Dateline: July 19, 2026 | Next update: July 26, 2026
A week with no new model launch but five significant product and platform moves. The headline is Claude for Teachers — free premium access for every verified US K–12 educator through June 2027, with live state standards grounding via the Chan Zuckerberg Initiative's Learning Commons and FERPA-aligned data terms co-developed with the AFT. The Admin API gained user-management endpoints in beta, letting enterprise IT teams manage org members programmatically for the first time. Claude Code shipped a focused workflow and safety update: /fork branches a conversation to a background session while you keep working in the foreground, a new WebSearch cap prevents runaway search loops in long-horizon agents, and a Unicode normalisation fix closes a permission-preview security issue. Claude chat gained the ability to search and recall past conversations. The Claude Code 50% usage-limit promotion ends tonight at 11:59 PM PT. And Opus 4.7 fast mode is five days from a hard removal — any pipeline still calling it after July 24 will break with no silent fallback. On the OpenAI side, the ChatGPT desktop app was redesigned around Chat, Work, and Codex; unified conversation search launched globally across all plans; custom instructions grew from 1,500 to 5,000 characters; and workspace-scoped Admin keys arrived for Enterprise and Education. Gemini shipped a Workspace optimisation for large PDF batches in Docs and a 15% TTFT improvement for long-context API requests. ⚠ gemini-experimental-1125 API routes are hard-deprecated on July 25 — no fallback.
Claude / Anthropic
✅ Claude for Teachers — free premium access for US K–12 educators
Claude for Teachers launched on July 14, giving verified US K–12 educators a full year of free premium Claude access — the same feature set enterprise customers pay for, including Claude Code and Cowork. The product is built around a specific pedagogical problem: AI tools can generate lesson plans, but they are not grounded in what a specific state actually requires students to know. Anthropic's answer is Learning Commons, a live database of academic standards across all 50 states developed by the Chan Zuckerberg Initiative, which Claude draws on directly rather than relying on training-data memory.
Free for verified US K–12 educators through June 30, 2027 — sign up at claude.com/solutions/teachers. What's included: full premium Claude capabilities; Learning Commons connector (live database of 150,000+ academic standards from all 50 states plus sub-skills, grounding lesson plans in state-specific requirements rather than model memory); a library of teaching skills co-developed with Learning Commons and evaluated for pedagogical rigor; access to evidence-based curricula including OpenSciEd and Illustrative Mathematics (IM v.360); Claude Code and Claude Cowork for agentic tasks (example: hand off exit-ticket review once, Claude reviews student outcomes and adapts the next day's plan automatically at 4 PM every school day); FERPA-aligned data terms co-developed with the American Federation of Teachers (AFT); nine edtech connectors; AI fluency training for teachers. Currently for individual educators only — a dedicated school and district offering is in development.
Access: free for verified US K–12 educators | Duration: through June 30, 2027 | Learning Commons: live 150,000+ standard database, all 50 states, Chan Zuckerberg Initiative | Curricula: OpenSciEd, Illustrative Mathematics IM v.360 | Privacy: FERPA-aligned terms (co-developed with AFT; verify district-level terms with your own counsel) | Agentic tools: Claude Code + Cowork included | Connectors: 9 edtech integrations | Scope: individual educators only; school/district offering in development | Competitor context: OpenAI also offers ChatGPT for Teachers on comparable terms through June 2027
Best for: US K–12 educators and edtech developers building on Claude — verify and sign up at claude.com/solutions/teachers
✅ Admin API — user management endpoints in beta
Enterprise IT teams can now manage Claude organisation members programmatically via the Admin API — a capability that has been a top request from large enterprise deployments since the platform launched. Previously, member and group management required navigating the Claude Console UI; the API now enables it at scale through existing identity management and provisioning pipelines.
New Admin API user-management endpoints available in beta for all Claude Enterprise organisations. Capabilities: list members and look them up by email address; change a member's role; remove members; send and withdraw invites; manage groups and their membership; read custom roles. Member and invite endpoints require no beta header. Group and custom-role endpoints require the anthropic-beta: ce-user-management-2026-07-13 header. An Admin API key with the read:org_audit scope can call every user-management GET endpoint. Full documentation at the User management reference in Claude Docs.
