Dateline: July 10, 2026 | Next update: July 17, 2026
A week defined by government and enterprise. Claude for Government launched in public beta with Claude Code and Cowork in a FedRAMP High environment on July 7 — Anthropic's clearest signal yet that it is competing seriously for US public-sector contracts. The same day: Cowork expanded to web and mobile (Max plan first), the Microsoft 365 connector gained write access for the first time, and Claude Science launched as a dedicated research platform. The Fable 5 subscription window was extended through July 12. Claude Code shipped a broad performance and stability update with 37% lower CPU use during streaming. On the OpenAI side, GPT-5.6 launched with three tiers (Sol, Terra, Luna) across ChatGPT, Codex, and the API on July 9, GPT-Live brought full-duplex voice to ChatGPT, and ChatGPT for PowerPoint reached GA for Business. Google expanded Managed Agents with async background tasks and remote MCP support; Gemini 3.5 Pro remains in limited preview; a faulty backend config caused a brief Gemini 2.5 API outage on July 9 that was quickly rolled back.
Claude / Anthropic
✅ Claude for Government — FedRAMP High public beta
Claude for Government Desktop launched in public beta on July 7, making Claude Code and Claude Cowork available to US federal agencies inside a FedRAMP High authorized environment for the first time. It is built on the same commercial product codebase, giving agencies access to new capabilities on the same shipping cadence as commercial users. Anthropic remains the contracted and billing party — no separate cloud-provider relationship is required to get started.
Claude Code in government: public sector teams can build and modernize software systems directly using Claude Code inside the FedRAMP High environment. Claude Cowork in government: agency staff can delegate memo creation, RFP reviews, casework, and decks to Claude working across local desktop files. Governance controls: conversation history stored locally on agency-managed devices; inference inside the FedRAMP High authorized environment; department-level administration with sub-agency seat and spend allocation; hash-chained tamper-evident audit logs reviewable by org admins; sensitive Anthropic-side operations require two-person approval; usage exports are metering data only. Billing: fixed spending increments with a hard not-to-exceed cap; burndown alerts before balance runs low. SCIM group mappings set rate limits, dollar caps, and allowed models per seat tier. FedRAMP Secure Configuration Guide published as a public document. Penetration-test summary available via Anthropic trust center (NDA). Deploys through standard agency MDM platforms.
Authorization: FedRAMP High (also available via Bedrock GovCloud FedRAMP High, Vertex Assured Workloads FedRAMP High, and AWS Secret region IL6) | NIST 800-171r3 attestation: available under NDA | Audit: hash-chained log, two-person approval for sensitive Anthropic ops | Billing: fixed increments, hard NTE cap, burndown alerts | SCIM: rate limits, dollar caps, model restrictions per group | Access: claude.com/solutions/government | C4G pricing: $60/seat/month; $1/month for federal/judicial/legislative | Claude Enterprise (non-FedRAMP): $20/seat/month + PAYG
Best for: US federal, judicial, and legislative agencies; defense contractors via Bedrock GovCloud or Vertex Assured Workloads; public sector IT and security teams
Claude Cowork — web and mobile expansion
Claude Cowork — which launched as a desktop-only research preview in January 2026 and went GA in April — expanded to web and mobile on July 7. Sessions now run remotely in the cloud, so work follows the user across devices. This is a meaningful architecture shift: previously, Cowork required a desktop device to remain online for sessions and scheduled tasks to run. Now neither is required. Anthropic published a usage analysis from 1.2 million anonymised Cowork sessions across 600,000+ organisations, revealing that more than 90% of Cowork use has nothing to do with software development.
Cowork is now available on claude.ai (web) and in the Claude app (iOS and Android sidebar), in addition to Claude Desktop. Beta rolling out over the next several weeks starting with Max plan. Remote sessions: sessions and files are saved to your Claude account and continue across devices — close your laptop and pick up on your phone. Scheduled tasks run with no device online. Chat and Cowork now share one unified home tab on web and desktop — one sidebar, one search, one place for projects and artifacts. Usage limits doubled through August 5 to encourage bigger tasks during the beta. Usage data (1.2M sessions, 600k+ orgs, last two weeks of May 2026): 33.4% of sessions were business process and operations (reports, checklists, spreadsheet reconciliation); 16.4% were content creation and copywriting; software development accounted for only 8.7%.
