Dateline: July 24, 2026 | Next update: July 31, 2026
A shorter five-day window with no model launch, but five meaningful updates across voice, developer tooling, and platform infrastructure. The headline is a major Claude voice mode overhaul — new speech models, richer intonation, faster response, expanded language support, and customisable voices — the largest voice update since the feature launched in May 2025. Alongside it: Claude Managed Agents gained effort controls, expanded webhooks covering environment and memory store lifecycle events, and session seeding via an initial events array; the agent-memory-2026-07-22 API header landed with coordinated SDK updates across all eight official Anthropic SDKs; API key expiration settings arrived in the Console; and the Anthropic Economic Index connector launched, letting anyone query Anthropic's labour-market data directly in Claude. Claude Code shipped subagent depth controls (default: no nested spawning), a corrected budget cap that halts running background agents when the limit is reached, a filesystem isolation toggle, and a fix for a quadratic message-normalisation slowdown that caused multi-second stalls in long sessions. ⚠ Opus 4.7 fast mode is removed as of today — any call using claude-opus-4-7 with speed: 'fast' returns a hard error with no fallback. On the other platforms: OpenAI launched Health in ChatGPT, extended ChatGPT Voice to Work and Codex on desktop, and added organisation and project spend limits to the API. Gemini shipped a parallel voice overhaul, thinking effort controls for agentic workflows, API key expiration, and Gemini Code Assist depth and performance fixes.
Claude / Anthropic
✅ Claude voice mode — major overhaul with new speech models
Anthropic shipped the largest voice mode update since the feature launched in May 2025. The overhaul is built on new underlying speech synthesis and recognition models, delivering meaningfully more natural intonation, lower latency, and broader language support. This brings Claude's voice experience substantially closer to the quality bar that competitors have been setting with dedicated voice products.
New speech models underpin the entire voice overhaul. Natural intonation: voice responses now adapt tone and cadence to conversational context rather than delivering flat, uniform speech. Faster response latency: reduced time from end of user speech to start of Claude's reply. Expanded language support: languages beyond English now available in voice mode — see the Claude app settings for the full list. Customisable voices: choose from multiple voice characters and adapt them to different use-case scenarios (customer support, education, casual conversation). Improved context retention in long voice sessions: Claude better tracks what has been said across a multi-turn voice session. More accurate speech recognition: lower error rate on complex queries and accented speech. Noise suppression support. Stronger safety and content controls during voice interactions. The new voice engine is designed for integration across different deployment environments, including customer support platforms and educational applications. Claude Code voice (/voice command): separate feature, rolling out to ~5% of Claude Code users since March 2026 and expanding.
Platform: Claude iOS and Android apps | Rollout: all plans, gradual | Voice toggle: tap microphone in the Claude app | Underlying change: new speech synthesis and recognition models | Languages: expanded beyond English, see app settings for full list | Customisation: multiple voice characters, scenario adaptation | Safety: enhanced content controls during voice interactions | Claude Code voice (/voice command): separate feature, ~5% of Claude Code users, expanding
Best for: Mobile users, customer-support teams building voice integrations, education platforms, and anyone who uses Claude hands-free
✅ Claude Managed Agents — effort controls, lifecycle webhooks, session seeding
Claude Managed Agents gained three capabilities this week that collectively give developers much finer control over agent behaviour and observability. Effort controls let you set how hard an agent thinks per turn at the configuration level, without per-request changes. Expanded webhooks cover environment and memory store lifecycle events, eliminating the need to poll for state changes. Session seeding lets you pre-populate an agent's session with initial events before it starts working.
Effort controls: set an effort level on a Managed Agents agent's model configuration by passing effort inside the model object at agent creation. Levels: low, medium, high (default), xhigh. Controls the depth of adaptive thinking per turn without requiring per-request parameter changes. Expanded webhook coverage: four new environment.* event types and three new memory_store.* event types — react to environment and memory store lifecycle changes without polling. Session seeding with initial events: pre-populate an agent session with context (tool results, prior conversation turns, structured data) before the agent's first user turn, enabling warm-start workflows. Optional version checks on updates: pass an expected version when updating an agent; the API returns a 409 conflict if the agent has changed since you last read it, preventing lost-update races. Event deltas for thread streams: thread stream events now include delta fields showing only what changed between events, reducing parsing overhead for consumers that track diffs. Task tool mode parameter deprecated: the mode parameter on the Task tool is now ignored; subagents inherit the parent session's permission mode by default.