Beta header for groups + custom roles: anthropic-beta: ce-user-management-2026-07-13 | No header required for: list/lookup members, change role, remove members, send/withdraw invites | Scope for read-only access: read:org_audit on Admin API key | Plans: Claude Enterprise (claude.ai) | Documentation: User management in Claude Docs
Best for: Enterprise IT and identity teams managing large Claude deployments programmatically — SCIM-provisioned users can now also manage groups and roles via API
Claude Code — /fork to background, WebSearch cap, /subtask
This week's Claude Code update focused on workflow control and safety guardrails rather than new headline features. The most useful change for power users is /fork: you can now branch a conversation into a background session and keep working in the foreground simultaneously. The most important safety change is a cap on WebSearch calls per session, preventing runaway search loops in long-horizon agentic workflows.
/fork now copies your current conversation into a new background session (its own row in claude agents) while you continue working in the foreground — the in-session subagent /fork used to launch is now renamed /subtask for clarity. Session-wide WebSearch limit added: default cap of 200 web searches per session, tunable via CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION env var, to stop runaway search loops in long-horizon agents. Auto-mode reset command: claude auto-mode reset restores the default auto-mode configuration with a confirmation prompt (pass --yes to skip). Auto-mode safety: new rule blocks tampering with session transcript files. /doctor now surfaces a complete setup checkup with actionable guidance. Fixed: startup hang when the Claude in Chrome extension is enabled but Chrome is not running. Fixed: 300ms delay revealing Settings tabs, Stats, and diff views. Fixed: reopening a just-stopped background session from agents view starting a blank conversation under the same session ID. Fixed: /loop hiding the session from /resume after a single use. Fixed: background jobs on LLM gateway auth (ANTHROPIC_AUTH_TOKEN + ANTHROPIC_BASE_URL) returning 'Not logged in' after daemon respawn. Fixed: claude agents jobs becoming permanently undeletable when git no longer recognises their worktree.
/fork: branches conversation to background session | /subtask: replaces old in-session /fork subagent | CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION: default 200, tunable | claude auto-mode reset: --yes flag to skip confirmation | Auto-mode: transcript tamper-block rule added | Fixed: Chrome extension startup hang | Fixed: 300ms async content reveal delay | Fixed: background session blank-conversation bug on reopen | Fixed: /loop hiding from /resume | Fixed: LLM gateway 'Not logged in' after daemon respawn | Fixed: permanently undeletable worktree-orphaned claude agents rows
Best for: Developers running parallel agentic sessions (/fork), long-horizon web research agents (WebSearch cap), and teams on LLM gateway auth (daemon respawn fix)
Claude Code — subagent streaming and security update
A second point release this week added subagent text streaming to external outputs, fixed a permission-preview security issue, and resolved a cluster of Bedrock, Vertex, Chrome, Windows, and hook reliability issues.
--forward-subagent-text flag and CLAUDE_CODE_FORWARD_SUBAGENT_TEXT env var: include subagent text and thinking in stream-json output for external consumers. Permission-preview security fix: bidirectional-override, zero-width, and look-alike quote characters are now neutralised in permission previews relayed to chat channels, preventing tool inputs from visually altering the approval message. Auto-mode fix: PreToolUse hook's ask decision now correctly floors auto-mode — a hook returning ask can no longer be overridden by auto-mode for unsandboxed Bash. Fixed: screen reader users losing the audible terminal bell after /terminal-setup or onboarding setup. Fixed: broad reliability improvements to Bedrock, Vertex, Chrome, and Windows session handling.
--forward-subagent-text / CLAUDE_CODE_FORWARD_SUBAGENT_TEXT: subagent text + thinking in stream-json | Security fix: Unicode normalisation on permission previews (bidi-override, zero-width, look-alike quotes neutralised) | Auto-mode: PreToolUse hook ask now floors decision (cannot be overridden for unsandboxed Bash) | Fixed: screen reader terminal bell after /terminal-setup | 'Claude Browser' and 'Claude Preview' MCP server names remain reserved ahead of upcoming Claude Desktop pane rename
Best for: Teams using stream-json output with subagents; security-conscious teams running auto-mode with unsandboxed Bash; screen reader users
Claude chat — search across past conversations
Claude gained the ability to search and build on previous conversations this week. Users can now reference earlier chats from within a session — Claude can look up what was discussed before and carry context forward. This is distinct from the existing Memory feature (which stores explicit facts across sessions): chat search retrieves actual prior conversation content on demand.