Platforms: claude.ai web, Claude iOS app, Claude Android app, Claude Desktop (existing) | Rollout: Max plan first, other plans to follow | Session model: remote cloud execution, files and state saved to account | Scheduled tasks: run with no device online | Usage limits: 2× through August 5 | Chat + Cowork: unified home tab, shared sidebar, shared Projects and Artifacts | Usage data: 1.2M sessions, 600k+ orgs, May 2026 sample
Best for: Max plan subscribers: available now. Enterprise teams running overnight or multi-day Cowork tasks no longer need a device to stay on.
Microsoft 365 connector — write tools enabled
The Microsoft 365 connector — available since late 2025 for read and search operations — gained write capabilities on July 7. Claude can now draft and send email, manage calendar events, update mailbox settings, and create or update files in OneDrive and SharePoint directly from a Cowork or chat session. Teams remains read-only. This is a posture change: from a tool you consult, to an agent you delegate to. Write tools must be explicitly enabled by an administrator.
New M365 write capabilities (all require Entra admin consent + org admin enable): Email — draft, send, organise, and manage drafts from Claude directly within your existing M365 permissions; Calendar — create, update, and delete calendar events; Mailbox settings — update mailbox configuration; OneDrive and SharePoint — create and update files. Read and search tools continue to work as before. Teams: read-only, unchanged. Before enabling, a Microsoft Entra administrator must consent to the updated permission set and an org admin must enable the tools for the organisation in the Claude admin console.
Write tools: email (draft/send/organise), calendar (create/update/delete), mailbox settings, OneDrive/SharePoint (create/update files) | Teams: read-only | Prerequisites: Microsoft Entra admin consent + org admin enable in Claude console | Plans: Claude Enterprise | Read/search: unchanged | Governance note: run a permissions audit before enabling and define a write-tools policy in advance
Best for: Enterprise teams who have already vetted their M365 permission surfaces and defined clear human-checkpoint policies for AI-delegated email and calendar actions
Claude Science — dedicated research platform
Anthropic launched Claude Science, a customisable application for scientific researchers that integrates the tools and packages researchers use most often, produces auditable research artifacts, and provides flexible access to computing resources. It is designed to lower the barrier between a research question and a working analysis pipeline, without the setup overhead of configuring Claude from scratch for scientific work.
Claude Science launches as a dedicated scientific research application. Key features: integrates commonly used research tools and packages (statistical, computational, domain-specific) out of the box; produces auditable artifacts suitable for research documentation and reproducibility; flexible compute resource access for workloads beyond standard chat limits. Positioned for researchers, scientific teams, and institutions. The Government of Alberta was disclosed as an early customer using Claude for cybersecurity vulnerability scanning across government systems.
Application type: customisable research platform built on Claude | Key capabilities: tool and package integrations, auditable artifacts, flexible compute access | Target users: academic researchers, institutional science teams, R&D organisations | Early customer: Government of Alberta (cybersecurity vulnerability scanning) | Distinct from: Claude Security (vulnerability scanning for enterprise code), Project Glasswing (Mythos-class vulnerability research)
Best for: Scientific researchers and institutional R&D teams needing a pre-configured Claude environment with research-grade tooling and artifact auditability
Monthly recap and focus settings
Anthropic added two new settings oriented around helping users be more intentional about how they use Claude. The monthly recap shows patterns in how you have been working with Claude; focus settings let you set quiet hours and break reminders. Both are in beta.
Monthly recap (Settings → Reflect): shows topics you spent time on, your most active day and peak hour, and observations about how you work with Claude. In beta for Free, Pro, and Max on web and Claude Desktop; requires Memory to be on. Cowork conversations will be included soon. Focus settings (Settings → Time and focus): optional break reminders and quiet hours. Both settings available on claude.ai and Claude Desktop.
