Dateline: August 21, 2026 | Next update: August 28, 2026
A two-week window with no new model launch but several platform-level moves that affect every developer and enterprise. The headline pricing story: on August 10, Anthropic made Claude Sonnet 5's $2/$10 per MTok pricing permanent, cancelling the September 1 increase to $3/$15 that teams had been budgeting for. Three APIs graduated from beta to GA simultaneously — Files API, Agent Skills, and Admin API user management. The Workbench and three experimental prompt endpoints retired on August 17 with no recovery path for missed exports. Managed Agents gained session budgets, an advisor model parameter, inference geo-pinning, and GitHub-hosted skills. The computer use and browser use toolsets left beta on August 19, the Python SDK reached 1.0 on August 20, and Claude Academy launched the same day. Anthropic is reportedly in talks to acquire AI infrastructure startup Decart for approximately $6 billion. On the OpenAI side, Ultrafast previewed as a GPT-5.6 Sol service tier running up to 14× Standard speed; ChatGPT Business gained Premium seats; GPT-5.6-Cyber and expanded Daybreak access launched for approved defenders; and internal evaluations that could not rule out the upcoming Astra model reaching the Critical cyber threshold led OpenAI to pause frontier reinforcement-learning training and expand monitoring. o3 leaves ChatGPT on August 26 and the official DALL·E GPT retires August 30. Google celebrated the Gemini app passing one billion active users and unveiled Pixel 11 AI features at Made by Google 2026.
Claude / Anthropic
✅ Claude Sonnet 5 — $2/$10 pricing made permanent
On August 10, Anthropic made Claude Sonnet 5's introductory $2/$10 per MTok the standard price, cancelling the increase to $3/$15 that had been scheduled for September 1. This is not a price cut — nobody's bill goes down. It is a withdrawn 50% increase that teams had already budgeted for.
Standard price: $2/$10 per MTok, permanent from August 10 | Previously scheduled for September 1: $3/$15, cancelled | Model string: claude-sonnet-5 | Subscription plans: unchanged, since they bill on plan limits rather than per token
Best for: Any team running Sonnet 5 in production. If you built Q4 projections on September rates, that money is now free for capacity instead.
★ Files API, Agent Skills and Admin API user management leave beta
Three APIs came out of beta on the same day. Requests to /v1/files, and Messages API requests that reference an uploaded file, no longer need the files-api-2025-04-14 header — and requests sent without it get the current response format, which adds file expiration (expires_in_seconds on upload, expires_at on the file object), page and next_page pagination, and an ids[] filter when listing files.
Agent Skills and the Skills API (`/v1/skills`) also left beta, dropping the `skills-2025-10-02` header. And the Admin API user-management endpoints for Claude Enterprise organisations — members, invites, groups and custom roles — came out of beta, so the `anthropic-beta: ce-user-management-2026-07-13` header is no longer required on group and custom-role requests. Those endpoints had only entered beta on July 14, so this is a fast graduation.
Files API: no `files-api-2025-04-14` header; `expires_in_seconds` on upload, `expires_at` on the file object; `page`/`next_page` pagination; `ids[]` filter on list | Agent Skills: no `skills-2025-10-02` header on `/v1/skills` | Admin API user management: no `ce-user-management-2026-07-13` header on group and custom-role requests; beta began July 14, 2026
Best for: API developers — drop the three beta headers. Requests without them receive the current response shapes, so update your SDKs to read the new fields rather than assuming the old format.
★ Computer use and browser use toolsets reach GA
The computer use tool left beta as the computer_toolset_20260801 toolset: no beta header, batch actions so several actions run in one turn, zoom enabled by default, and per-member configuration through configs. Earlier beta versions stay available, but upgrading an existing integration changes the request shape and tool handling.
Alongside it, Anthropic launched the browser use toolset (`browser_toolset_20260801`), a client toolset for driving a browser your own application hosts. It works inside a browser viewport rather than a whole desktop, and it reads the page itself — accessibility tree, elements, forms and tabs — adding element references, form input, tab management, download reporting and opt-in file upload on top of screenshot-and-click control. That distinction matters: reading the page structure is more reliable than inferring controls from pixels.