Claude can now search across your previous conversations and use that context in the current session — reference what you discussed in an earlier chat without re-explaining it. Claude uses memory to build on previous context automatically where relevant. Rolling out to all plans on claude.ai web and Claude Desktop. Cowork conversation search is coming soon.
Feature: conversational search across past sessions | Platform: claude.ai web, Claude Desktop | Rollout: all plans (gradual) | Distinct from: Memory (explicit stored facts); this retrieves prior conversation content on demand | Cowork search: coming soon
Best for: All Claude users who return to ongoing topics, projects, or questions across multiple conversations
⚠ Claude Code 50% usage-limit promotion — ends tonight
The Claude Code 50% weekly usage-limit promotion was extended through tonight and expires at 11:59 PM PT. After this point, limits revert to standard levels with no changes to plans or billing. Free plans and consumption-based Enterprise seats are not eligible. No action is required — the increase is automatic for eligible plans.
50% Claude Code weekly usage limit promotion ends tonight, July 19 at 11:59 PM PT. Eligible plans: Pro, Max, Team, legacy seat-based Enterprise. After tonight, limits return to standard. Consumption-based Enterprise and Free plans are not eligible.
Best for: Run your heaviest Claude Code sessions today before 11:59 PM PT if you are on an eligible plan
⚠ Opus 4.7 fast mode — removed July 24 (5 days away)
Opus 4.7 fast mode was deprecated on June 25 and will be removed on July 24 — five days from today. Unlike Opus 4.6 (which silently falls back to standard speed), Opus 4.7 will return a hard error. Any pipeline calling claude-opus-4-7 with speed: 'fast' after July 24 will break. Migrating to Opus 4.8 also brings a significant cost reduction: $10/$50 per MTok versus $30/$150 on Opus 4.7.
July 24 is five days away. Search your codebase now for claude-opus-4-7 combined with speed or fast. Change the model ID to claude-opus-4-8. Opus 4.8 fast mode: $10/$50 per MTok — 3× cheaper than Opus 4.7 fast mode ($30/$150). After July 24, claude-opus-4-7 + speed: 'fast' returns a hard error (no silent fallback). Next deadline after this: Opus 4.1 retirement, August 5, 2026.
Removal: July 24, 2026 | Affected: claude-opus-4-7 with speed: 'fast' | Error type: hard error (not silent fallback) | Migration: change model to claude-opus-4-8, keep speed: 'fast' and beta header fast-mode-2026-02-01 | Cost savings: $10/$50 per MTok (Opus 4.8 fast) vs $30/$150 (Opus 4.7 fast) | Opus 4.6 fast mode: already removed June 29 (silently falls back, no error) | Next retirement: Opus 4.1 (claude-opus-4-1-20250805), August 5, 2026
Best for: API developers and Claude Code users with fast-mode Opus workloads — immediate action required
Plans and Pricing
No pricing changes this week. Key consumption notes: Claude for Teachers is free for verified US K–12 educators through June 2027. Fable 5 on subscription plans moved to usage-credits-only after July 12. Sonnet 5 introductory pricing ($2/$10 per MTok) continues through August 31. Cowork usage limits doubled through August 5.
Claude for Teachers: free through June 30, 2027 for verified US K–12 educators | Fable 5 subscription: usage credits only from July 13 | Fable 5 API: $10/$50 per MTok, Batch $5/$25 | Sonnet 5: $2/$10 per MTok through August 31 (standard $3/$15 from September 1) | Opus 4.8: $5/$25 per MTok | Opus 4.7 fast mode: $30/$150 per MTok — migrating to Opus 4.8 saves 67% | Cowork limits: 2× through August 5 | Claude Code 50% limit promo: ends tonight July 19 | Next retirement: Opus 4.1, August 5
Best for: Educators: sign up at claude.com/solutions/teachers today. API developers: migrate Opus 4.7 fast mode to Opus 4.8 before July 24. Use Sonnet 5 at $2/$10 introductory rate through August 31.
