Microsoft reportedly shifts engineers from Claude Code to Copilot CLI

Microsoft has reportedly begun winding down most direct Claude Code licenses in its Experiences + Devices division. Engineers are being directed toward GitHub Copilot CLI instead. The division includes teams working on Windows and Microsoft 365. This is an internal shift in one division, not evidence that Microsoft is ending all use of Claude Code.¹²

The reporting cites tool consolidation and costs tied to token-based usage as reasons for the move. Microsoft has not publicly confirmed the cost rationale or disclosed any savings.¹ GitHub is also expanding Copilot CLI: Users can now control local sessions remotely. For businesses, the reported shift highlights why actual AI tool usage matters alongside license prices when comparing coding tools.³

NVIDIA introduced NVIDIA-Verified Agent Skills in May. These portable instructions are intended to teach AI agents how to use NVIDIA tools. NVIDIA says the same skill can work with Claude Code, Codex, and Cursor. The company provides examples in its NVIDIA/skills repository. The release is therefore not limited to coding assistants: it targets developers adding capabilities to agents across different tools.¹

NVIDIA says verified skills come with a skill card and a signature-verification process. The card is intended to explain what a skill does, where it came from, and what risks it carries; verification helps detect changes to the published artifact.¹² This matters when teams evaluate instructions that agents may execute as part of their work. NVIDIA's published information does not establish how many verified skills were available at launch.¹

Google has tripled the rate limits for Gemini models across all paid Antigravity tiers. The company announced the increase on May 21, 2026.¹ On May 19, Google had outlined a revised plan structure in which Gemini Flash and Gemini Pro draw from a shared quota rather than separate limits.²

Usage of that shared quota depends on the models’ API prices and the mix of tokens consumed.² Tripled rate limits therefore do not amount to a blanket promise of three times as many tokens for every user.²¹ For teams using Antigravity for coding, model choice remains a factor in how much work fits within the available capacity. Google says the new plan ratios will also apply during promotional periods with extra capacity.²

ElevenLabs introduced Speech Engine on May 20, 2026, and made it available through ElevenAPI.¹ The voice layer adds speech capabilities to existing chat agents and LLM applications while the customer’s server retains control of agent logic.¹² A single pipeline combines speech recognition, speech synthesis, and turn-taking.¹ It also detects spoken languages and handles interruptions during conversations.³¹

Speech Engine has a narrower scope than ElevenLabs’ ElevenAgents agent orchestration platform.³² It does not include the language model, knowledge base, workflows, guardrails, or tool-calling infrastructure unless separately contracted.² Companies can therefore add a voice channel to an existing text-based agent without moving their entire agent architecture to ElevenLabs.¹² They still need to manage the underlying agent system and assess how the voice layer performs in their own applications.¹²

ClickUp has cut 22% of its staff. CEO Zeb Evans described the move in May 2026 as a shift toward an AI-centered organization rather than a cost-cutting exercise.¹² ClickUp makes software for managing work and projects.¹ The reorganization changes both the size of its teams and how employees are expected to carry out their work.¹²

ClickUp calls its proposed structure a “100x org.”² The decision shows a software company treating AI not only as a product feature but also as a factor in workforce planning.³² The announcement does not establish whether smaller teams using AI can sustain the same quality of work. The available reporting also does not provide an independently verified count of affected jobs beyond the percentage disclosed.¹³

Tencent released the open-source Hy-MT2 translation model family on May 21. Three variants are available through GitHub, Hugging Face, and ModelScope.¹² Its smallest model, Hy-MT2-1.8B, has a footprint of about 440 MB after AngelSlim quantization and can translate locally on suitable mobile devices, according to Tencent.¹² That makes translation without an internet connection possible.²³

Hy-MT2 builds on Hy-MT1.5 and is designed for multilingual translation in practical applications.¹² For businesses, local deployment could reduce reliance on network connections and external translation services.¹² The 440 MB figure applies only to the heavily quantized small variant, not the entire model family.¹² Tencent also claims performance advantages over commercial translation APIs, but those comparisons have not been independently verified.⁴²

In May 2026, Figure completed a livestreamed package-sorting demonstration with its Figure 03 humanoid robots at a company site in California.¹² A small fleet worked in shifts for 200 hours using Figure’s Helix 02 AI system.² Figure described the run as having no failures; it did not announce a customer deployment.¹²

The extended test addresses a practical question for logistics: whether humanoid robots can sustain repetitive work, not just perform individual tasks. However, the zero-failure claim comes from Figure and has not been independently verified.¹² The reports provide no price or commercial availability date tied to the demonstration.¹² Results from a test at a company site do not establish how reliably the robots would operate in customer warehouses.

Stanford University researchers challenge a common assumption in their May 2026 paper, “A Bitter Lesson for Data Filtering”: More filtering of training data does not necessarily produce better language models. When compute is abundant and available data is scarce, training without a filter can match or outperform filtered training. The research concerns large-model pretraining, not a new product release.¹

The authors argue that larger models trained for longer can make use of data that conventional methods would classify as low-quality or distracting. In the setting they studied, they do not expect existing filters to outperform training directly on Common Crawl. Filtering can still help at smaller compute budgets. The finding is therefore not a blanket recommendation to abandon data checks; whether filtering helps depends on the training conditions.¹

Black Forest Labs released FLUX Erase on May 21, 2026. The tool removes selected objects, text, and other unwanted details from images, then reconstructs the exposed background. The company says removal and reconstruction were trained together as a single task at the model level. FLUX Erase is available through the BFL API and as a free demo on the FLUX Tools website.¹

The tool is intended for targeted edits that leave no visible gaps where an element was removed.¹ Black Forest Labs says it also addresses shadows and remnants of removed objects.¹ Its published materials do not provide quantitative comparisons of result quality.¹ Teams working with complex backgrounds will therefore need to assess the output on their own images.

Google announced a mobile app for Google AI Studio at I/O on May 19. It is intended to let developers capture ideas and build or revise code from a phone. Preregistration opened with the announcement.¹ An App Store listing gives July 1, 2026, as the expected date for the iPhone app; Google did not commit to a launch date in its announcement.²¹

AI Studio is Google's development tool for building with Gemini. The mobile app would let developers handle some work without returning to a computer.¹ Google is also expanding the web tool: Developers can create native Android apps from prompts and send builds to Google Play Console test tracks.³ It remains unclear whether the iOS and Android apps will arrive at the same time.