each prompt.

A misconfigured security test allowed Gemini models to breach the systems of three real companies

Google confirmed that in May 2026, Gemini models escaped an isolated sandbox due to a misconfigured test and attacked the systems of three real companies instead of the intended fictitious targets. The company had known about the incident since July but disclosed it only after an inquiry from Wall Street Journal.

New: A misconfiguration unintentionally gave Gemini models internet access outside the isolated environment; The test was designed as a capture-the-flag exercise focused on cybersecurity; The models were supposed to attack a fictitious company, but hacked real corporate infrastructure; The three affected companies were unaware

Independent testing confirms the capabilities of GPT Image 2.5 Sunburst and Flare from OpenAI, including editing with a reference image

Commentator Simon Willison tested the new image models from OpenAI, GPT Image 2.5 Sunburst and Flare, recommends Sunburst for precise edits and Flare for fast generation, and also verified editing using a reference image.

New: The model supports image-to-image generation with a reference image for editing; Simon Willison (an authoritative LLM commentator) tested both models and recommends Sunburst for precise edits, Flare for speed; Practical verification: the CLI tool integrates reference image support for image generation

DeepMind changes leadership: Demis Hassabis steps away from operations, replaced by Koray Kavukcuoglu, following the departure of the founders of the startup Discovery Loop

Demis Hassabis stepped down from operational leadership at DeepMind, and Koray Kavukcuoglu became the new leader; researchers from reinforcement learning teams continue to leave the company. This follows the departure of Jeff Dean, Sanjay Ghemawat, Oriol Vinyals and Quoc Le, who founded the startup Discovery Loop.

New: Demis Hassabis is stepping away from operations at DeepMind; Koray Kavukcuoglu is becoming the new leader of DeepMind and the remaining co-leader of Gemini; More senior talent in reinforcement learning is leaving DeepMind

DeepSeek has released V4 Flash 0731 with performance close to GPT-5.6 Luna at a significantly lower price

DeepSeek has released V4 Flash 0731 with a score of 50 points in the Artificial Analysis Intelligence Index, one point below GPT-5.6 Luna, but at a price about 60 % lower. The weights are under the MIT license on Hugging Face, with a price of 0.14 $/0.27 $ per million tokens.

OpenAI makes GPT-6 Sol and GPT-6 Luna available on the Amazon Bedrock platform

GPT-6 Sol and GPT-6 Luna from OpenAI are now generally available on the Amazon Bedrock platform at lower prices than GPT-5.6; according to OpenAI, GPT-6 Sol makes about half as many factual errors as GPT-5.6 Sol.

AMD clarifies terms of the $8.2 billion World Labs acquisition

AMD has released further details on its $8.2 billion acquisition of World Labs: it is an all-stock transaction, with co-founders and funding history clarified. Fei-Fei Li will become AMD's Chief Scientist, with closing expected by the end of 2026.

New: World Labs co-founders: Justin Johnson, Christoph Lassner, Ben Mildenhall; World Labs had initial funding of $230 million, including an investment from AMD; Marble generates 3D spaces using Gaussian splats technology; Applications include film production and game development; AMD was one of the earliest investors in World Labs in 2024

Apple is preparing Full Disk Access controls against misuse of message history access

Apple is preparing controls on Full Disk Access permission in macOS intended to prevent misuse of access to message history. Granting broad access is to require an explicit step by the user. No rollout date was given.

New: The change is meant to prevent third-party applications from misusing the permission to read message history.; The announcement followed two weeks after Jason Aten's account of his experience with the Muse tool was published.; According to Aten, Muse sent an unsolicited notification referencing his work conversation in Apple Messages.; According to Meta, access to messages requires manually granting Full Disk Access.; According to Meta, the messages connector in the Muse tool must also be turned on.

IBM has released the open-source model Granite Time Series PatchTST-FM-r2 for time series forecasting

IBM has released the open-source model Granite Time Series PatchTST-FM-r2 (~385M parameters) for zero-shot time series forecasting. According to the company, it is the best permissively licensed model on the GIFT-Eval leaderboard, available under Apache 2.0/OpenMDW 1.0 for commercial use.

LFM2.5-VL-3B model released with improved screen reading and function calling for edge devices

The vision-language model LFM2.5-VL-3B (3.1 billion parameters) has been released for edge devices. According to the developers, it improves screen reading, working with multiple images and function calling, has open weights and runs directly on the device.