Effort: pass effort inside model object at agent creation (low / medium / high / xhigh) | Webhooks: 4 environment.* event types + 3 memory_store.* event types | Session seeding: initial_events array at session creation | Version check: pass expected_version on update; 409 on conflict | Event deltas: delta fields in thread stream events | Header: managed-agents-2026-04-01 | Task tool mode: deprecated, ignored; subagents inherit parent permission mode
Best for: Teams building production Managed Agents workflows needing cost-to-quality tuning (effort), real-time observability (webhooks), warm-start context (session seeding), or concurrent-update safety (version checks)
API — agent-memory-2026-07-22 header + all-SDK updates
A new beta header for Managed Agents memory stores landed this week, stabilising how memory listing behaves. All eight official Anthropic SDKs were updated simultaneously to send the new header by default — an unusually broad coordinated platform update. Existing page cursors issued without the header are not valid with it; restart pagination from the first page when adopting.
New beta header: agent-memory-2026-07-22. Changes to GET /v1/memory_stores/{id}/memories: results now return in a stable, server-defined order (order_by and order parameters are ignored); depth accepts only 0, 1, or omitted (other values return a 400 error); path_prefix must end with / and matches whole path segments rather than substrings. Conflict: sending both agent-memory-2026-07-22 and managed-agents-2026-04-01 returns a 400 error. From July 22, managed-agents-2026-04-01 adopts the same list behaviour automatically on memory endpoints. All SDKs updated: Python 0.116.0, TypeScript 0.110.0, Go 1.56.0, Java 2.48.0, Ruby 1.55.0, PHP 0.36.0, C# 12.35.0, CLI 1.16.0 now send agent-memory-2026-07-22 on all memory store calls by default.
New header: agent-memory-2026-07-22 | Endpoint: GET /v1/memory_stores/{memory_store_id}/memories | Breaking: depth values other than 0 or 1 now 400; path_prefix must end with /; order_by and order ignored | Conflict: sending both headers returns 400 | July 22 cutover: managed-agents-2026-04-01 adopts same behaviour on memory endpoints | SDK versions: Python 0.116.0, TS 0.110.0, Go 1.56.0, Java 2.48.0, Ruby 1.55.0, PHP 0.36.0, C# 12.35.0, CLI 1.16.0 | Migration: restart pagination from page one; replace managed-agents-2026-04-01 with agent-memory-2026-07-22 in any explicit beta overrides on memory store calls
Best for: All API developers using Managed Agents memory stores — update SDKs and replace the old header in any explicit beta overrides
✅ API — API key expiration settings in the Claude Console
Anthropic added API key expiration controls to the Claude Console — a security hygiene feature that has been a common request from enterprise security teams and developers following standard credential rotation policies. Previously, all API keys had indefinite lifetimes unless manually revoked.
API key expiration is now available when creating API keys or Admin API keys in the Claude Console. Options: preset durations, a custom duration, or Never (the previous default behaviour). For keys with a lifetime of at least 7 days, Anthropic emails the creator before the key expires. Existing keys are not affected — they retain indefinite lifetimes unless manually revoked. The Admin API reports each key's expiration in the expires_at field on the key object. Recommended practice: set an expiration aligned with your organisation's credential rotation policy on all new keys going forward.
Location: Claude Console → API keys (create new key) | Options: preset durations, custom duration, Never | Email reminder: sent before expiry for keys with lifetime ≥7 days | Existing keys: unaffected, still indefinite unless manually revoked | Admin API: expires_at field on key object | Applies to: API keys and Admin API keys
Best for: Enterprise security teams requiring credential rotation; developers wanting to enforce key lifetimes aligned with SOC 2 or ISO 27001 controls
✅ Anthropic Economic Index — connector for Claude
The Anthropic Economic Index — a dataset measuring how AI is actually being used across the labour market, published periodically by Anthropic's research team — is now directly queryable in Claude via a connector. Previously, the data was available only through research reports. Now anyone can ask Claude questions grounded directly in the Index data without reading the underlying reports first.
Anthropic Economic Index connector is live in the Claude connector directory. Enable in one click from the Connectors menu in any claude.ai conversation — works with any Claude model, no install required. Query the data conversationally: ask broad questions ('What does the Index say about my industry?') or drill into specifics, and ask Claude to surface the underlying data behind any answer. The Index covers how Claude is being used across occupations and tasks. The June 2026 edition added hourly telemetry, artifact classifiers, and a 9,700-person Economic Index Survey linking usage to workers' beliefs about AI and job security. International context: Canada country brief published (2.6% global share, second-highest per-capita adoption after US); India local pricing launched (Claude Pro ₹2,000/month annual).