ChatGPT / OpenAI
Dateline: July 19, 2026 | Next update: July 26, 2026
A week focused on product consolidation, enterprise administration, and cross-device continuity rather than a new model launch. OpenAI redesigned the ChatGPT desktop app around three increasingly distinct modes — Chat, Work, and Codex — while allowing cloud Work conversations and Projects to move across web, mobile, and desktop. ChatGPT gained unified search across conversations, Projects, images, and files for all plans globally, and paid users received a much larger custom-instructions allowance (1,500 → 5,000 characters). Enterprise and Education workspaces gained workspace-scoped Admin keys, expanded analytics history, and support for synchronised apps inside Enterprise Key Management environments. ChatGPT returned to WhatsApp in the European Economic Area. Codex on iOS added inline visualisations and stronger task controls. Codex CLI releases concentrated on safety and reliability, including broader detection of destructive terminal commands and corrected GPT-5.6 model metadata. No new model or API pricing change was announced during the reporting period.
✅ ChatGPT desktop app — unified Chat, Work, and Codex
OpenAI updated the ChatGPT desktop app to clarify the relationship between its conversational, agentic, and coding products. A new global switcher lets users move between ChatGPT and Codex, while the ChatGPT side is divided into Chat for normal conversational work and Work for longer tasks completed end to end. This is primarily a product-architecture change: OpenAI is making the desktop app the central surface for both immediate assistance and longer-running delegated work.
The desktop app now combines Chat and Work conversations in one Recents view, with support for sorting, filtering, and pinning. Existing Projects appear directly inside the app, and users can begin either a standard conversation or a Work thread using a Project's context. Cloud Work conversations now synchronise across desktop, web, and mobile — a user can start a long-running task on one device and continue steering it from another. Local conversations remain stored on the computer and do not receive this cross-device continuity. Codex remains a separate view, with its existing workflows and history unchanged.
Platforms: macOS, Windows | Plans: all ChatGPT plans | Navigation: global ChatGPT/Codex switcher | ChatGPT modes: Chat and Work | Recents: unified Chat + Work history with sorting, filtering, and pinning | Projects: available directly in the desktop app | Cloud Work: synchronised across web, mobile, and desktop | Local conversations: remain device-local | Codex: separate view; workflows and history unchanged
Best for: Users adopting ChatGPT Work for multi-stage research, analysis, or artifact creation, and teams that need to start a task on desktop and continue supervising it from another device
✅ Unified search across chats, Projects, images, and files
ChatGPT now provides one search experience across a user's conversations, Projects, generated or uploaded images, and documents. Previously, users often had to remember whether information lived in a conversation, Project, or file. The new sidebar search treats these as one retrievable personal information layer.
Users can search broadly or filter by content type, then open the relevant chat, Project, image, or file directly from the result. Available globally on every ChatGPT plan — Free, Go, Plus, Pro, Business, Enterprise, and Education — rather than being restricted to paid or managed workspaces. Strengthens the practical value of Projects and the user file library by making older material easier to locate.
Platforms: ChatGPT web, iOS, and Android | Plans: all (Free, Go, Plus, Pro, Business, Enterprise, Education) | Indexed surfaces: chats, Projects, images, and documents | Entry point: ChatGPT sidebar | Controls: content-type filtering | Result behaviour: opens the original chat, Project, image, or file
Best for: Researchers, consultants, writers, and teams with substantial ChatGPT histories who need to recover previous work without navigating individual Projects or conversations manually
✅ Custom instructions limit raised to 5,000 characters
OpenAI increased the custom-instructions limit from 1,500 to 5,000 characters for most paid and managed ChatGPT plans. Custom instructions allow users to define persistent preferences around tone, structure, working methods, background context, and response behaviour.
Plus, Pro, Business, Enterprise, and Education users can now save up to 5,000 characters. Users can provide fuller style guides, organisational terminology, recurring constraints, and detailed rules without compressing them into a short prompt — reducing repeated prompting in professional workflows. Free and Go plans are not listed as receiving the expanded allowance.
Previous limit: 1,500 characters | New limit: 5,000 characters | Eligible plans: Plus, Pro, Business, Enterprise, and Education | Use cases: persistent response style, domain context, formatting rules, behavioural preferences, workflow guidance | No model change attached to this release
Best for: Power users, organisations with established house styles, researchers with recurring methodological requirements, and consultants who need ChatGPT to follow detailed default operating instructions
✅ Enterprise Admin keys and expanded analytics
Administrators can now create workspace-scoped Admin keys inside the Global Admin Console, usable with supported ChatGPT and Codex administration APIs — reducing reliance on manual console workflows for large deployments. Admin keys support group management, spend controls, cost reporting, and analytics. Permissions depend on the administrator's workspace role; keys cannot be used for model inference.
