Dateline: May 4, 2026 | Next update: May 11, 2026
A packed 10-day window: Claude Security launched in public beta for enterprise customers, Claude Code shipped its biggest weekly feature drop of the year (xhigh effort, cloud Routines, /ultrareview, usage tracking, native binaries), Claude Design added creative connectors for Adobe and Blender, the 1M context beta retired for older models, and Anthropic opened a Sydney office. For technical users, the period also brought a major Claude Code CLI overhaul and the Managed Agents memory beta going live.
Claude / Anthropic
Claude Security — public beta for enterprise
Anthropic launched Claude Security in public beta this week — a dedicated vulnerability scanning product powered by Opus 4.7. It finds and explains software flaws, ranks them by severity and confidence, and generates targeted patch instructions that can be opened directly in Claude Code on the web. Partners including CrowdStrike, Microsoft Security, Palo Alto Networks, SentinelOne, TrendAI, and Wiz are embedding Opus 4.7 into their own security products.
Claude Security is in public beta for Enterprise. New in this version vs the earlier research preview: multi-stage validation to reduce false positives, scheduled scan support, directory-level scanning, the ability to dismiss or export findings, and send results to project management tools. Non-verified users have guardrails applied; cybersecurity professionals can join the Cyber Verification Program for fuller access.
Powered by Opus 4.7 | Enterprise public beta | Integration: findings exportable + sendable to PM tools | Cyber Verification Program for professionals | Partners: CrowdStrike, Microsoft Security, Palo Alto Networks, SentinelOne, TrendAI, Wiz
Best for: Security teams, DevSecOps, enterprises running vulnerability scanning at scale
Claude Code — biggest feature drop of the year (Week 16)
Claude Code's Week 16 digest is the largest single update of 2026 so far. Opus 4.7 is now the default model on Max and Team Premium plans. A new xhigh effort level is the recommended setting for most coding work. Cloud-based Routines let you automate templated agents on a schedule or GitHub event. /ultrareview runs parallel multi-agent code review in the cloud. /usage shows exactly what is consuming your limits. And the CLI now ships as native binaries for faster startup.
Opus 4.7 set as default model on Max and Team Premium. New xhigh effort level added — now the recommended setting for most coding work — with an interactive /effort slider. Routines: fire templated cloud agents from a schedule, GitHub event, or API call via Claude Code on the web. /ultrareview: parallel multi-agent cloud code review. /usage: see what is driving your usage limits. CLI moves to native binaries. claude project purge [path] added to wipe all Claude Code state for a project. /model picker now lists models from your gateway's /v1/models endpoint when using a custom base URL.
Default model for Max + Team Premium: claude-opus-4-7 | Effort levels: low / medium / high / xhigh (max removed) | ANTHROPIC_BEDROCK_SERVICE_TIER env var for Bedrock tier selection (default, flex, priority) | /resume now finds sessions by pasting a PR URL (GitHub, GitLab, Bitbucket, GitHub Enterprise) | ENABLE_PROMPT_CACHING_1H for 1-hour cache TTL on Bedrock/Vertex/Foundry | claude project purge: --dry-run, -y, -i, --all flags
Best for: Developers running long agentic coding sessions, teams doing parallel code review, CI/CD pipelines
Claude Code — broad CLI stability + security overhaul
Alongside the Week 16 feature drop, Anthropic shipped a separate broad CLI update covering security hardening, permission improvements, OAuth reliability, and Windows fixes. The update also brings expanded OpenTelemetry logging and a smarter model picker for gateway setups.
Stronger sandbox and permission safeguards. OAuth login overhauled — 401 retry loops fixed, MCP servers with omitted expires_in no longer require hourly re-auth. Windows: CLAUDE_ENV_FILE and SessionStart hook env files now apply; drive-letter path rules correctly root-anchored. /color now syncs session accent color to claude.ai/code when Remote Control is connected. ExitWorktree tool added to leave EnterWorktree sessions. CLAUDE_CODE_DISABLE_CRON env var added to stop scheduled cron jobs mid-session. HTML comments in CLAUDE.md now hidden from Claude when auto-injected.