Enable: Connectors menu in claude.ai | Works with: any Claude model, any conversation | Data: Anthropic Economic Index (labour-market AI usage data) | June 2026 edition: hourly telemetry, artifact classifiers, 9,700-person Economic Index Survey | No install required | Canada brief: 2.6% global share, second-highest per-capita adoption after US | India: Claude Pro ₹2,000/month annual pricing launched
Best for: Researchers, journalists, policymakers, HR and workforce teams wanting to ground AI-and-labour questions in primary Anthropic data rather than secondary reporting
Claude Code — subagent depth controls, filesystem isolation, quadratic slowdown fix
Claude Code shipped a focused stability and safety update this week. The most impactful change for users running long or complex sessions is the fix for a quadratic message-normalisation cost that was causing multi-second stalls and slow /resume on sessions with many turns. The most important safety addition is subagent depth control — subagents no longer spawn nested subagents by default.
Subagent spawn depth: subagents no longer spawn nested subagents by default — set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH to allow deeper nesting. This prevents accidental runaway agent trees and makes resource use predictable by default. Budget cap fix: --max-budget-usd now correctly stops background subagents when the cap is reached — new spawns are denied and running background agents are halted. Filesystem isolation control: new sandbox.filesystem.disabled setting skips filesystem isolation while keeping network egress control, for environments where filesystem sandboxing conflicts with tooling. Quadratic slowdown fix: fixed a bug where message normalisation cost grew quadratically with the number of turns, causing multi-second stalls and slow /resume in long sessions. Security fix: bidirectional-override and zero-width character neutralisation now also applied to /subtask inputs relayed to the agents view, closing the same permission-preview spoofing vector for subtasks that was fixed for tool inputs the previous week. Task tool mode parameter deprecated: the mode parameter on the Task tool is now ignored; subagents inherit the parent session's permission mode by default.
CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH: controls nested subagent depth (default: 0, no nesting) | sandbox.filesystem.disabled: skips filesystem isolation, keeps network egress control | --max-budget-usd: now halts running background agents when cap reached (new spawns denied) | Quadratic fix: message normalisation O(n²) → O(n) per turn | Security fix: Unicode normalisation extended to /subtask inputs in agents view | Task tool mode: deprecated, ignored; subagents inherit parent permission mode
Best for: Users with long Claude Code sessions experiencing stalls or slow /resume: update immediately. Teams running multi-depth agent trees: set CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH explicitly. Budget-conscious teams running background subagents: --max-budget-usd now works correctly.
⚠ Opus 4.7 fast mode — removed today (July 24)
Opus 4.7 fast mode has been removed as of today. Any API call or Claude Code session using claude-opus-4-7 with speed: 'fast' will now return a hard error. There is no silent fallback. The migration to Opus 4.8 fast mode also saves 67% on fast-mode costs — $10/$50 per MTok versus $30/$150 on Opus 4.7.
Opus 4.7 fast mode is removed as of July 24. Hard error: requests to claude-opus-4-7 with speed: 'fast' return an error immediately, with no fallback to standard speed. Action required: change model ID to claude-opus-4-8, keep speed: 'fast' and beta header fast-mode-2026-02-01. Cost benefit of migrating: Opus 4.8 fast mode is $10/$50 per MTok vs $30/$150 for Opus 4.7 fast mode — 67% cheaper. Next deadline: Opus 4.1 (claude-opus-4-1-20250805) retires August 5, 2026 — 12 days away. Check your codebase for claude-opus-4-1-20250805 now.
Removed: July 24, 2026 | Affected: claude-opus-4-7 with speed: 'fast' | Error type: hard error, no fallback | Migration: claude-opus-4-8 + speed: 'fast' + beta header fast-mode-2026-02-01 | Opus 4.8 fast mode pricing: $10/$50 per MTok | Opus 4.7 fast mode was: $30/$150 per MTok | Opus 4.6 fast mode: already removed June 29 | Next retirement: Opus 4.1 (claude-opus-4-1-20250805), August 5, 2026
Best for: Any team still on Opus 4.7 fast mode: fix the model string now. August 5 is 12 days away — also check for claude-opus-4-1-20250805 in your codebase.
Plans and Pricing
No new pricing changes this week. Key active windows: Sonnet 5 introductory pricing ($2/$10 per MTok) continues through August 31. Cowork usage limits doubled through August 5. Fable 5 on subscription plans requires usage credits. Claude for Teachers is free for verified US K–12 educators through June 2027.