Toolsets: `computer_toolset_20260801`, `browser_toolset_20260801` | Computer use: batch actions, `zoom` on by default, per-member `configs`, earlier beta versions still available | Browser use: runs in a viewport your application hosts, reads the accessibility tree, elements, forms and tabs; element references, form input, tab management, download reporting, opt-in file upload | Models: Claude Fable 5, Mythos 5, Opus 5, Sonnet 5 and Opus 4.8 on the Claude API | Migration: see the `computer_20251124` migration guide before upgrading
Best for: Teams building agents that operate internal tools and portals. Browser use is the one to look at first — hosting the viewport yourself keeps execution inside your own infrastructure.
★ Managed Agents — session budgets, advisor model, geo-pinning, GitHub skills
Four additions, each closing a production-readiness gap. Session budgets set a hard cap on a session's spend, priced at public list rates: a session that reaches its budget pauses with the budget_reached stop reason instead of starting new model requests, and changing or removing the budget resumes it. Deployments accept the same budget and apply it to every session they start.
An advisor lets a session's primary thread consult a model at least as capable as the agent's own for strategic guidance mid-turn, configured as a `{"type": "advisor"}` entry in the agent's multiagent roster. Inference geo-pinning controls where model inference runs, set as `inference_geo` inside the `model` object when creating an agent or overridden for a single session. And sessions can load skills from a GitHub repository: when a session mounts a repository, any skills in its root `.claude/skills` directory are discovered automatically at session start.
Session budgets: hard cap at public list rates; `budget_reached` stop reason; changing or removing the budget resumes the session; deployments apply the same budget per session | Advisor: `{"type": "advisor"}` in the multiagent roster, naming the model to consult; must be at least as capable as the agent's model | Geo-pinning: `inference_geo` in the `model` object at agent creation, overridable per session | GitHub skills: skills in the repository's root `.claude/skills` directory, discovered at session start when the session mounts the repository
Best for: Anyone running autonomous agents where a runaway session is a real cost risk — budgets are the guardrail, and they stop the session rather than merely warning you.
★ Python SDK 1.0 — the first stable major release
The anthropic Python SDK reached 1.0.0 on August 20, its first stable major release, after a dense run of 0.12x releases through the preceding fortnight.
Version: `anthropic` 1.0.0, published August 20, 2026 | Preceding releases in this window: 0.121.0 on August 7, 0.122.0 on August 13, and 0.123.0, 0.124.0 and 0.125.0 all on August 19 | Minimum version for Managed Agents session budgets: 0.121.0
Best for: Python teams — pin deliberately rather than drifting. A 1.0 line means the API surface is now expected to hold, which is the point of the version number.
★ Compliance API reaches local sessions, and a workspace ID header
The Compliance API now returns transcripts of Cowork and Claude Code sessions that run on your users' own machines, in beta for Claude Enterprise organisations. GET /v1/compliance/apps/sessions/local lists sessions across the organisation, and two further endpoints retrieve one session's metadata and its transcript, using the existing Compliance Access Key and the read:compliance_user_data scope.
Separately, the Claude API now returns an `anthropic-workspace-id` response header on every response, carrying the `wrkspc_`-prefixed ID of the workspace the request's API key or access token resolved to — including the Default Workspace. It is a small thing that removes a real annoyance: attributing spend across workspaces without inspecting the key itself.
Compliance API: `GET /v1/compliance/apps/sessions/local`, `GET /v1/compliance/apps/sessions/local/{session_id}`, `GET /v1/compliance/apps/sessions/local/{session_id}/messages` | Auth: existing Compliance Access Key with the `read:compliance_user_data` scope | Scope: Cowork and Claude Code sessions running on users' machines; beta, Claude Enterprise | Header: `anthropic-workspace-id`, `wrkspc_`-prefixed, on every API response
Best for: Compliance teams who could see claude.ai sessions but not the ones running on laptops, and platform teams attributing API spend per workspace.
⚠ The legacy Workbench retired August 17 — no recovery path
The legacy Workbench at platform.claude.com/workbench was sunset on August 17, along with the experimental prompt tools APIs for generating, improving and templatizing prompts.