New env vars: CLAUDE_CODE_DISABLE_CRON, ANTHROPIC_BEDROCK_SERVICE_TIER | ExitWorktree tool | Fixed: OAuth 401 retry loop with CLAUDE_CODE_DISABLE_EXPERIMENTAL_BETAS=1 | Fixed: plugin install not re-resolving dependency at wrong version | Fixed: Remote Control sessions archiving on transient JWT refresh | OpenTelemetry logging expanded
Best for: Developers using Remote Control, MCP servers, Windows, Bedrock/Vertex/Foundry custom gateway setups
Claude Design — creative app connectors added
Claude Design, which launched on April 17, has expanded rapidly. Anthropic added connectors for Adobe (Photoshop and other apps), Blender, Ableton, Affinity, and Autodesk Fusion this week — putting Claude inside the file formats and tools that creative and product teams already use. This positions Claude Design beyond one-pager generation and into active design and 3D production pipelines.
New Claude Design connectors: Adobe apps (including Photoshop), Blender, Ableton, Affinity, Autodesk Fusion. Claude can now read files directly from these tools during design sessions. The web capture tool (grab elements from your live site) and multi-design-system support were also confirmed as generally available within Design this week.
Connectors: Adobe, Blender, Ableton, Affinity, Autodesk Fusion | Export: PDF, URL, PPTX, Canva | Multi-design-system support per team | Enterprise: off by default, admin-enabled | More integrations signalled as coming weeks
Best for: Designers, 3D artists, creative agencies, product and marketing teams
Claude Managed Agents — memory now in public beta
Memory for Claude Managed Agents moved to public beta this week. Agents can now persist state across sessions without developers building their own memory layer — a significant reduction in harness complexity for long-running agent workflows.
Managed Agents memory is in public beta. Agents retain context and state across sessions natively, removing the need for custom memory infrastructure. Use the existing managed-agents-2026-04-01 beta header — no new header required. Full integration guide available in Using agent memory in Claude Docs.
Beta header: managed-agents-2026-04-01 (unchanged) | Durable state across sessions | See: Using agent memory in Claude Docs | Complements secure sandboxing, built-in tools, and server-sent event streaming already in Managed Agents
Best for: Developers building long-horizon agents that need to remember context between sessions
API Platform — 1M context beta retired for older models
The 1M token context window beta has now been retired for Claude Sonnet 4.5 and Claude Sonnet 4. The context-1m-2025-08-07 beta header has no effect on these models as of April 30. Requests exceeding the standard 200k context window now return an error. Developers must migrate to Sonnet 4.6 or Opus 4.6, where 1M context is generally available at standard pricing with no beta header.
1M context beta (context-1m-2025-08-07) retired April 30 for Sonnet 4.5 and Sonnet 4. Requests exceeding 200k tokens on these models now return an error. Migrate to Sonnet 4.6 or Opus 4.6 for continued 1M context support — no beta header required, no long-context surcharge.
Retired header: context-1m-2025-08-07 | Affected models: claude-sonnet-4-5, claude-sonnet-4-20250514 | Next deprecation: Sonnet 4 + Opus 4 full retirement June 15, 2026 | Migration targets: claude-sonnet-4-6, claude-opus-4-6, claude-opus-4-7
Best for: API developers — action required if still on Sonnet 4 or Sonnet 4.5 with long context
Anthropic — Sydney office + ANZ expansion
Anthropic named Theo Hourmouzis as General Manager of Australia and New Zealand and officially opened its Sydney office this week. This follows the NEC Japan partnership announced the prior week, and signals continued international infrastructure investment alongside the AWS compute deal.
Sydney office officially open. Theo Hourmouzis appointed GM for Australia and New Zealand. Anthropic now has a direct regional presence to serve ANZ enterprise customers.