Monthly recap: Settings → Reflect | Plans: Free, Pro, Max (beta) | Requires: Memory enabled | Cowork integration: coming soon | Focus settings: Settings → Time and focus | Features: break reminders, quiet hours | Platforms: claude.ai web, Claude Desktop
Best for: Any Claude user who wants visibility into their usage patterns or wants to set work/rest boundaries around Claude use
Claude Code — 37% CPU reduction, worktree fixes, login warnings
Claude Code's point releases this week focused on performance and reliability rather than new features. The most impactful change is a 37% reduction in CPU use during streaming, alongside a batch of worktree and background agent fixes that had been causing sessions to silently drop or re-run work from scratch.
Performance: CPU usage during streaming responses reduced ~37% by coalescing text updates to 100ms intervals. Long-session memory growth from terminal output cache reduced. Login-expiry warnings added so sessions warn before credentials expire mid-task rather than failing silently. Clearer agent status and manual mode badges in the agents view. Transcript protection: auto mode now blocks tampering with session transcript files. /doctor now runs a full setup checkup with actionable output. Auto-update downloads use lower memory. Fixed: returning to claude agents silently stopping running subagents and re-running the prompt from scratch — their work now carries over correctly. Fixed: memory and per-turn CPU regression in interactive sessions where the context-usage indicator was re-analysing the entire transcript after every turn. Fixed: background agents inheriting a stale PATH from the daemon instead of the dispatching shell (caused missing tools on Windows). Fixed: background sessions dropping a shell-exported ANTHROPIC_BASE_URL (sent API keys to default endpoint, failed with 401). Fixed: Bash failing with 'argument list too long' in repos with many git worktrees. Fixed: worktree-isolated subagents sometimes running shell commands in the parent checkout instead of their own worktree. Fixed: worktree creation rejecting nested repositories in multi-repo workspaces.
CPU: ~37% reduction in streaming by 100ms text-update coalescing | Memory: long-session terminal output cache growth reduced | Login expiry: warning before credentials expire | Transcript protection: auto-mode blocks transcript tampering | /doctor: full setup checkup | Fixed: claude agents silently re-running prompt from scratch | Fixed: context-usage indicator per-turn full transcript re-analysis | Fixed: stale PATH in background agents (Windows) | Fixed: ANTHROPIC_BASE_URL dropped in background sessions | Fixed: git worktree 'argument list too long' | Fixed: worktree subagents running in parent checkout | Fixed: nested repo worktree creation rejection | Reserved: 'Claude Browser' and 'Claude Preview' MCP server names ahead of Claude Desktop pane rename
Best for: All Claude Code users — the claude agents worktree fix is significant. Update to the latest version if you run background agents in multi-repo workspaces or on Windows.
⚠ Fable 5 — free subscription window extended to July 12
The Fable 5 free subscription window — originally set to transition to usage-credit pricing on July 8 — was extended through July 12 alongside the Cowork web and mobile launch. After July 12, Fable 5 requires usage credits on subscription plans. The 50%-of-weekly-limits cap applies during the extended window.
Fable 5 remains free for paid subscription plans (up to 50% of weekly usage limits) through July 12, 2026. Extension granted alongside the Cowork web and mobile rollout. After July 12: usage credits required on subscription plans. API pricing unchanged: $10/$50 per MTok standard, $5/$25 per MTok Batch API.
Extended free window: through July 12, 2026 | Cap: 50% of weekly usage limits | After July 12: usage credits required | API: $10/$50 per MTok standard, $5/$25 Batch (unchanged) | Mythos 5: US Glasswing organisations only, no change
Best for: Subscription users: use Fable 5 freely through July 12 within the 50% weekly cap. After July 12, budget for usage credits or fall back to Sonnet 5 or Opus 4.8.
Plans and Pricing
No new model launched this week. Operative changes: Fable 5 extended free through July 12 (then usage credits); Cowork usage limits doubled through August 5 for all plans. Claude for Government: $60/seat/month or $1/month for qualifying federal/judicial/legislative agencies. Sonnet 5 introductory pricing ($2/$10 per MTok) continues through August 31.