Retired on August 17: the legacy Workbench and the `/v1/experimental/generate_prompt`, `/v1/experimental/improve_prompt` and `/v1/experimental/templatize_prompt` endpoints. Anthropic gave a month's notice on July 17. If any automation still calls those three endpoints, it is broken now, and the replacement is a standard Messages API meta-prompt approach.
Best for: Anyone whose tooling called the experimental prompt endpoints. This is the deprecation from this window that actually breaks running code.
★ Claude Academy — a structured place to learn the tool
Anthropic launched Claude Academy, a learning platform built around its 4D AI Fluency Framework, with courses, tutorials and use cases organised around real problems rather than feature tours. It includes guidance on human-agent team dynamics and department-specific learning paths, and it tracks course completion with badges.
URL: academy.claude.com, also reachable from the Claude profile menu | Structure: 4D AI Fluency Framework — Delegation, Description, Discernment, Diligence | Content: courses, tutorials, use cases, department-specific paths, human-agent team dynamics | Progress: course completion tracking and badges | Audience: individuals through to leaders running organisational change | Emphasis: human agency, durable mindsets over specific features, ethical disclosure of AI use, hands-on practice
Best for: Anyone onboarding staff onto Claude at scale who has been writing their own training deck. Start with the AI Fluency framework courses rather than the tool-specific ones.
⚠ Anthropic reported in talks to buy Decart for about $6 billion
Bloomberg reported that Anthropic was in talks to acquire Decart, an Israeli startup whose optimisation stack extracts more throughput from NVIDIA, AWS Trainium and Google TPU hardware, for roughly $6 billion. The strategic logic was inference cost — the same cost line that sits underneath every pricing decision Anthropic makes.
This was press reporting, not an Anthropic announcement, and it was explicitly described as talks rather than a closed deal. It is worth reading with that caveat, because the deal did not survive: Bloomberg subsequently reported that Anthropic walked away after completing due diligence. We cover that in the September 11 issue.
Best for: Informational. The signal to take from it is not the acquisition but the motive — inference cost is the constraint Anthropic is spending against.
Plans and pricing
The significant change is the permanent Sonnet 5 rate. Claude Code's 50% higher weekly limits ran through August 19 and have now expired.
Sonnet 5: $2/$10 per MTok, permanent from August 10; the September 1 increase to $3/$15 is cancelled | Claude Code limits: the 50% boost expired August 19 | Retired: Claude Opus 4.1 (`claude-opus-4-1-20250805`) on August 5 — requests to it now return an error | Retired: the legacy Workbench and three experimental prompt endpoints on August 17
Best for: Set Sonnet 5 budgets at a permanent $2/$10 and remove any September rate-increase assumption. If anything still calls Opus 4.1, it has been erroring since August 5.
ChatGPT / OpenAI
★ GPT-5.6 Sol Ultrafast — up to 14× faster frontier inference
OpenAI previewed Ultrafast, a new API service tier designed to run GPT-5.6 Sol at dramatically higher inference speeds — up to 14 times the speed of Standard processing and as many as 750 output tokens per second, while retaining the intelligence of the full model.
The significance is less about benchmark capability than latency. Developers have traditionally had to choose smaller models when they need near-real-time responses. Ultrafast removes some of that trade-off, making the flagship reasoning model practical for latency-sensitive coding, voice, financial analysis and interactive agents.
Ultrafast launches first through the OpenAI API and is powered by Cerebras infrastructure. It is a distinct processing tier rather than a separate model: developers still use GPT-5.6 Sol but request substantially faster inference. OpenAI describes this as an early preview rather than a universal replacement for Standard processing.
Model: GPT-5.6 Sol | Service tier: Ultrafast | Platform: OpenAI API | Maximum stated speed-up: up to 14× Standard | Maximum stated generation speed: up to 750 output tokens/sec | Infrastructure partner: Cerebras | Model capability: same GPT-5.6 Sol | Status: early preview
Best for: Developers building real-time or latency-sensitive applications that need GPT-5.6 Sol-level intelligence without conventional frontier-model response times.
★ ChatGPT Business — Premium seats with 5× more usage
OpenAI announced a new Premium seat for ChatGPT Business, aimed at heavy users who repeatedly hit the capacity of a Standard seat. It provides five times more usage and removes the five-hour usage limit while keeping users inside the same centrally administered workspace.