ANZ GM: Theo Hourmouzis | Region: Australia and New Zealand | Follows: NEC Japan partnership (April 24), AWS $25B compute deal (April 20), Google/Broadcom TPU expansion (April 6)
Best for: ANZ enterprise customers, regional partnership inquiries
Plans and Pricing
No consumer or API pricing changes this period. A note for GitHub Copilot users: the promotional pricing for Opus 4.7 on Copilot ended April 30 — the premium request multiplier reverted to 15x. No changes to direct Anthropic API pricing.
Opus 4.7: $5/$25 per MTok (unchanged) | Sonnet 4.6: $3/$15 per MTok | Haiku 4.5: low-cost tier | GitHub Copilot: Opus 4.7 promo ended April 30, multiplier now 15x | Managed Agents: contact enterprise sales
Best for: No action needed on pricing — unless on GitHub Copilot with Opus 4.7
ChatGPT / OpenAI
Dateline: May 04, 2026 | Next update: May 11, 2026
Over the past week, OpenAI has continued expanding professional use cases while refining system reliability, multimodal performance, and agent autonomy.
GPT-5.3 Standard — default model
More stable performance across mixed (text + image) tasks.
Improved consistency when handling multimodal inputs.
Context ~128k | Output ~4k–8k | Improved cross-modal alignment
Best for: General use
GPT-5.3 Pro — high-reasoning model
Better handling of complex, multi-domain reasoning.
Improved integration of external knowledge during long reasoning chains.
Context ~200k (est.) | Enhanced reasoning coherence across domains
Best for: Deep analysis
GPT-5.3 Mini — fallback model
More efficient and less noticeable fallback usage.
Reduced quality drop when switching between models mid-task.
Context ~64k | Improved dynamic routing
Best for: Quick tasks
Agent Mode
Handles multi-step workflows.
Improved ability to maintain direction over extended, multi-step tasks.
Better long-horizon task tracking | Reduced drift
Best for: Task delegation
Deep Research
Combines browsing and reasoning.
Improved synthesis across conflicting or low-confidence sources.
Enhanced conflict resolution in ranking pipeline
Best for: Research
Memory & Projects
Persistent context across chats.
Improved contextual recall across longer project timelines.
Better long-term memory weighting | Reduced redundancy
Best for: Ongoing workflows
Advanced Voice Mode
Improved conversational flow.
More natural tone adaptation in longer conversations.
Improved prosody modelling | Better conversational continuity
Best for: Voice interaction
ChatGPT for Clinicians
Expansion into healthcare workflows continues.
Early refinements based on usage — better structuring of clinical summaries and improved reliability in medical-context responses.
Iterative tuning on domain-specific outputs | Strengthened guardrails
Best for: Clinical support (non-diagnostic assistance)
Other Features & Pricing
Continued improvements in multimodal reasoning (text, image, document inputs). General system stability improvements across tools and integrations. No major updates to Sora, Codex, or pricing.
What this means
This week reinforces a clear pattern: OpenAI is moving from capability expansion to capability consolidation — while quietly scaling into real-world domains.
Three things stand out:
First, multimodal reliability is becoming a priority. Improvements here suggest that OpenAI sees the future not as text-first, but as fully integrated inputs (documents, images, context) — especially for professional use.
Second, Agent Mode is becoming more autonomous in practice, not just in theory. The improvements around task persistence and reduced drift indicate movement toward systems that can handle longer workflows with less user correction.
Third, the continued iteration on ChatGPT for clinicians shows that last week's launch was not experimental — it's being actively developed. This is how OpenAI is likely to enter other sectors: launch early, refine quickly, and expand laterally into adjacent professional domains.
Overall, this is a deployment phase week. The technology itself is stabilising, while the strategic focus is shifting toward embedding AI into high-value, real-world workflows.
Gemini
Date: May 4, 2026 | Next update: May 11, 2026
Following the transition period at the end of April, this week focuses on the expansion of Gemini into new hardware environments — specifically the living room via Google TV — and major breakthroughs in DeepMind's healthcare and scientific research initiatives.