Fable 5: free (50% weekly cap) through July 12; usage credits from July 13 | Sonnet 5: $2/$10 per MTok through August 31, then $3/$15 | Opus 4.8: $5/$25 per MTok | Haiku 4.5: low-cost tier | Cowork limits: 2× through August 5 | Claude for Government: $60/seat/month; $1/month for federal/judicial/legislative | Opus 4.7 fast mode removal: July 24 | Opus 4.1 retirement: August 5
Best for: Maximise Fable 5 use before July 12. Take advantage of doubled Cowork limits through August 5 for larger tasks. Government teams: request access at claude.com/solutions/government.
ChatGPT / OpenAI
Dateline: July 10, 2026 | Next update: July 17, 2026
A major model and product week for OpenAI. GPT-5.6 launched across ChatGPT, Codex, and the API on July 9, introducing three tiers — Sol, Terra, and Luna — plus new max and ultra reasoning options for more demanding work. GPT-Live launched on July 8, bringing a more natural full-duplex voice experience to ChatGPT Voice. ChatGPT for PowerPoint became generally available for Business workspaces on July 6. OpenAI also moved the App Directory into the new Plugin Directory, positioning plugins as the main way to discover repeatable workflow capabilities. Codex joined the ChatGPT desktop app, Codex iOS gained deeper task-management tools, and OpenAI published national-security principles for government partnerships.
GPT-5.6 — new frontier model family (Sol, Terra, Luna)
OpenAI launched GPT-5.6, its new frontier model family spanning three tiers: Sol (flagship), Terra (lower-cost, competitive with GPT-5.5), and Luna (fastest and most affordable). This is a meaningful product shift — OpenAI is moving towards durable model tiers rather than a single headline model name. The launch also introduces max reasoning for GPT-5.6 and ultra mode for complex multi-agent work.
GPT-5.6 is available across ChatGPT, Codex, and the API. Plus, Pro, Business, and Enterprise users can access GPT-5.6 Sol in ChatGPT through medium and higher effort settings. Pro and Enterprise users can also select GPT-5.6 Sol Pro for the highest-quality responses on complex work. In ChatGPT Work and Codex, Free and Go users get GPT-5.6 Terra; paid users can choose between Sol, Terra, and Luna. New: max reasoning for GPT-5.6 and ultra mode for eligible users on complex multi-agent work.
Models: GPT-5.6 Sol, GPT-5.6 Terra, GPT-5.6 Luna | Surfaces: ChatGPT, ChatGPT Work, Codex, OpenAI API | Reasoning: max (all tiers); ultra (eligible users) | API: Responses API supports Programmatic Tool Calling and beta multi-agent execution | Pricing: Sol $5/$30 per MTok; Terra $2.50/$15; Luna $1/$6 | Prompt caching: explicit cache breakpoints, 30-minute minimum cache life, writes at 1.25× uncached input rate, reads retain 90% discount
Best for: Developers, enterprise teams, and advanced ChatGPT users needing higher-quality reasoning, stronger coding, and better long-horizon agentic performance. Evaluate Sol/Terra/Luna by workload rather than defaulting to the flagship.
GPT-5.6 becomes preferred model in Microsoft 365 Copilot
OpenAI announced that GPT-5.6 will become the preferred model in Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat, and Cowork, bringing the new model family directly into the productivity tools used by millions of enterprise workers.
Microsoft expects GPT-5.6 to improve drafting, analysis, presentation creation, and cross-functional collaboration. OpenAI frames the update as delivering more useful work per token and stronger performance per dollar in existing Microsoft workflows.
Platform: Microsoft 365 Copilot | Apps: Word, Excel, PowerPoint, Copilot Chat, Cowork | Model: GPT-5.6 via OpenAI API | Focus: document drafting, spreadsheet analysis, presentation generation, collaborative work
Best for: Microsoft 365 enterprise customers and productivity teams already using Copilot for daily workflows
GPT-Live — new full-duplex voice model
OpenAI launched GPT-Live, a new generation of voice models designed to make conversations with ChatGPT feel more natural. The key change is full-duplex architecture: GPT-Live can listen and speak at the same time, allowing more fluid back-and-forth rather than rigid turn-taking. At launch, GPT-Live uses GPT-5.5 behind the scenes for complex reasoning or work tasks.