Premium costs $125 per user per month billed monthly, or $100 per user per month billed annually. Standard seats remain $25 monthly or $20 annually. Crucially, organisations do not have to upgrade the whole workspace: Standard and Premium seats can coexist, and administrators can assign or reassign the higher-capacity seats to employees whose workloads justify them. OpenAI also offered the first 10,000 eligible Business customers who joined the waitlist by August 20 up to $500 in promotional workspace credits. That window has now closed.
Premium monthly: $125/user/month | Premium annual: $100/user/month | Standard monthly: $25/user/month | Standard annual: $20/user/month | Included capacity: 5× Standard | Five-hour limit: removed on Premium | Workspace: Standard + Premium seats can be mixed | Promotion: up to $500 workspace credits; waitlist deadline August 20
Best for: ChatGPT Business organisations with a smaller group of extremely heavy Work, Codex or agentic users. Upgrade those users rather than paying Premium pricing across the whole workforce.
★ GPT-5.6-Cyber and expanded Daybreak access
OpenAI substantially expanded its cybersecurity offering with GPT-5.6-Cyber, a cybersecurity-specific model, and two controlled Daybreak access levels. The move follows OpenAI's conclusion that frontier models are approaching capabilities that could materially change both offensive and defensive cybersecurity.
Daybreak Blue provides approved defenders access to frontier general-purpose models including GPT-5.6 Sol, with safeguards designed for authorised defensive work — vulnerability discovery, secure-code review, malware analysis, incident response and patch validation. More specialised cyber work can receive tighter controlled access through the Daybreak programme. The broader strategy is to give trusted defenders access to advanced cyber capabilities before similar capabilities become widely available to attackers, treating access controls, identity, scope, logging, monitoring and human oversight as part of the product. On August 11, Daybreak extended into AWS: both Blue and Red access levels are now available through Amazon Bedrock, letting organisations standardised on AWS bring these models into existing cloud security architectures.
New model: GPT-5.6-Cyber | Programme: OpenAI Daybreak | Blue: frontier general-purpose models including GPT-5.6 Sol with defensive safeguards | Workflows: vulnerability discovery, secure-code review, malware analysis, incident response, patch validation | AWS: Daybreak Blue and Red via Amazon Bedrock from August 11 | Governance: identity verification, defined testing scope, monitoring, logging, human oversight
Best for: Cybersecurity teams, incident-response organisations and enterprises that need frontier defensive capabilities under controlled governance — particularly AWS-heavy estates.
★ ChatGPT for Teens — dedicated under-18 experience
OpenAI launched ChatGPT for Teens, a dedicated experience that automatically applies stronger protections when a user states that they are 13–17 or OpenAI's systems estimate that they are under 18. The product combines educational features with more restrictive model behaviour in areas where younger users face greater risks.
The learning side combines Study Mode with quizzes, Learning Visualizations and a new Study Hours setting that allows teens or parents to specify periods when Study Mode should be enabled by default. Responsible-homework reminders recognise when a user appears to be trying to bypass the learning process and redirect them towards guided problem solving. Teen accounts receive stronger protections around self-harm, violence, eating disorders, dangerous activities and explicit material. OpenAI also explicitly instructs the teen experience not to encourage emotional dependence, use romantic language, or imply that ChatGPT has feelings or consciousness. Additional features include break reminders, sensitive-image warnings and teen-specific onboarding. Parents with linked accounts retain controls including Quiet Hours and limited safety notifications.
Age scope: 13–17 or system-estimated under 18 | Assignment: automatic when age criteria are met | Learning: Study Mode, quizzes, Learning Visualizations, Study Hours, responsible-homework reminders | Safety: stronger protections for self-harm, eating disorders, violence, dangerous activities and sexual/graphic content | Relational safeguards: no romantic language, no emotional-dependence encouragement, no claims of consciousness | Parent controls: Quiet Hours, selected settings, limited high-risk safety notifications
Best for: Families, educators and policymakers tracking how generative AI products are adapting for minors. For teen users, the key change is that safety settings are default rather than optional.
★ ChatGPT — interactive quizzes, editable Project memory, Google Drive in Library
ChatGPT gained interactive quizzes that let users answer questions directly inside a conversation rather than asking ChatGPT to generate a static quiz. Projects also became more flexible, and Google Drive became directly browsable from Library.
