Dateline: May 22, 2026 | Next update: May 29, 2026
Another dense week anchored by three headline stories: Andrej Karpathy (OpenAI co-founder, Tesla AI director) joined Anthropic's pre-training team on May 19; self-hosted sandboxes and MCP tunnels launched for Claude Managed Agents at the Code with Claude London event on the same day; and the Gates Foundation — already announced May 14 — became fully detailed as Anthropic's largest mission-aligned partnership ($200M over four years). PwC expanded its Anthropic alliance to certify 30,000 professionals on Claude, and the API gained cache diagnostics in public beta.
Claude / Anthropic
Andrej Karpathy joins Anthropic — pre-training team
Andrej Karpathy — OpenAI founding team member, former Tesla senior director of AI, and the most widely followed AI educator in the world — joined Anthropic on May 19. He will work under pre-training team lead Nick Joseph with a specific mandate: start a new team that uses Claude itself to accelerate pre-training research. This is a direct bet that AI-assisted experimentation — running more ablations, summarising internal results, writing infrastructure code, and proposing training recipes — is the next competitive lever in the frontier model race.
Karpathy starts at Anthropic on May 19. He will lead a new group within pre-training focused on using Claude to accelerate the training research loop. He paused his education startup Eureka Labs to join. He had been publicly enthusiastic about Claude Code since January 2026. His hire comes as Anthropic's annualised revenue run rate reaches $30B and Claude Code alone generates approximately $2.5B in annualised revenue.
Team: pre-training | Lead: Nick Joseph | Mission: AI-accelerated pre-training research | Background: OpenAI founding (2015), Tesla Autopilot (2017–2022), OpenAI (2023), Eureka Labs (2024–2026) | Education work: plans to resume in time
Best for: Watch this space — Karpathy's focus on AI-accelerated pre-training research is a long-term signal, not an immediate product change
Claude Managed Agents — self-hosted sandboxes + MCP tunnels
Announced at Anthropic's first international developer conference — Code with Claude London — self-hosted sandboxes and MCP tunnels together solve the core blocker for enterprise agent production deployments: data and code leaving the customer's security perimeter. The agent orchestration loop stays on Anthropic's infrastructure; tool execution and private service access move inside the customer's perimeter.
Self-hosted sandboxes (public beta): tool execution now runs on your own infrastructure or through managed providers — Cloudflare, Daytona, Modal, or Vercel. Your network policies, audit logging, and security tooling apply by default. Resource sizing and runtime images are fully customer-controlled. MCP tunnels (research preview): a lightweight gateway you deploy makes a single outbound encrypted connection — no inbound firewall rules, no public endpoints. Agents reach internal databases, private APIs, ticketing systems, and knowledge bases as MCP tools without those services being exposed to the public internet. Also shipping: MCP server and tool configurations can now be updated on an active session (no restart required); large tool outputs over 100K tokens are automatically spilled to a sandbox file, with the model receiving a truncated preview and the path to read the full content.
Self-hosted sandboxes: public beta | Supported providers: Cloudflare (microVMs), Daytona, Modal, Vercel | MCP tunnels: research preview, request access via Claude Console workspace settings | Orchestration stays on Anthropic infra | Live MCP config updates: yes, no session restart required | Large output spill: >100K tokens auto-spilled to sandbox file | Both under managed-agents-2026-04-01 header
Best for: Enterprises in regulated industries (finance, healthcare, legal, government) requiring agent execution inside their own security perimeter
Gates Foundation — $200M four-year partnership
Anthropic committed $200 million over four years — in grant funding, Claude usage credits, and embedded engineering support — to a partnership with the Bill and Melinda Gates Foundation. It is the Gates Foundation's largest AI commitment to date (4x the $50M it gave OpenAI in January for a narrower African clinic programme). The work is led by Anthropic's Beneficial Deployments team.
Healthcare focus: accelerating vaccine and drug candidate screening for neglected diseases (including polio, HPV, preeclampsia), and improving malaria and tuberculosis treatment deployment forecasts with the Institute for Disease Modeling. The partnership will serve the approximately 4.6 billion people in low- and middle-income countries who lack access to essential health services. Education focus: tools to improve K–12 outcomes in the US, sub-Saharan Africa, and India, including African-language data collection and labelling released as public goods. Economic mobility: support for solopreneur and small-business programmes. Public goods deliverables include open health datasets, evaluation benchmarks, and knowledge graphs.