GPT-Live powers the new ChatGPT Voice experience globally. It can keep the conversation flowing while delegating harder tasks to a frontier model in the background. GPT-Live-1 and GPT-Live-1 mini are rolling out to ChatGPT users; API access planned but not yet broadly available.
Models: GPT-Live-1, GPT-Live-1 mini | Architecture: full-duplex voice | Surface: ChatGPT Voice | Background model at launch: GPT-5.5 | API: planned, not yet broadly available | Use cases: natural conversation, language practice, hands-free help, longer voice interaction
Best for: ChatGPT Voice users, accessibility workflows, language learning, coaching, and hands-free productivity
ChatGPT for PowerPoint — GA for Business
ChatGPT for PowerPoint is now generally available for Business workspaces. Teams can create and revise editable presentations directly inside PowerPoint, ask questions about deck structure, improve narrative flow, and use Skills and enabled apps to build slides from repeatable workflows and connected sources. Business usage remains free through August 6.
Platform: Microsoft PowerPoint | Plan: ChatGPT Business | Admin controls: enabled by workspace admins | Pricing: free through August 6; then flexible-pricing/credit-pool model | Pool: Business plans include usage for Workspace Agents, ChatGPT for Excel, and ChatGPT for PowerPoint through the general Codex agentic usage pool
Best for: Business teams producing recurring decks, leadership briefings, sales presentations, and client-facing materials. Review PowerPoint and Workspace Agent usage before August 6.
Workspace Agent pricing begins
The free period for Workspace Agents ended on July 6 and credit-based pricing began. Workspace Agent runs now use token-based pricing based on input tokens, cached input tokens, and output tokens. Admins can view workspace-agent activity and usage in the admin console.
Applies to: Workspace Agents in Business, Enterprise, and Edu | Pricing: token-based credit usage | Metering: input tokens, cached input tokens, output tokens | Admin visibility: workspace-agent activity and usage in admin console
Best for: Workspace admins and finance teams managing agentic-workflow spend now that the free period has ended
Plugin Directory replaces App Directory
OpenAI migrated the App Directory to the new Plugin Directory. Plugins are now the primary way to discover workflow capabilities across ChatGPT and Codex. A plugin can include skills, apps, and app templates; existing app connections are unaffected.
Directory: Plugin Directory replaces App Directory | Components: skills, apps, app templates | Admin controls: Workspace settings → Plugins; app permissions still governed in Apps | Surfaces: ChatGPT web, desktop, ChatGPT Work, and ChatGPT Codex | Existing app connections: unaffected
Best for: Workspace admins and builders creating repeatable AI workflows across connected tools
Codex in the ChatGPT desktop app
Codex is now part of the ChatGPT desktop app on macOS and Windows. Existing Codex app users can update as usual while keeping their projects, settings, and workflows. New capabilities include Markdown and code editing directly in the app, inline annotations, GitHub PR review in the sidebar, and multi-repository projects.
Platforms: ChatGPT desktop for macOS and Windows | Features: Markdown/code editing, inline annotations, selected-content revision, GitHub PR review sidebar, multi-repository projects | Performance: faster Computer Use with GPT-5.6, clearer task activity, plugin management moved into Settings, better mobile connection reliability
Best for: Developers and technical teams who want Codex integrated directly into the main ChatGPT desktop environment
Codex CLI 0.144 — approval modes, MCP auth, GPT-5.6 readiness
Version 0.144.0 added usage-credit visibility, a new writes app-approval mode (allows declared read-only actions while prompting before write actions), interactive MCP authentication by default, and warnings when Ultra reasoning may increase usage quickly. Version 0.144.1 followed with installer and code-mode reliability fixes.