Eligible unshared Projects can now switch between default memory and project-only memory without creating a new Project. Project-only memory keeps context contained: ChatGPT can use conversations from the Project but does not reference memories or chats outside it, and Project information is not added to memory used elsewhere. Shared Projects remain project-only and cannot be switched; ChatGPT Work is unavailable inside project-only Projects. Separately, users who connect the Google Drive plugin can now browse Drive files and folders directly from Library, including items shared with them, and pull them into a conversation through the composer or @ mentions without re-uploading. Google Docs, Sheets and Slides can remain open alongside the conversation. Eligible paid users can also receive homepage suggestions based on conversation history and connected tools; Free and Go users gained the ability to select Think on the web.
Quizzes: consumer + Edu, web/mobile | Project memory: existing eligible unshared Projects can switch default ↔ project-only | Shared Projects: remain project-only | Work: unavailable in project-only Projects | Drive: files, folders and directly shared items via Library, composer or @ mention; re-upload not required | Homepage suggestions: eligible paid users | Think: Free and Go on web
Best for: Students using ChatGPT for active learning, professionals needing separation between Project-specific context and general memory, and Google Workspace users doing document-heavy research.
★ ChatGPT Business — Computer History on macOS, restaurant reservations
OpenAI introduced Computer History, an optional macOS feature that allows ChatGPT and Codex to use context from selected applications and websites. Rather than capturing screenshots or recordings, it records interaction events that can later help ChatGPT understand what the user was working on. Separately, ChatGPT gained real-time restaurant reservation search.
Computer History is off by default. A Business workspace administrator must first enable access, after which individual members decide whether to opt in. Users can pause collection, choose which applications and sites are included, and inspect or delete their timeline data; private browsing is excluded. At launch it is not available in the EEA, United Kingdom or Switzerland — a restriction particularly relevant for European enterprise deployments. Restaurant reservations: ChatGPT can now search real-time availability inside a conversation via OpenTable (global), Resy (US) and Yelp (US and Canada). Users provide a location, date, time, party size and preferences, and ChatGPT returns available slots rather than simply recommending restaurants. Rolling out across all ChatGPT plans on mobile, web and desktop; ChatGPT Work is excluded.
Computer History — platform: ChatGPT macOS | Plan: Business | Default: off | Admin must enable, user must opt in | Captured: interaction events | Not captured: screenshots, screen recordings, microphone or system audio | Private browsing excluded | Initial exclusions: EEA, UK, Switzerland | Reservations — partners: OpenTable (global), Resy (US), Yelp (US + Canada); all plans; Work excluded
Best for: Business users who want ChatGPT to retain desktop context across applications — European organisations should note the EEA/UK/Switzerland exclusion at launch.
⚠ Enterprise connected-app sync — individual connections retired
OpenAI retired individually authorised sync connections for connected enterprise apps. New individual-user sync connections stopped being available on August 10, and existing connections were disabled on August 14, with deletion of associated synced data beginning at that point.
The change does not eliminate enterprise sync — it moves organisations towards administrator-managed connections. Google Drive can use the Google Drive plugin and, where indexed knowledge is needed, domain-wide delegated admin sync. SharePoint follows a similar administrator-managed model, while GitHub moves to the non-synchronised GitHub plugin. Existing individual sync connections should no longer be treated as active; enterprise administrators who relied on them need to ensure replacement plugins or administrator-managed sync are configured. OpenAI said replacement plugin availability for GitLab Issues and Azure Boards would be communicated separately.
Best for: Enterprise and Edu admins — verify that important knowledge sources have moved from individual-user sync to supported administrator-managed integrations.
⚠ Astra may reach Critical cyber capability — OpenAI slows frontier training
The most consequential research and safety development of the period concerns Astra, an upcoming OpenAI model. Internal evaluations led OpenAI to conclude on August 7 that it could no longer rule out Astra reaching the Critical cybersecurity capability threshold under its Preparedness Framework.