Workspace-scoped Admin keys are now available in the Global Admin Console under Credentials → Admin keys. The Global Admin Console now displays up to 120 days of credit and Codex analytics history. Spend Controls and the Usage Limits page remain in ChatGPT Workspace settings. Administrators can continue to set usage limits, including a No limit option where available. Organisations can automate governance and reporting without creating a credential that can also submit prompts or consume model tokens.
Plans: ChatGPT Enterprise and Education | Key type: workspace-scoped Admin keys | Creation path: Global Admin Console → Credentials → Admin keys | Supported functions: group management, Spend Controls API, cost reporting, analytics | Inference: not supported | Permission model: based on workspace admin role | Analytics history: up to 120 days for credit and Codex usage | Usage-limit configuration: remains under Workspace settings
Best for: Enterprise IT, identity, FinOps, and AI-governance teams managing multiple workspaces, monitoring Codex consumption, or automating administrative reporting
✅ Synced apps now available in Enterprise Key Management workspaces
Apps with synchronisation are now available in Enterprise and Education workspaces that use OpenAI's Enterprise Key Management configuration. These apps can pre-index permitted external information so ChatGPT can retrieve relevant material more efficiently during conversations and research tasks.
Enterprise and Education administrators no longer have to choose between EKM-enabled workspace security and the productivity benefits of synchronised apps. The update applies across all apps that support sync rather than to one specific connector. Existing app permissions, role controls, and workspace governance remain relevant when deciding which integrations to enable.
Plans: Enterprise and Education | Requirement: Enterprise Key Management enabled | Feature: apps with sync | Scope: all currently supported synchronised apps | Purpose: pre-index authorised external content for faster contextual retrieval | Governance: existing app permissions and workspace controls continue to apply
Best for: Regulated enterprises, universities, and public institutions that require customer-controlled encryption but also want searchable, synchronised organisational data inside ChatGPT
✅ ChatGPT returns to WhatsApp in the EEA
ChatGPT became available again through WhatsApp in the European Economic Area. Users can start without creating or linking a ChatGPT account by messaging OpenAI's verified 1-800-CHATGPT contact. The WhatsApp experience supports text, image uploads, voice notes, image creation, and multiple languages. Linking an existing ChatGPT account is optional but provides higher usage limits.
Restores a low-friction ChatGPT access route in the EEA and makes the service available to users who may not want to install a separate application or create an account immediately. OpenAI also lists Kakao in South Korea and Viber in supported markets as other messaging-platform access points. Availability is determined by the country code associated with the user's WhatsApp number.
Platform: WhatsApp | Region: European Economic Area | Account requirement: none for initial access | Optional linking: ChatGPT account for higher limits | Inputs: text, images, voice notes | Outputs: text and image generation | Rollout: gradual; usage limits apply
Best for: EEA users seeking lightweight mobile access, multilingual conversations, image assistance, and voice-note interaction without initially creating a ChatGPT account
✅ Codex on iOS — inline visualisations and stronger task controls
Codex tasks on iOS can now display visualisations inline, reducing the need to leave the task view to understand chart, diagram, or structured visual output results. Task creation and management from ordinary conversations were also improved, with more reliable links to newly created Codex tasks and refined tool activity and progress indicators.
Inline visualisations in Codex tasks. Improved task creation, management, and linking from conversations. Composer controls remain visible above the keyboard during long prompts or when larger accessibility text settings are used. Fixed: Fast mode selection and restoration on a per-task basis; initial prompts ignoring the selected approval preset; interface elements becoming unresponsive during swipe gestures.
Version: ChatGPT for iOS 1.2026.188 | New capability: inline visualisations in Codex tasks | Task controls: improved creation, management, and linking from conversations | Accessibility: composer controls remain visible with long prompts and larger text sizes | Fixed: Fast mode persistence, approval-preset handling, autocomplete backgrounds, unresponsive task rows
Best for: Developers and analysts who monitor Codex work from mobile, particularly when tasks produce charts or require approvals and follow-up instructions away from a desktop
✅ Codex CLI 0.144.6 — destructive-command safety and GPT-5.6 metadata
Version 0.144.5 expanded detection of dangerous commands — including additional forced forms of rm — and improved the explanation shown when a command is rejected. Version 0.144.6 refreshed bundled instructions for GPT-5.6 Sol, Terra, and Luna and corrected their listed context windows to 272,000 tokens. Version 0.144.2 restored the previous Guardian auto-review policy after OpenAI rolled back a prompting regression.