Version: Codex CLI 0.144.0 / 0.144.1 | Install: npm install -g @openai/codex@0.144.1 | New: reset-credit details, writes approval mode, MCP interactive auth, hosted app-server login redirects, Ultra reasoning usage warning | Fixes: Intel macOS Code Mode crash, Windows sandbox deletion, terminal control sequence corruption, expired hosted connector auth refresh, proxy/CA support for Responses WebSockets
Best for: Codex CLI users, enterprise developers, and security-conscious teams managing agent permissions across apps and MCP tools
Codex iOS — deeper task management from mobile
Codex on iOS received a major task-management update. Users can now create, search, open, fork, and manage Codex tasks directly from a conversation, with filters for staged, unstaged, branch, and last-turn changes, plus branch comparison controls.
Version: ChatGPT for iOS 1.2026.181 | Features: Codex task creation/search/open/fork/manage, branch/change filters, selected-transcript composer insertion, attachment previews, Photos/Camera picker, SSH host support, usage/credit details | Fixes: task loading, foreground recovery, reconnects, host-pairing preservation, stale images, microphone permission alerts
Best for: Codex users who manage long-running agent tasks from mobile or need to approve and steer work away from their desk
⚠ OpenAI publishes national-security principles
OpenAI published its National Security Principles, setting out how it approaches government and national-security partnerships. Democratic societies should be able to use AI for cyber defence, biosecurity, public services, and critical-infrastructure protection, but deployments must reinforce democratic accountability, human judgement, and the rule of law. Restrictions include no use of OpenAI technology for mass domestic surveillance, directing autonomous weapons systems, or high-stakes automated decisions.
Scope: government, national security, law enforcement partnerships | Areas: cyber defence, biological security, critical infrastructure, public services | Restrictions: no mass domestic surveillance, no autonomous weapons targeting, no high-stakes automated decisions | Partnerships referenced: Daybreak (cyber-defence trusted access with allied governments and EU institutions); GPT-Rosalind (public-health and biodefence missions)
Best for: Policymakers, defence analysts, AI governance teams, and organisations tracking how frontier AI labs are positioning themselves in sensitive government use cases
Plans and Pricing
The biggest pricing change this week is GPT-5.6 API pricing and the start of Workspace Agent credit-based pricing. GPT-5.6 Sol: $5/$30 per MTok; Terra: $2.50/$15; Luna: $1/$6. GPT-5.6 also introduces explicit cache breakpoints, a 30-minute minimum cache life, cache writes at 1.25× the uncached input rate, and cache reads retaining the 90% cached-input discount. ChatGPT for PowerPoint free through August 6 for Business; Workspace Agent runs on token-based credit pricing from July 6.
GPT-5.6 Sol: $5/$30 per MTok | GPT-5.6 Terra: $2.50/$15 | GPT-5.6 Luna: $1/$6 | Cache writes: 1.25× uncached input rate | Cache reads: 90% discount | ChatGPT for PowerPoint Business: free through August 6 | Workspace Agents: token-based credit pricing from July 6
Best for: Developers should evaluate Sol/Terra/Luna by workload. Business admins: review PowerPoint and Workspace Agent usage before free periods end or credit consumption increases.
Gemini (Google)
Dateline: July 10, 2026 | Next update: July 17, 2026
A week marked by architectural expansions and an unexpected infrastructure incident. Google expanded Managed Agents within the Gemini API on July 7 with asynchronous background tasks, remote MCP server support, and a unified Interactions API. Gemini 3.5 Pro remains in limited enterprise preview as an architectural rebuild continues, now targeting a July 17 release. On July 9, a faulty backend configuration accidentally triggered premature deprecation errors for legacy Gemini 2.5 models worldwide; Google rolled back the change within hours and confirmed the October 16 deprecation date is unchanged.
Managed Agents in Gemini API — async tasks and remote MCP
Google expanded the scope of Managed Agents within the Gemini API, enabling developers to build, scale, and deploy resilient autonomous workflows. The framework now allows stateful AI agents to run continuously within secure, isolated Google-hosted Linux sandboxes without requiring an active client device. The most notable inclusion is native support for remote MCP configurations, allowing external tools and corporate repositories to connect to Google's hosted sandboxes using open-source, vendor-agnostic protocols. A unified Interactions API coordinates interactions between foundation models, custom functions, and Google's built-in tools.