Sonnet 5: $2/$10 per MTok through August 31, then $3/$15 | Fable 5 API: $10/$50 per MTok standard, $5/$25 Batch | Opus 4.8: $5/$25 per MTok | Opus 4.8 fast mode: $10/$50 per MTok (2.5× speed) | Opus 4.7 fast mode: removed today | Cowork limits: 2× through August 5 | Fable 5 subscription: usage credits only | Claude for Teachers: free through June 30, 2027 | Next retirement: Opus 4.1, August 5 — migrate to Opus 4.8
Best for: Check for claude-opus-4-1-20250805 in your codebase now — August 5 is 12 days away. Continue using Sonnet 5 at introductory rates through August 31.
ChatGPT / OpenAI
Dateline: July 24, 2026 | Next update: July 31, 2026
A shorter five-day window with no new frontier-model launch, but three meaningful product and platform releases. The headline is Health in ChatGPT, which is beginning to roll out to eligible adults in the United States and allows users to securely connect medical records and Apple Health data. ChatGPT Voice also expanded into Work and Codex on the desktop app, allowing users to start and coordinate longer agentic or coding tasks through natural speech. On the developer side, OpenAI introduced organisation- and project-level API spend limits, including enforceable hard caps that stop requests once a defined monthly budget is reached. Codex on iOS gained inline Mermaid diagrams, interactive forms, unsent-prompt recovery, and improved controls for blocked or usage-limited tasks. The week also included several resolved service disruptions affecting ChatGPT, image generation, Workspace Agent, and the gpt-image-2 API model. No new consumer subscription or model API pricing changes were announced.
✅ Health in ChatGPT — connected health records and Apple Health
OpenAI began rolling out a dedicated health experience inside ChatGPT. Eligible users can connect supported medical records and Apple Health data, view key health information in one dashboard, and ask questions grounded in their own records rather than relying only on details manually entered into a conversation. The initial rollout is limited to logged-in Free, Go, Plus, and Pro users in the United States who are at least 18 years old.
Connected medical records, Apple Health information, and conversations that use this data are not used to train OpenAI's foundation models or for targeted advertising. The Health experience can display laboratory results, medications, activity, sleep, and other supported health data. Users can ask ChatGPT to help understand test results, identify trends, prepare questions for a medical appointment, or track personal wellness goals. OpenAI describes the feature as a support tool rather than a substitute for professional medical care, diagnosis, or treatment. Users control which sources they connect and when ChatGPT may use that information. The experience is initially available on web and iOS; OpenAI has not announced Android availability or an international rollout date.
Eligibility: logged-in US users aged 18+ | Plans: Free, Go, Plus, and Pro | Platforms: web and iOS | Supported context: compatible health records and Apple Health | Dashboard data: laboratory results, medication, activity, sleep, and other supported records | Training: connected health data and associated conversations excluded from foundation-model training | Advertising: not used for targeted ads | Clinical limitation: not intended for diagnosis or treatment | Rollout: gradual
Best for: US users who want to consolidate health information, prepare for medical appointments, understand their records, or ask more informed questions of healthcare professionals
✅ ChatGPT Voice — expands to Work and Codex on desktop
ChatGPT Voice is now available inside both Work and Codex in the ChatGPT desktop app. Users can start a new task by speaking naturally, interrupt the voice response, and ask ChatGPT to begin or coordinate work using the tools and permissions available in the selected environment. Voice can now act as an interface for starting and steering agentic work rather than merely providing spoken answers.
In Work, users can verbally initiate longer research, analysis, and artifact-creation tasks. In Codex, developers can use speech to describe coding work, change direction, or coordinate tasks without typing every instruction. The available actions still depend on the tools, permissions, and workspace controls attached to the selected experience. Voice instructions operate within the same authorised tools and controls available when the task is initiated through text — the permission model is unchanged. Mobile availability for this specific Work and Codex integration has not been announced.
Platform: ChatGPT desktop app | Supported experiences: Work and Codex | Input: natural speech | Interaction: supports interruption and follow-up steering | Permissions: limited to tools and actions authorised in the selected Work or Codex environment | Models and limits: governed by the user's applicable ChatGPT Voice access and plan limits | Mobile availability: not announced for this specific Work and Codex integration
Best for: Developers, analysts, and professionals who want to initiate or supervise longer ChatGPT tasks hands-free from the desktop app
✅ OpenAI API — organisation and project spend limits
OpenAI added monthly spend limits at both the organisation and project level. API administrators can now define budgets for an entire organisation or impose separate limits on individual projects, making it easier to monitor costs and assign different spending boundaries to teams, products, or environments.