Update to 0.144.6: npm install -g @openai/codex@0.144.6. GPT-5.6 Sol, Terra, and Luna context windows corrected to 272,000 tokens in bundled metadata. Version 0.144.5: expanded destructive-command detection including forced rm variants, with clearer denial reasons. Version 0.144.2: Guardian auto-review prompting regression rolled back. Versions 0.144.3–0.144.4: no substantive user-facing additions.
Latest version: 0.144.6 | Install: npm install -g @openai/codex@0.144.6 | GPT-5.6 metadata: Sol, Terra, and Luna context windows corrected to 272,000 tokens | Version 0.144.5: expanded destructive-command detection, forced rm variants, clearer denial reasons | Version 0.144.2: Guardian auto-review prompting regression rolled back | Versions 0.144.3–0.144.4: no substantive user-facing additions
Best for: Codex CLI users, security teams, and developers using autonomous or approval-based terminal execution — update to 0.144.6 for the complete set of fixes
⚠ Enterprise desktop app — brief sign-in errors (resolved)
Some Enterprise users who did not have Codex role-based access enabled encountered a codex_cli_workspace_disabled error when signing in to the new ChatGPT desktop app. OpenAI temporarily advised affected customers to use ChatGPT Classic while a fix was prepared. The issue briefly recurred on July 18 before OpenAI reported full recovery.
Resolution: update to the latest version of the ChatGPT desktop application. Enterprise desktop-app administrators should ensure managed deployments distribute the corrected application version before broad migration from ChatGPT Classic.
Affected population: Enterprise users without Codex RBAC enabled | Error: codex_cli_workspace_disabled | Workaround during incident: ChatGPT Classic | Resolution: update to latest desktop application | Final status: fully recovered July 18
Best for: Enterprise desktop-app administrators — verify managed deployments are on the corrected version before migrating users from Classic
⚠ GPT-5.6 Sol — brief Codex capacity errors (resolved)
Codex users experienced elevated server-overload errors when using GPT-5.6 Sol on July 17. OpenAI applied a mitigation and reported full recovery at 8:42 PM that day. The incident does not indicate a model retirement or continuing availability restriction — it reflects near-term capacity pressure that can accompany a newly launched flagship model across agentic coding workloads.
No migration action is currently required. Teams tracking reliability of newly released models should note this was a transient capacity event, now fully resolved.
Plans and Pricing
No new model, consumer-plan, or API pricing changes between July 11 and July 19. GPT-5.6 pricing remains as introduced the previous week: Sol at $5/$30 per MTok, Terra at $2.50/$15, Luna at $1/$6. The most significant operational changes concern enterprise administration and access. Enterprise and Education customers gained Admin keys and up to 120 days of credit and Codex analytics. Synchronised apps became available in EKM-enabled workspaces. Paid users received larger custom-instruction capacity; unified content search launched across every plan.
New pricing changes this week: none | New model launch: none | GPT-5.6 Sol: $5/$30 per MTok | GPT-5.6 Terra: $2.50/$15 | GPT-5.6 Luna: $1/$6 | Custom instructions: 5,000 characters for Plus, Pro, Business, Enterprise, and Education | Unified search: all plans | Enterprise analytics: up to 120 days | Codex CLI: update to 0.144.6
Best for: No pricing migration required. Enterprise administrators: create narrowly scoped Admin keys, review synchronised-app permissions in EKM workspaces, and update managed desktop deployments. Codex CLI users: move to version 0.144.6.
Gemini (Google)
Dateline: July 19, 2026 | Next update: July 26, 2026
A quiet week for new model launches. Google focused on internal platform optimisations and stability improvements across Workspace integrations. The main engineering work centred on refining context-window retrieval efficiency for long-context requests and preparing for upcoming infrastructure changes. ⚠ Legacy gemini-experimental-1125 API routes are deprecated on July 25 — a hard 404 with no fallback.