Stateful agents run in secure Google-hosted Linux sandboxes with no device required. Native remote MCP support: connect external programming suites, third-party frameworks, and corporate repositories via open-source tool protocols. Unified Interactions API: coordinates foundation models, custom functions, and built-in Google tools (Search, Maps) — prevents agents from overriding system prompts or breaching operational parameters during background cycles. Supports long-horizon multi-hour reasoning, programmatic research pipelines, and background data processing.
Platform: Gemini API / Google AI Studio / Google Cloud | Models: gemini-3.5-flash and antigravity-preview-05-2026 | Runtime: gated Linux container sandboxes with ephemeral storage | Protocol: open-source MCP remote endpoints | Core interface: Interactions API with granular function calling | Task state: stateful persistence for long-horizon background async queues
Best for: Backend engineers and enterprise automation teams designing autonomous agents with persistent background processing or multi-cloud database integration
⚠ Gemini 3.5 Pro — limited preview extended, target July 17
Gemini 3.5 Pro has missed its early-summer release targets and remains in limited Vertex AI enterprise preview. The delay stems from an executive decision to discard the existing baseline architecture in favour of a ground-up pre-training cycle, targeting fixes for mathematical logic regressions, SVG compilation issues, and token consumption inefficiency observed during multi-turn testing. The new architecture introduces a 2M token context window and an integrated Deep Think reasoning layer designed to manage sub-agents efficiently.
Platform: Vertex AI (Gemini Enterprise Agent Platform) | Model: gemini-3.5-pro (pre-commercial preview) | Context: 2M token input window | Reasoning: dual-layer with integrated Deep Think module | Orchestration: native governance of gemini-3.5-flash sub-agents | Target release: revised to July 17, 2026
Best for: Enterprise technology officers and procurement teams tracking Gemini 3.5 Pro before committing production architecture budgets
⚠ Gemini 2.5 API outage — rollback completed
On July 9, developers globally experienced a production outage when calls to gemini-2.5-flash and gemini-2.5-pro unexpectedly returned 404 errors declaring the models "no longer available" — contradicting Google's documented deprecation date of October 16, 2026. The cause was a flawed backend configuration script deployed during routine maintenance that forced the API router to treat active models as retired. Google rolled back the change within hours, restoring full availability, and issued a formal apology. Teams are advised to plan migration toward gemini-3.1-flash-lite and gemini-3.5-flash ahead of the autumn deprecation window.
Affected: gemini-2.5-flash, gemini-2.5-flash-lite, gemini-2.5-pro | Error: HTTP 404 ("model no longer available") | Root cause: flawed backend configuration script deployed during routine maintenance | Resolution: completed backend rollback | Official deprecation date: October 16, 2026 (unchanged)
Best for: DevOps engineers managing active codebases on Gemini 2.5 — review error logs from July 9 and audit migration schedules ahead of the October deprecation.
Plans and Pricing
No structural price changes to Gemini model pricing this week. Gemini 3.5 Flash introductory pricing ($2/$10 per MTok) remains active through August 31. Managed Agent sandbox costs scale by compute-per-minute of the underlying container instance. Deprecation roadmap: Imagen 4 and early Gemini 3 Image endpoints shut down August 17; early Veo architectures already turned off June 30 (migrate to Veo 3.1 immediately); Gemini 2.5 family deprecation reconfirmed for October 16.