Commitment: $200M over 4 years | Components: grant funding + Claude usage credits + Anthropic engineering support | Led by: Anthropic Beneficial Deployments team | Partner: Bill and Melinda Gates Foundation | Public goods: open datasets, benchmarks, knowledge graphs released to the broader research community | Comparison: Gates Foundation gave OpenAI $50M in Jan 2026 for narrower African clinic work
Best for: Global health researchers, educators, nonprofits, and policymakers — informational for commercial users
PwC — expanded strategic alliance
Anthropic and PwC disclosed a major expansion of their strategic alliance. The rollout starts with PwC's US teams and extends to hundreds of thousands of professionals globally, anchored by a joint Center of Excellence and a certification programme targeting 30,000 PwC professionals trained and certified on Claude.
Claude Code and Claude Cowork rolling out across PwC's global workforce. Joint Anthropic–PwC Center of Excellence established. 30,000 PwC professionals to be trained and certified on Claude. Deployment begins with US teams, expanding globally. This makes PwC one of the largest single enterprise deployments of Claude to date.
Products: Claude Code + Claude Cowork | Starting geography: US, expanding globally | Center of Excellence: joint Anthropic–PwC | Certification target: 30,000 professionals
Best for: Enterprise buyers evaluating large-scale professional services deployment; consulting and advisory firms
API — cache diagnostics in public beta
Prompt caching is one of the most impactful cost-reduction tools on the Claude API, but until this week there was no way to know why a cache miss happened. Cache diagnostics, now in public beta, tells you exactly where the prompt cache prefix diverged from the previous turn — making it possible to debug and fix caching issues without guesswork.
Pass diagnostics.previous_message_id on a Messages request and the API returns a cache_miss_reason field explaining where the cache prefix diverged. Enables systematic debugging of prompt cache misses across multi-turn agentic workflows.
Beta header: cache-diagnosis-2026-04-07 | Parameter: diagnostics.previous_message_id | Response field: cache_miss_reason | Works on: Messages API | Use case: debug prompt cache misses in agentic and multi-turn workflows
Best for: API developers optimising cost and latency through prompt caching
Claude Code — Agent Teams, plugin fixes, session tools
Claude Code shipped a broad point release focused on Agent Teams stability, plugin improvements, and session tooling polish. Agent Teams — collaborative multi-agent sessions where named teammates work together on a task — received several reliability fixes after non-ASCII names caused API failures and crashed sessions.
Agent Teams fixes: teammates with non-ASCII names no longer fail every API call due to invalid header encoding. Plugin marketplace: add/update now respects CLAUDE_CODE_PLUGIN_PREFER_HTTPS; /plugin returns to the Installed list after enable/disable/uninstall actions. /doctor now shows an exec-form example when a command hook is missing the command field. Skill-listing truncation moved out of startup notifications — run /doctor for the full breakdown. Pre-response stream stall recovery improved — now retries streaming once rather than falling back to the slower non-streaming path. SDK/headless MCP startup is up to 2 seconds faster with slow MCP servers (pre-wait now overlaps startup instead of blocking). Fixed: infinite loop where a skill using context: fork repeatedly re-invoked itself. Fixed: /review using a deprecated projectCards GraphQL query that errored on repos with Classic Projects.
Agent Teams: non-ASCII name header encoding fixed | CLAUDE_CODE_PLUGIN_PREFER_HTTPS: now respected on marketplace add/update | MCP headless startup: up to 2s faster | Stream stall: retry-streaming-once before fallback | Fixed: context: fork infinite loop | Fixed: /review GraphQL projectCards deprecation | Fixed: stale 'Failed to install Anthropic marketplace' banner | Fixed: PR badge not updating after gh pr create in-session
Best for: Developers using Agent Teams, MCP servers in headless SDK mode, plugin marketplace
Usage limits — Pro and Max tightening reported
Anthropic quietly tightened Claude usage limits for Pro and Max subscribers. The API was not affected. This is a consumer-plan change only and came shortly after the doubling of Claude Code five-hour limits via the SpaceX compute deal announced May 6. Anthropic has not published a detailed breakdown of the new limits.
Pro and Max plan usage limits tightened as of approximately May 14. API pricing and limits unchanged. The net effect on Claude Code five-hour limits relative to the May 6 doubling is unclear — monitor your usage dashboard if you are a heavy Pro or Max user.
Affected: Pro ($20/mo) and Max ($100–200/mo) consumer plans | API: unaffected | Source: Axios, May 14, 2026 | No official Anthropic statement on specifics | Practical impact: most visible on long Claude Code sessions without context auto-compaction
Best for: Heavy Pro and Max subscribers — check your usage dashboard; no action needed for API users
Plans and Pricing
No API pricing changes this week. The usage limit tightening on Pro and Max plans is the only consumer-facing change and has not been quantified officially. Claude Platform on AWS consumption pricing via AWS Marketplace remains the same as native Claude API rates.