Core models — Gemini 3.1 series
Gemini Robotics-ER 1.5 (retired): The 1.5-preview model was officially sunset on April 30, 2026. Developers should now be fully migrated to Gemini Robotics-ER 1.6-preview to maintain access to instrument reading and spatial reasoning capabilities.
Stability & scaling: Following the end-of-month migrations, infrastructure has been optimized for the Gemini 3.1 family, with no reported outages in API batch processing during the past week.
Gemini on Google TV — new
Google has integrated Nano Banana (Gemini 3.1 Flash Image) and Veo capabilities directly into Google TV. Users can now generate images, manipulate photos, and create AI-powered video content directly on their living room screens.
Nano Banana & Veo integration on Google TV — generate images and create AI video content from your TV.
Best for: Living room creativity
Healthcare research
Google DeepMind has prioritized "triadic care" research, where AI agents assist in patient care journeys under physician authority. Evaluations in primary care settings have demonstrated high factual grounding with zero critical errors in 97/98 cases.
Healthcare AI co-clinician research expanded — new evaluations showing high accuracy in primary care settings.
Best for: Scientific & health research
Scientific partnerships
Google expanded its open-resource initiatives, focusing on data mining and bioscience models designed to catalyze global scientific discovery.
Plans and migration deadlines
Whisk to Flow transition (finalized): The April 30, 2026 migration deadline has passed. All experimental assets not migrated to the Flow studio have been processed; users encountering missing files should contact support immediately.
Upcoming feature rollouts: With the successful integration of Gemini into television platforms, expect further "ambient AI" features for other smart home devices to be discussed at upcoming developer forums.
Service status: The post-April 30 period has been stable, with no significant disruptions to API or Workspace services reported. Usage caps remain in effect for the new Flex and Priority inference tiers.
Microsoft Copilot
Dateline: April 24–May 4, 2026 | Next update: May 11, 2026
This period brought a mix of product improvements, reliability upgrades, and expanded integrations across the Copilot ecosystem.
Copilot for Code — expanded developer tooling
Improved multi-file context awareness in pull request reviews — Copilot now tracks dependencies across files when suggesting changes. Usage dashboard added — developers can see token consumption and request breakdowns directly in IDE extensions.
Context window: up to 200k tokens for Pro tier | New /usage command in GitHub integration | Expanded support for JetBrains IDEs
Best for: Developers managing large codebases and CI/CD pipelines
Copilot for Office — design & productivity updates
PowerPoint Designer integration — Copilot can now generate slide layouts directly from Word documents or meeting transcripts. Excel formula explanations — natural language breakdowns of complex formulas now available in sidebar.
Export formats: PPTX, DOCX, XLSX | Admin toggle for enterprise rollout | Integration with SharePoint and Teams confirmed
Best for: Enterprise teams creating presentations, reports, and structured data workflows
Copilot Security — enterprise preview
Copilot added vulnerability explanation mode — security findings are now explained in plain language alongside technical details. Integration with Sentinel and Defender for Cloud for automated patch recommendations.
Findings exportable to Azure DevOps and Jira | Multi-stage validation reduces false positives | Directory-level scanning supported
Best for: Security teams embedding AI into vulnerability management workflows
Copilot Voice — natural conversation improvements
Adaptive prosody — speech output adjusts tone based on conversation length and context. Meeting summarization in Teams voice calls now includes action item extraction.
Expanded prosody modeling | Action item tagging integrated with Microsoft Planner | Voice mode latency reduced by ~15%
Best for: Professionals using Copilot in meetings and live collaboration
Copilot Platform — reliability & scaling
Retired legacy 1M context beta for older models; standard 200k context now enforced. Migration required to newer models for long-context support. Expanded telemetry logging for enterprise deployments.
Deprecated header: context-1m-2025-08-07 | Migration targets: latest Copilot API models | Expanded OpenTelemetry hooks
Best for: API developers and enterprise IT teams
Plans and Pricing
No consumer or API pricing changes this period. Promotional pricing for extended context models ended April 30. Standard multipliers now apply across GitHub and Microsoft 365 integrations.