Gemini 3.5 Flash: $2/$10 per MTok through August 31 | Managed Agent sandbox: base API token cost + container runtime overhead | Imagen 4 + Gemini 3 Image sunset: August 17, 2026 | Veo 2.0/3.0 standard: already ended June 30 (migrate to Veo 3.1) | Gemini 2.5 family sunset: October 16, 2026 (unchanged)
Best for: Enterprise engineers who need to migrate off deprecated image and video endpoints before August 17 while taking advantage of locked-in Gemini 3.5 Flash promotional rates
Microsoft Copilot
Dateline: July 10, 2026 | Next update: July 17, 2026
Microsoft's biggest announcements this week mirrored Anthropic's: Copilot for Government launched in FedRAMP High public beta on July 7 alongside Copilot Cowork's expansion to web and mobile, write tools for Microsoft 365, Copilot Science, and a suite of wellbeing and performance updates. The week rounds out a major push into both the public sector and enterprise productivity.
✅ Copilot for Government — FedRAMP High public beta
Copilot for Government Desktop launched in public beta, making Copilot Code and Copilot Cowork available to US federal agencies inside a FedRAMP High authorized environment. Agencies gain access to commercial-grade Copilot features with government-specific governance controls: local conversation storage, tamper-evident audit logs, department-level administration, SCIM group mappings, and fixed spending increments with hard caps.
Authorization: FedRAMP High | NIST 800-171r3 attestation available | Audit logs: hash-chained | Billing: fixed increments, hard NTE cap | SCIM: group mappings for rate limits and model restrictions | Pricing: $60/seat/month; $1/month for federal/judicial/legislative
Best for: US federal, judicial, and legislative agencies; defense contractors; public sector IT and security teams
Copilot Cowork — web and mobile expansion
Copilot Cowork expanded to web and mobile, allowing tasks to run remotely in the cloud without requiring a device to stay online. Unified home tab merges Chat and Cowork into one sidebar, search, and project space.
Available on copilot.microsoft.com (web) and Copilot mobile apps (iOS/Android). Remote sessions: work continues across devices; scheduled tasks run even if no device is online. Unified home tab merges Chat and Cowork. Usage limits doubled through August 5.
Platforms: web, iOS, Android, desktop | Remote cloud execution | Files and state saved to account | Usage limits: 2× until August 5 | Unified sidebar and projects
Best for: Enterprise teams running overnight or multi-day Cowork tasks without device dependency
Microsoft 365 — write tools enabled
The Microsoft 365 connector gained write capabilities for the first time. Copilot can now draft and send email, manage calendar events, update mailbox settings, and create or update files in OneDrive and SharePoint. Teams remains read-only.
Write tools: email (draft/send/organise), calendar (create/update/delete), mailbox settings, OneDrive/SharePoint (create/update) | Teams: read-only | Prerequisites: Entra admin consent + org admin enable | Governance note: permissions audit recommended before enabling
Best for: Enterprise teams delegating email, calendar, and file tasks to Copilot with clear human-checkpoint policies
Copilot Science — dedicated research platform
Microsoft launched Copilot Science, a customisable application for researchers integrating scientific tools and packages, producing auditable artifacts, and providing flexible compute resources. Early customer: Government of Alberta (cybersecurity vulnerability scanning).
Capabilities: tool integrations, artifact auditability, flexible compute | Target: academic researchers, institutional R&D
Best for: Scientific researchers and institutional R&D teams needing reproducible, audit-ready AI pipelines
Monthly recap and focus settings
Monthly recap (Settings → Reflect) shows usage patterns, peak hours, and topics. Focus settings (Settings → Time and focus) add break reminders and quiet hours.
Best for: Any Copilot user wanting visibility into usage patterns and healthier work boundaries
Copilot Code — performance and reliability
Copilot Code updates focused on performance this week: ~35% CPU reduction during streaming by batching text updates, login-expiry warnings, transcript protection, clearer agent status badges, and fixes for background agents, worktree handling, and multi-repo sessions.
CPU: ~35% reduction | Transcript tamper protection | /doctor setup checkup | Fixed: stale PATH inheritance, ANTHROPIC_BASE_URL drops, git worktree errors
Best for: Developers running background agents in multi-repo workspaces or Windows environments
Plans and Pricing
Cowork usage limits doubled through August 5. Copilot for Government: $60/seat/month or $1/month for qualifying agencies. No consumer pricing changes this week.
Best for: Enterprises maximising Cowork tasks before August 5; government teams requesting FedRAMP High access