Opus 4.7: $5/$25 per MTok | Sonnet 4.6: $3/$15 per MTok | Haiku 4.5: low-cost tier | Claude Platform on AWS: consumption via AWS Marketplace | Gates Foundation and PwC: enterprise/mission-aligned pricing, contact Anthropic | Self-hosted sandboxes + MCP tunnels: included in Managed Agents (beta)
ChatGPT / OpenAI
Dateline: May 22, 2026 | Next update: May 29, 2026
Over the past week, OpenAI has continued improving agent reliability, multimodal workflows, and enterprise integration, while quietly refining system consistency across professional use cases.
GPT-5.3 Standard — default model
More reliable handling of long and mixed-format conversations.
Improved consistency when switching between text, document, and image-based tasks.
Context ~128k | Output ~4k–8k | Improved multimodal context management
Best for: General use
GPT-5.3 Pro — high-reasoning model
More stable performance on layered analytical tasks.
Improved reasoning continuity in very long outputs.
Context ~200k (est.) | Reduced degradation across extended reasoning chains
Best for: Deep analysis
GPT-5.3 Mini — fallback model
Faster and smoother lightweight responses.
Improved routing precision between Mini and Standard models.
Context ~64k | Better dynamic inference allocation
Best for: Quick tasks
Agent Mode
Handles multi-step workflows.
Improved reliability in executing structured workflows over longer sessions.
Better task persistence | Reduced workflow interruption rates
Best for: Task delegation
Deep Research
Combines browsing and reasoning.
Improved handling of conflicting information across multiple sources.
Enhanced synthesis and source-weighting pipeline
Best for: Research
Memory & Projects
Persistent context across chats.
Better prioritisation of active-project context over older conversational memory.
Improved relevance filtering | More efficient memory retrieval
Best for: Ongoing workflows
Advanced Voice Mode
More natural transitions between conversational turns.
Improved turn-taking latency | Better conversational pacing
Best for: Voice interaction
ChatGPT for Clinicians
Continued expansion of healthcare-oriented workflows.
Improved structure and readability in generated clinical summaries and administrative outputs.
Further domain-specific tuning | Reinforced medical safety guardrails
Best for: Clinical support (non-diagnostic assistance)
Enterprise & Workflow Integrations
Improved reliability across integrations involving documents, collaborative workflows, and persistent projects. Better orchestration across connected tools and memory systems.
Best for: Organisational deployment
Plans and Pricing
No significant changes this week. Pricing stable | API structure unchanged.
Gemini (Google)
Date: May 22, 2026 | Next update: May 29, 2026
This week marked the kickoff of Google I/O 2026, officially shifting the ecosystem into the "Agentic Gemini Era." Google introduced a brand-new model generation, multimodal video foundational architectures, and autonomous cloud agents.
Gemini 3.5 Flash — new default model
Launched as the default engine across the Gemini app and Google Search. Runs four times faster than other frontier models in output tokens per second and outperforms Gemini 3.1 Pro across key logic and coding benchmarks.
Gemini 3.5 Flash is now the default model across the Gemini app and Google Search. 4x faster output token throughput than competing frontier models. Outperforms Gemini 3.1 Pro on logic and coding benchmarks. Infrastructure successfully managed massive traffic spikes post-I/O keynote with no API downtime.
Best for: High-speed agent workflows, general-purpose everyday use
Gemini Omni — new multimodal family
Introduced as Google's premier cross-modal creative model. The first rollout, Gemini Omni Flash, natively combines text, audio, images, and video inputs to generate and text-edit high-quality video outputs while maintaining perfect scene, character, and physics continuity. All output includes SynthID watermarking.
Gemini Omni Flash: native text + audio + image + video input, high-quality video output with scene/character/physics continuity. SynthID watermarking on all generated content. Available via Flow and YouTube Shorts.
Best for: Native video generation and editing, creative multimodal workflows
Gemini 3.5 Pro — coming next month
Announced to be in final testing, with an official deployment scheduled for next month.
Best for: Watch this space — no action needed yet
Gemini Spark — autonomous cloud agent
An always-on, autonomous agent platform powered by Gemini 3.5 Flash. Running on Google Cloud VMs, Spark can review credit card statements for hidden subscriptions, track school updates from emails, compile notes into Docs, and safely perform multi-step actions across third-party apps like Instacart and OpenTable — requiring final user confirmation for purchases.
Gemini Spark: always-on autonomous background agent running on Google Cloud. Supports multi-step actions across third-party apps with user confirmation required for purchases. Continuous cloud automation without needing to stay in an active session.