✅ Gemini Workspace — large PDF batch optimisation
Google rolled out an optimisation patch for Gemini across Google Docs and Drive. The update resolves latency issues when compiling large multi-source summaries, and processing speed for files over 50 MB within attached Drive folders has improved. Gemini can now synthesise information from up to 20 PDFs concurrently within the Google Docs sidebar without triggering timeout errors.
Enhanced data-handling for large file batches in Google Drive — up to 20 PDFs concurrently in the Google Docs sidebar without timeout. Fixed: bullet-point structures losing indentation when exporting from Gemini drafts into active Docs. Parallel processing pipeline for PDF batch reads.
Access: Workspace Enterprise and Labs tiers | Optimisation: parallel processing pipeline for PDF batch reads | Max concurrent file evaluation: 20 active pointers | Bug fix: Markdown-to-AST layout fix for indentation retention on export
Best for: Enterprise users managing heavy research documentation workflows inside Google Workspace
✅ API — context window retrieval improvements
Google pushed a mid-cycle update to the RAG framework for Gemini models in AI Studio, targeting performance consistency at the outer edges of the 2M+ token context window. The update reduces first-token latency by up to 15% on requests exceeding 1 million tokens and stabilises memory footprint per concurrent session, preventing unexpected 503 errors during peak API hours.
Optimised KV-caching mechanisms to reduce first-token latency (TTFT) by up to 15% on requests exceeding 1 million tokens. Memory footprint per concurrent session stabilised, preventing unexpected 503 service unavailable errors during peak API hours.
Framework update: dynamic context routing and localised attention weighting | Latency reduction: ~15% improvement on TTFT for requests exceeding 1M tokens | Errors addressed: 503 transient dropouts under high concurrency | Platform: Vertex AI and Google AI Studio
Best for: Developers building long-form document analysis tools or complex agentic pipelines requiring sustained context history
⚠ gemini-experimental-1125 deprecated July 25
Google issued an advisory for an upcoming two-hour infrastructure maintenance window. Legacy preview endpoints designated from late 2025 will be permanently deprecated following this window. Any service calls hitting these legacy paths after the maintenance window will return a 404 error — there is no fallback to standard models.
July 25 brings a hard deprecation of all gemini-experimental-1125 API routes. Any service calls hitting these legacy paths after the maintenance window will return a 404 error instead of falling back to standard models. Update your environment configurations immediately. Point production pipelines to gemini-1.5-pro or gemini-1.5-flash production tags.
Maintenance window: July 25, 2026 at 02:00–04:00 UTC | Deprecated routes: gemini-experimental-1125 paths | Error type: hard 404 error | Action required: point production pipelines to gemini-1.5-pro or gemini-1.5-flash production tags | Competitor context: Anthropic has a hard error deadline on July 24 for Opus 4.7 fast mode removal
Best for: DevOps engineers and platform administrators managing production-level API implementations — update gemini-experimental-1125 routing before July 25
Plans and Pricing
No pricing changes for Gemini API tiers this week. Standard billing remains consistent across Flash and Pro models. Deprecation map: gemini-experimental-1125 is deprecated July 25. Gemini 2.5 family deprecation remains confirmed for October 16, 2026. Gemini 3.5 Flash introductory pricing ($2/$10 per MTok) remains active through August 31.
Gemini 3.5 Flash: $2/$10 per MTok through August 31 | gemini-experimental-1125: deprecated July 25 (hard 404) | Gemini 2.5 family: deprecated October 16, 2026 | No model pricing changes this week
Best for: API developers running on gemini-experimental-1125 — migrate now. Take advantage of Gemini 3.5 Flash introductory pricing through August 31.
Microsoft Copilot
Dateline: July 19, 2026 | Next update: July 26, 2026
Microsoft Copilot's biggest moves this week mirrored Anthropic's: Copilot for Educators launched July 14 with free premium access for US K–12 teachers through June 2027, grounded in live state standards via Learning Commons. The Admin API gained user-management endpoints. Copilot Code shipped workflow and safety updates (/fork, /subtask, WebSearch cap, permission-preview security fix). Copilot Chat gained search across past conversations. The 50% usage-limit promotion ended July 19. ⚠ Opus 4.7 fast mode is removed July 24 — migrate to Opus 4.8 now.