Spend controls can be used either for monitoring or enforcement. A monitoring limit helps teams compare actual expenditure against a defined monthly threshold. A hard limit turns the budget into an operational boundary: when an organisation or project reaches its configured hard cap, subsequent API requests fail rather than continuing to generate charges. This gives administrators stronger control over unexpected usage spikes, experimental workloads, and autonomous systems that might otherwise continue consuming tokens. OpenAI's release note does not specify whether existing organisations receive a default cap — administrators should review their API platform settings directly.
Platform: OpenAI API | Scope levels: organisation and project | Period: monthly | Control types: monitoring limits and enforceable hard limits | Hard-limit behaviour: API responses fail once the organisation or project reaches the configured cap | Use cases: environment separation, departmental budgets, customer-level controls, experimental agents, and protection against unexpected usage | Pricing impact: no change to model prices; this is a spend-governance feature
Best for: API administrators, finance teams, developers running autonomous workloads, and companies that need enforceable cost controls across multiple teams or projects
Codex on iOS — Mermaid diagrams, forms, and prompt recovery
OpenAI released a substantial Codex update for iOS, adding support for inline Mermaid diagrams and interactive forms inside task transcripts. Mermaid support allows Codex to display structured diagrams — flowcharts, sequence diagrams, system relationships — directly in the mobile task view. Interactive forms make it easier for Codex to request structured information rather than relying on long free-text exchanges.
Inline Mermaid diagrams in task transcripts. Interactive forms inside Codex tasks. Unsent prompts restored when users switch between tasks, hosts, or workspaces — reducing the risk of losing partially written instructions. Task lists improved to prioritise recent activity and show when a host is unavailable during task creation. Blocked or usage-limited runs can now be resumed through improved goal controls. Additional improvements: plan-progress displays, Fast controls, inline dictation, remote onboarding, composer guidance, and iPad navigation. Fixes: duplicate task-list entries that could close the app, iOS 18 task actions, attachment-menu spacing, and duplicate transcription indicators.
Platform: ChatGPT for iOS | New rendering: inline Mermaid diagrams | Interaction: forms inside Codex tasks | Prompt handling: unsent prompts restored across tasks, hosts, and workspaces | Task ordering: recent activity prioritised | Host selection: unavailable hosts shown at task creation | Run recovery: blocked and usage-limited goals can be resumed | Fixes: duplicate task-list crash, iOS 18 task actions, layout and transcription indicators
Best for: Developers managing Codex tasks from iPhone or iPad, particularly workflows involving architecture diagrams, structured inputs, or frequent movement between hosts and workspaces
⚠ Service disruptions — ChatGPT, image generation, and Workspace Agent (all resolved)
OpenAI experienced several service disruptions during the reporting period. On July 19, elevated errors affected ChatGPT conversations, Voice, and Work Mode — identified at 3:01 PM, fully resolved at 5:05 PM. Image generation experienced multiple disruptions July 21–24, including an extended incident July 21–22 and elevated gpt-image-2 errors resolved at 12:09 AM on July 24. Workspace Agent experienced elevated errors on July 22. ChatGPT saw additional elevated error rates on July 23, fully resolved at 8:41 PM.
All listed incidents were resolved before the end of the reporting period. Operational inference from the cluster of incidents: developers using image generation, Work, or Workspace Agent in production should maintain retry logic, exponential backoff, appropriate timeouts, and graceful failure handling rather than assuming every request will complete successfully on its first attempt. This is a resilience recommendation, not an OpenAI statement.
July 19 ChatGPT: conversations, Voice, Work Mode | Identified: 3:01 PM | Resolved: 5:05 PM | July 21–22 image generation: extended disruption | July 22 Workspace Agent: elevated errors | July 23 ChatGPT: elevated errors, mitigated 6:22 PM, resolved 8:41 PM | July 24 gpt-image-2: elevated errors and latency, resolved 12:09 AM | Recommended: retries, exponential backoff, timeouts, graceful failure, idempotent application logic
Best for: Teams using ChatGPT Work, image generation, or Workspace Agent in production — review retry and fallback handling following this week's cluster of disruptions
Plans and Pricing
No new consumer subscription, model, or token pricing changes between July 19 and July 24. The principal financial update was the addition of monthly organisation- and project-level API spend limits — these change how organisations can govern expenditure but do not alter the underlying price of individual API models. Health in ChatGPT is beginning to roll out to eligible Free, Go, Plus, and Pro users in the United States at no separately announced fee; availability is currently restricted by age, geography, platform, and account eligibility.