Best for: Autonomous background tasks, continuous cloud automation
The 25-year Search upgrade
Google Search rolled out its biggest interface overhaul in over two decades, replacing the standard search box with an expanded AI Search Box. Users can input full natural language queries alongside images, video clips, and entire Chrome tabs simultaneously.
AI Search Box replaces the classic Google search box. Accepts text, images, video clips, and full Chrome tabs as simultaneous inputs. May 21 core ranking algorithm update running concurrently.
Best for: Advanced web search, multimodal research queries
Gemini in Chrome — Android and Desktop
Coming to mobile next month, it introduces auto browse to automate digital chores (pulling event ticket details to book local parking). Desktop users gain Skills in Chrome, which saves complex multi-tab prompts into one-click reusable tools.
Auto browse (mobile, next month): automates digital chores from within Chrome. Skills in Chrome (desktop): saves complex multi-tab prompts as reusable one-click tools.
Best for: Advanced web and coding agents, reusable browser automation
Android XR and Smart Glasses
Google teased an Android XR collaboration with Samsung, Gentle Monster, and Warby Parker, showcasing prototype smart glasses capable of real-world text translation, Gemini voice chatting, and real-time audio translation.
Prototype smart glasses: real-world text translation, Gemini voice chat, real-time audio translation. Collaboration: Samsung, Gentle Monster, Warby Parker.
Best for: Watch this space — prototype stage, no release date announced
Plans and Pricing
AI subscription tier restructuring announced at Google I/O:
| Plan | Price | Key Includes |
|---|---|---|
| AI Ultra | $99/mo | 5x higher usage limits vs. standard $20 Pro plan, 20TB storage, priority Antigravity developer tools |
| AI Ultra | $200/mo | 20x higher usage limits, exclusive Project Genie access (interactive 3D from Street View) |
Antigravity 2.0 replaces legacy developer platforms and evolves into a full agentic ecosystem with a standalone desktop application and CLI. All Gemini CLI users are urged to migrate to Antigravity CLI immediately.
Microsoft Copilot
Dateline: May 22, 2026 | Next update: May 29, 2026
The biggest story of the week is model integration: Microsoft rolled out GPT-5.5 models across Copilot experiences, significantly improving reasoning, summarization, and writing quality.
GPT-5.5 models integrated into Copilot
GPT-5.5 models now power Copilot, offering faster responses, deeper reasoning, and smarter writing assistance across all enterprise workflows.
Models: GPT-5.5 Instant + GPT-5.5 Thinking now power all Copilot tiers. Improvements: reasoning quality, context understanding, summarization accuracy. Enterprise workflows benefiting: email drafting, meeting summaries, document creation, data analysis, presentations.
Models: GPT-5.5 Instant + GPT-5.5 Thinking | Applies to: all Microsoft 365 Copilot tiers | Effective: immediate from May 19, 2026
Best for: All Copilot users — especially enterprise teams needing higher-quality outputs
Outlook email grounding in Copilot Chat
Users can now add emails or text from threads directly into Copilot Chat prompts, enabling context-aware answers without switching between apps.
Implicit grounding in Outlook: insert email sections into Copilot Chat. Summarize, analyze, or extract action items from email content directly in the chat interface.
Best for: Business users managing large volumes of email
PDF opening inside Copilot Chat
PDFs now open directly inside Copilot Chat, allowing summarization, highlighting, and Q&A without switching apps.
PDFs open inline in Copilot — no need to leave the chat. Summarization and extraction supported. Reduces workflow interruptions for document-heavy tasks.
Best for: Researchers, analysts, legal teams, document-heavy workflows
App Launcher "Waffle" returns
The classic "Waffle" app launcher is back, improving navigation across Outlook, Word, Excel, Teams, OneDrive, and PowerPoint.
Restores the older launcher design by popular request. Quick access to all Microsoft 365 apps from a single consistent entry point.
Best for: Longtime Microsoft 365 users, productivity-focused teams
Researcher + Notebooks upgrades
Researcher now provides deeper summaries and smarter document analysis. Copilot Notebooks gained new organisational features for managing research workflows.
Researcher: improved information gathering and analysis. Notebooks: enhanced organisation and summarization. Now competitive with standalone AI research assistants.
Best for: Knowledge workers, students, enterprise research teams
Plans and Pricing
No pricing changes this week. Updates focus on feature expansion and model improvements. GPT-5.5 included in all Copilot tiers at no additional cost. Outlook grounding, PDF opening, Waffle, and Researcher upgrades all included in Microsoft 365 — no action needed.