✅ Copilot for Educators — free premium access for US K–12 teachers
Copilot for Educators launched July 14, giving verified US K–12 teachers a full year of free premium Copilot access, including Copilot Code and Copilot Cowork, plus live grounding in state standards via the Learning Commons connector.
Free premium Copilot for verified US K–12 teachers through June 2027. Includes: Learning Commons connector (live database of 150,000+ academic standards across all 50 states); teaching skills library co-developed with Learning Commons; evidence-based curricula (OpenSciEd, Illustrative Mathematics IM v.360); FERPA-aligned data terms co-developed with the American Federation of Teachers; Copilot Code and Copilot Cowork for agentic tasks. Individual educators only — school and district offering in development.
Access: free through June 30, 2027 | Learning Commons: live standards database, 150,000+ standards, all 50 states | Curricula: OpenSciEd, IM v.360 | Privacy: FERPA-aligned | Tools: Copilot Code + Cowork included | Scope: individual educators only; school/district offering in development
Best for: US K–12 educators and edtech developers building on Copilot
✅ Admin API — user management endpoints in beta
Enterprise IT teams can now manage Copilot org members programmatically via the Admin API. Endpoints for listing members, role changes, removals, invites, group management, and custom roles. An Admin API key with the read:org_audit scope provides read-only access.
Beta header required for groups + custom roles | No header required for member/invite endpoints | Scope for read-only access: read:org_audit | Documentation: Copilot Docs
Best for: Enterprise IT and identity teams managing large Copilot deployments programmatically
Copilot Code — /fork, WebSearch cap, and safety updates
Copilot Code shipped workflow control and safety guardrails this week: /fork branches a conversation to a background session while you keep working in the foreground; /subtask replaces the old in-session /fork; a 200-search-per-session WebSearch cap prevents runaway search loops in long-horizon agents; a permission-preview Unicode fix closes a security issue. Fixes: Chrome extension startup hang, background session blank-conversation bug, LLM gateway auth issues, permanently undeletable orphaned agents.
/fork: branches conversation to background session | /subtask: replaces old in-session /fork | CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION: default 200 | Security fix: Unicode normalisation on permission previews | Fixed: Chrome extension startup hang, background session blank-conversation, LLM gateway daemon respawn, undeletable worktree-orphaned rows
Best for: Developers running parallel agentic sessions and long-horizon research agents
Copilot Code — subagent streaming and reliability
Forward-subagent-text flag and env var for including subagent text and thinking in stream-json output. Auto-mode fix: PreToolUse hook ask now correctly floors auto-mode — cannot be overridden for unsandboxed Bash. Broad reliability improvements for Bedrock, Vertex, Chrome, and Windows session handling.
--forward-subagent-text / CLAUDE_CODE_FORWARD_SUBAGENT_TEXT: subagent text + thinking in stream-json | Auto-mode: PreToolUse hook ask floors decision | Fixed: screen reader terminal bell, Bedrock/Vertex/Chrome/Windows reliability
Best for: Teams using stream-json output with subagents; security-conscious teams on auto-mode with unsandboxed Bash
Copilot Chat — search across past conversations
Copilot can now search across past conversations and use prior context in the current session. Distinct from Memory (explicit stored facts) — this retrieves actual prior conversation content on demand. Cowork conversation search is coming soon.
Best for: All Copilot users who revisit ongoing topics or projects across multiple sessions
⚠ Copilot Code 50% usage-limit promotion — ended July 19
The 50% Claude Code weekly usage-limit promotion ended tonight at 11:59 PM PT. Eligible plans: Pro, Enterprise, Team. Limits have reverted to standard.
⚠ Opus 4.7 fast mode — removal July 24
Opus 4.7 fast mode will be removed July 24. Pipelines calling claude-opus-4-7 with speed: 'fast' will return a hard error after that date — no silent fallback. Migrate to Opus 4.8 fast mode ($10/$50 per MTok) versus $30/$150 on Opus 4.7. Next deadline: Opus 4.1 retirement, August 5, 2026.
Best for: API developers and Copilot Code users with fast-mode workloads — immediate migration required
Plans and Pricing
No pricing changes this week. Copilot for Educators free through June 2027. Cowork usage limits doubled through August 5. Opus 4.7 fast mode removed July 24. Opus 4.1 retired August 5.
Best for: Educators: sign up for free premium Copilot. Developers: migrate Opus 4.7 fast mode before July 24.