Best for: API customers: configure organisation and project budgets rather than relying only on retrospective billing review. Eligible US users: evaluate Health without a separately announced subscription. Production teams using image generation or agentic tools: add operational fallbacks following this week's resolved disruptions.
Gemini (Google)
Dateline: July 24, 2026 | Next update: July 31, 2026
A focused five-day window bringing key updates across multimodal interactions, developer platform tooling, agent orchestration, and context management in Gemini. The headline is a major rollout of Gemini Live and Multimodal Voice — introducing enhanced low-latency speech reasoning, expanded dialect and language coverage, and customisable voice profiles across web and mobile. Alongside it: Gemini Code Assist added inline workspace context controls and subagent depth governance; the Gemini API introduced explicit thinking effort controls and updated SDKs for memory store integrations; Google AI Studio launched API key lifecycle and expiration parameters; and the Google Search and Trends grounding connector was updated to allow real-time economic and telemetry queries directly inside Gemini chats.
✅ Gemini Live & Multimodal Voice — major overhaul with new speech models
Google released a significant update to Gemini Live and multimodal voice capabilities. Powered by new foundational audio-to-audio and speech synthesis models, the overhaul brings lower latency, expressiveness adjustments, and wider global language support to hands-free Gemini interactions.
Next-gen speech synthesis: delivers adaptive pitch, dynamic intonation, and natural conversational cadence tailored to context. Ultra-low latency: streamlined endpointing reduces time-to-first-audio-response for fluid multi-turn voice sessions. Expanded global support: new languages and regional accents added, configurable in application settings. Customisable voice personalities: select from distinct voice presets optimised for learning, productivity, or casual conversations. Noise suppression and safety: upgraded acoustic processing for background noise reduction and real-time audio safety filters, including integrated echo cancellation.
Platform: Gemini web interface, iOS, and Android | Rollout: gradual deployment across Free, Advanced, and Workspace tiers | Acoustics: integrated echo cancellation and noise suppression | Voice customisation: multiple character profiles with variable speech rates and conversational tones
Best for: Mobile users, hands-free workflows, language learners, and real-time voice interaction scenarios
✅ Gemini Developer API & Agents — effort controls, lifecycle webhooks, session seeding
The Gemini API received major updates for agentic workflows, granting developers granular control over thinking allocation, lifecycle events, and warm-start context seeding.
Thinking effort controls: set thinking_budget or effort parameters directly on API calls or agent definitions (low, medium, high, xhigh) to control reasoning depth per turn. Expanded webhooks: receive event notifications for background agent tasks, environment updates, and memory state changes without polling. Session and context seeding: pre-load conversation sessions with past memory, structured tool outputs, or initial state prior to the user's first turn. Concurrency protection: optional entity tagging (etag / version checks) on agent state updates returns a 409 Conflict if the session was modified elsewhere. Coordinated SDK updates across Python, Node.js, Go, Java, C#, and Ruby.
Config: pass thinking_config object with effort_level parameter during session initialisation | Webhooks: events cover agent execution states, memory store indexing, and external tool execution | Version: v1beta, header gemini-agents-2026-07 | SDKs: coordinated updates across Python, Node.js, Go, Java, C#, and Ruby
Best for: Developers building complex, long-running agentic systems requiring tight cost, latency, and state controls
✅ Google AI Studio & API — API key expiration controls
Google added key lifecycle governance features to Google AI Studio, allowing security administrators and developers to set auto-expiration limits on API keys.
Expiration presets: choose fixed validity durations (30, 60, 90 days, or a custom date) when generating keys. Automated expiry alerts: email notifications sent to key creators prior to expiration for keys valid for 7 or more days. Existing keys unaffected: previously created keys maintain their existing lifecycles unless updated manually. Expiration dates exposed via the expires_at timestamp in standard key management endpoints.
Location: Google AI Studio → API Keys menu | Options: preset durations (30/60/90 days) or custom expiration date | Email alert: sent before expiry for keys with lifetime ≥7 days | Existing keys: unaffected | Metadata: expires_at timestamp in key management endpoints
Best for: Enterprise security teams enforcing SOC 2 or ISO 27001 compliance and automated credential rotation policies
Gemini Code Assist — workspace controls, subagent depth limits, performance fixes
Gemini Code Assist received stability improvements, context window optimisations, and safety controls for agent execution inside local workspaces.