On August 18, OpenAI disclosed that this assessment, together with lessons from the OpenAI–Hugging Face security incident, had already affected its model-development schedule. The company temporarily slowed frontier scaling, including a two-week pause in reinforcement-learning training for models intended for deployment; its largest planned frontier RL run remained on hold at publication. OpenAI now requires its strictest security environment for Astra and cyber-model workloads. Monitoring has been expanded to all RL training and evaluations involving tools for models at Sol capability or above, as well as all Astra inference that uses tools. The monitoring stack uses token-level activation classifiers and increasingly capable automated investigators to detect unauthorised access, data theft, destructive behaviour and attempts to defeat safeguards, with a target alert time within 30 minutes. OpenAI estimates the monitoring overhead at roughly 20% of the inference compute being monitored.
Best for: AI safety researchers, cybersecurity leaders and organisations tracking frontier-model timelines. The important implication is that capability growth is now materially changing OpenAI's internal training and deployment process.
⚠ The Defender's Window — OpenAI urges rapid AI-assisted security adoption
OpenAI published a broader warning that organisations have a limited period in which to strengthen their defences before increasingly capable AI systems make automated exploitation substantially easier. The company argues that the same capabilities can advantage defenders — but only if security teams deploy them quickly.
The recommendation is not simply to purchase an AI security tool. OpenAI argues for improving foundational controls while gradually introducing AI into vulnerability discovery, code review, alert triage, incident response and remediation, with autonomy expanding as organisations gain confidence. Recommended progression: foundational controls → AI-assisted code and security review → advisory scanning → live alert triage → narrowly scoped automated remediation. The warning follows OpenAI's assessment that it underestimated the real-world cyber capability demonstrated during the OpenAI–Hugging Face incident.
Best for: CISOs and enterprise security teams. The recommendation is to begin deploying AI-assisted defensive workflows now rather than waiting for autonomous offensive capability to become commonplace.
⚠ o3 retires from ChatGPT August 26 — DALL·E GPT August 30
OpenAI o3 is five days from its scheduled retirement in ChatGPT, following the 90-day sunset announced in May. The official DALL·E GPT retires four days after that.
The o3 retirement applies to ChatGPT only — OpenAI has not announced a corresponding API retirement. Users who still rely on o3-specific prompts or workflows should test them against GPT-5.6 reasoning models before August 26. The DALL·E GPT retirement does not remove image generation from ChatGPT: users should move to ChatGPT Images, and user-created GPTs with image generation enabled are unaffected. OpenAI recommends saving any images you want to retain before August 30.
Best for: Paid ChatGPT users still selecting o3 manually — move repeatable workflows to GPT-5.6 before August 26, and preserve anything from DALL·E GPT before August 30.
Plans and pricing
The major commercial change in this two-week window is ChatGPT Business Premium. GPT-5.6 model API list pricing did not change; the major API development is instead the Ultrafast preview, which creates a new performance tier where latency matters.
ChatGPT Business Standard: $20/user/month annual, $25 monthly | Premium: $100/user/month annual, $125 monthly | Premium capacity: 5× Standard | Five-hour usage limit: removed on Premium | Premium promotion: waitlist credit offer ended August 20 | GPT-5.6 API list-price change: none this period | Ultrafast: new GPT-5.6 Sol high-speed API tier | o3 ChatGPT retirement: August 26 | DALL·E GPT retirement: August 30
Best for: Business customers should identify the small number of employees whose usage justifies Premium rather than upgrading entire workspaces. API developers with latency-sensitive GPT-5.6 Sol workloads should evaluate Ultrafast.
Gemini / Google
★ Gemini 3.7 Flash reaches general availability
Gemini 3.7 Flash became generally available on August 13 as gemini-3.7-flash, offered at an introductory price through the end of the year. That end date is the part worth writing down: anything you cost out on this model today is cheaper than it will be on January 1.
Model ID: `gemini-3.7-flash` | Status: generally available from August 13, 2026 | Pricing: introductory rate through December 31, 2026
Best for: Developers moving volume workloads onto Flash — budget the post-December rate now rather than discovering it in the new year.
★ Ask Gemini arrives in Google Chat
Google introduced Ask Gemini in Chat, a unified command interface powered by Workspace Intelligence, covering search, content creation and task management from inside the Chat window.