Subagent nesting depth: set maximum subagent spawn depth via GEMINI_CODE_MAX_SUBAGENT_DEPTH to prevent uncontrolled recursively spawned agent tasks — default 0, preventing nested spawns unless configured. Quadratic context stall fix: resolved a parsing slowdown in long-running coding chats, restoring rapid workspace /resume times. Context normalisation overhead reduced from O(n²) to O(n) per turn. Budget and file controls: enforced strict spending limits across sub-tasks and introduced refined workspace file isolation parameters.
Environment variable: GEMINI_CODE_MAX_SUBAGENT_DEPTH (default: 0, preventing nested spawns unless configured) | Performance: context normalisation overhead reduced from O(n²) to O(n) per turn | Platforms: VS Code, JetBrains IDEs, web
Best for: Software engineers working in large codebases who require fast context retrieval and controlled agent tooling
⚠ Deprecations & Model Migration
Legacy preview endpoints and earlier experimental model iterations are entering sunset windows. Any service calls hitting deprecated routes will fail rather than falling back silently.
Legacy preview endpoints entering sunset windows — ensure all production workflows pass valid model versions and supported headers to prevent request interruptions. Fast-mode migration: ensure fast-tier requests are routed to gemini-2.5-flash or current stable equivalents. Header cutovers: update beta headers on memory and agent endpoints to match the July 2026 SDK specifications.
Best for: DevOps engineers and platform administrators managing production-level API implementations — audit gemini-experimental endpoints and update beta headers now
Plans and Pricing
No pricing changes for Gemini API tiers this week. Standard billing remains consistent across Flash and Pro models. Gemini 3.5 Flash introductory pricing ($2/$10 per MTok) remains active through August 31. Gemini 2.5 family deprecation remains confirmed for October 16, 2026. Google One AI Premium includes full access to Gemini Live, expanded Google Workspace integration, and advanced model usage.
Gemini 3.5 Flash: $2/$10 per MTok through August 31 | Gemini 2.5 family: deprecated October 16, 2026 | Google One AI Premium: includes Gemini Live, Workspace integration, advanced model access | Education and research: Gemini for Education and academic researcher credits continue under current grant structures | No model pricing changes this week
Best for: Take advantage of Gemini 3.5 Flash introductory pricing through August 31. Audit gemini-experimental and legacy endpoints — sunset windows now active.
Microsoft Copilot
Dateline: July 24, 2026 | Next update: July 31, 2026
Microsoft Copilot's biggest moves this week mirrored Anthropic's across the board: the largest Copilot voice mode overhaul since launch; Managed Agents gaining effort controls, lifecycle webhooks, and session seeding; the agent-memory-2026-07-22 header landing with coordinated SDK updates; API key expiration controls arriving in the Copilot Console; and the Copilot Economic Index connector making Microsoft's labour-market AI dataset directly queryable. Copilot Code shipped the same stability and safety update as Claude Code — subagent depth controls, budget cap fix, filesystem isolation toggle, and a quadratic slowdown fix. ⚠ Opus 4.7 fast mode is removed today — hard error, migrate to Opus 4.8 now.
✅ Copilot Voice Mode — major overhaul with new speech models
Copilot shipped its largest voice mode update since launch. The overhaul is built on new speech synthesis and recognition models, delivering more natural intonation, lower latency, and broader language support.
Natural intonation: Copilot voice now adapts tone and cadence to conversational context. Faster response latency: reduced time between user speech and Copilot reply. Expanded language support: multiple languages now available in voice mode. Customisable voices: choose from multiple voice characters for different scenarios (support, education, casual). Improved context retention in long voice sessions. Stronger safety and content controls.
Platform: Copilot iOS + Android apps | Voice toggle: microphone icon | Underlying change: new speech synthesis + recognition models | Languages: expanded beyond English | Customisation: multiple voice characters | Safety: enhanced content controls
Best for: Mobile users, customer-support teams, education platforms, and anyone using Copilot hands-free
✅ Copilot Managed Agents — effort controls, lifecycle webhooks, session seeding
Copilot Managed Agents gained three capabilities giving developers finer control over agent behaviour and observability, matching the Claude Managed Agents update shipped this week.
Effort controls: set low/medium/high/xhigh thinking depth per turn at configuration level — no per-request changes needed. Expanded webhook coverage: four environment.* event types and three memory_store.* event types — react to lifecycle changes without polling. Session seeding: pre-populate agent sessions with initial events before the first user turn, enabling warm-start workflows. Version checks on updates: pass expected_version; 409 on conflict. Event deltas for thread streams reduce parsing overhead. Task tool mode parameter deprecated — subagents inherit parent session permission mode.
Effort: pass effort inside model object (low / medium / high / xhigh) | Webhooks: 4 environment.* + 3 memory_store.* event types | Session seeding: initial_events array at session creation | Version check: expected_version → 409 on conflict | Event deltas in thread stream events | Task tool mode: deprecated, ignored
Best for: Teams building production workflows needing cost-to-quality tuning, real-time observability, warm-start context, or concurrent-update safety
API — agent-memory-2026-07-22 header + SDK updates
The new beta header stabilises memory listing behaviour. All SDKs updated simultaneously to send the new header by default. Breaking changes: depth restricted to 0 or 1; path_prefix must end with /; order_by and order are ignored. Conflict: sending both headers returns 400. Restart pagination from page one when adopting.
Header: agent-memory-2026-07-22 | Endpoint: GET /v1/memory_stores/{id}/memories | Breaking: depth restricted to 0/1; path_prefix must end with /; order_by/order ignored | Conflict: sending both headers returns 400 | SDK versions updated across Python, TS, Go, Java, Ruby, PHP, C#, CLI
Best for: API developers using Managed Agents memory stores — update SDKs and replace old header overrides
✅ API — API key expiration settings in Copilot Console
API key expiration controls added to the Copilot Console. Options: preset durations, custom duration, or Never. Email reminders before expiry for keys with a lifetime of 7 or more days. Existing keys unaffected. Admin API exposes the expires_at field.
Best for: Enterprise security teams requiring credential rotation; developers enforcing key lifetimes aligned with SOC 2 or ISO 27001
✅ Copilot Economic Index — connector for Copilot
Copilot Economic Index connector launched, enabling direct conversational queries of Microsoft's AI labour-market dataset. Supports broad and specific queries; surfaces the underlying data behind any answer. The June 2026 edition includes hourly telemetry, artifact classifiers, and a 9,700-person survey linking usage to workers' beliefs about AI and job security. International context: Canada country brief published; India local pricing launched.
Best for: Researchers, journalists, policymakers, HR and workforce teams grounding AI-and-labour questions in primary Microsoft data
Copilot Code — subagent depth controls, filesystem isolation, quadratic slowdown fix
Copilot Code shipped the same focused stability and safety update as Claude Code this week.
Subagent spawn depth: subagents no longer spawn nested subagents by default — set depth explicitly to allow deeper nesting. Budget cap fix: --max-budget-usd now correctly halts background agents when the cap is reached. Filesystem isolation toggle: skip filesystem isolation while keeping network egress control. Quadratic slowdown fix: message normalisation now O(n) per turn — multi-second stalls in long sessions resolved. Security fix: Unicode normalisation extended to /subtask inputs in agents view.
Subagent depth: default 0 (no nesting) | sandbox.filesystem.disabled: skips filesystem isolation | --max-budget-usd: halts running background agents | Quadratic fix: O(n²) → O(n) | Security: Unicode normalisation on /subtask inputs
Best for: Users with long Copilot Code sessions experiencing stalls; teams running multi-depth agent trees; budget-conscious teams running background subagents
⚠ Opus 4.7 fast mode — removed today (July 24)
Opus 4.7 fast mode removed as of today. Any API call or Copilot Code session using Opus 4.7 with speed: 'fast' now returns a hard error — no silent fallback. Migration to Opus 4.8 fast mode saves 67% on costs ($10/$50 per MTok versus $30/$150). Next deadline: Opus 4.1 (claude-opus-4-1-20250805) retires August 5, 2026 — 12 days away.
Best for: Any team still on Opus 4.7 fast mode: fix the model string now. Also check for claude-opus-4-1-20250805 in your codebase — August 5 is 12 days away.
Plans and Pricing
No new pricing changes this week. Key active windows: Sonnet 5 introductory pricing ($2/$10 per MTok) through August 31. Cowork usage limits doubled through August 5. Copilot for Teachers free through June 2027. Next retirement: Opus 4.1, August 5.
Best for: Developers: migrate Opus 4.7 fast mode before today's deadline. Educators: take advantage of free Copilot access. Enterprises: maximise Cowork tasks before August 5.