Surface: Google Chat | Powered by: Workspace Intelligence | Functions: search, content creation, task management | Editions: Business Standard/Plus, Enterprise Standard/Plus, AI Pro for Education
Best for: Workspace teams who run coordination out of Chat rather than email.
★ Made by Google 2026 — the Pixel 11 line, built around Gemini
At Made by Google 2026, Google unveiled Pixel 11, Pixel 11 Pro, Pixel 11 Pro XL and Pixel 11 Pro Fold — a line the company describes as designed for Gemini Intelligence. Google also recorded presentations in Google Slides through Google Vids on August 20, extending the Vids integration into Slides.
Devices: Pixel 11, Pixel 11 Pro, Pixel 11 Pro XL, Pixel 11 Pro Fold | Positioning: designed for Gemini Intelligence | Workspace, August 20: record presentations in Google Slides with Google Vids, all editions
Best for: Anyone evaluating Android hardware where the assistant is the point rather than an add-on.
★ The Gemini app passes one billion monthly users
Google confirmed that the Gemini app passed one billion monthly users, which it calls the fastest-growing product in the company's history.
Scale: more than 1 billion people using the Gemini app every month | Google's own characterisation: fastest-growing product in its history
Best for: Context rather than action. It matters mainly as a measure of how much default familiarity you can assume when you put Gemini in front of staff or customers.
Microsoft Copilot
★ Agent Plugins 1.0 — one plugin format, governed outside any single vendor
Agent Plugins 1.0 is an open standard that packages agent skills and MCP servers into one installable plugin, governed independently of any single vendor. Build a plugin once and it works across compatible agent clients, with vendor-specific features kept in namespaced directories.
Status: generally available on all Copilot plans | Surfaces: VS Code, Copilot CLI, GitHub Copilot SDK, GitHub Copilot app | Distribution: installable from the Awesome Copilot marketplace | Named participants: AWS, Anysphere, Microsoft, OpenAI, Vercel, with Google as a core maintainer
Best for: Teams that have been rebuilding the same skill or MCP server for each tool. The governance point matters more than the format: a standard maintained across vendors is one you can adopt without betting on a single client.
★ MAI-Code-1.1-Flash — native vision at 73% lower list price
Microsoft's small-tier coding model gained native vision support and better coding quality, at what Microsoft states is a 73% lower list price than MAI-Code-1-Flash. Free and Student users get it through auto-selection; everyone else can select it manually.
Surfaces: Copilot CLI, the cloud agent, the GitHub Copilot app, Chat on GitHub, VS Code, Visual Studio, GitHub Mobile, JetBrains IDEs, Eclipse and Xcode. Annual subscribers are charged at a 0.25× premium request multiplier; usage-based billing runs at provider list pricing. Microsoft announced the deprecation of the model it replaces on the same day.
Model: MAI-Code-1.1-Flash | Capability: native vision, improved coding quality | Price: Microsoft states 73% below MAI-Code-1-Flash list | Plans: Free, Student, Pro, Pro+, Max, Business, Enterprise | Billing: 0.25× premium request multiplier for annual subscribers; provider list pricing under usage-based billing | Also August 11: upcoming deprecation of MAI-Code-1-Flash announced
Best for: High-volume completion and small-task work, where a 73% list-price cut compounds. If you pinned the older Flash model anywhere, that announcement is your notice.
★ GitHub Copilot arrives in Slack and Microsoft Teams
Mentioning @GitHub in Slack now answers questions about your code and GitHub activity, triages issues, investigates failures, implements changes and opens pull requests, posting a link back to the conversation for review. Available to organisations on Copilot Business and Copilot Enterprise.
The Microsoft Teams equivalent shipped the same day: cloud agent sessions started by mentioning `@GitHub` in a channel or direct message, where team members can ask questions, add context and direct the agent's work together. Participants with repository write access can trigger Copilot to make changes. Available with paid GitHub Copilot plans, in public preview.
Slack: `@GitHub` mention; answers about code and GitHub activity, issue triage, failure investigation, changes and pull requests; Copilot Business and Enterprise; public preview | Teams: `@GitHub` in channels or DMs, shared agent sessions, write access required to trigger changes; paid Copilot plans; public preview
Best for: Teams whose incident and triage conversations already happen in chat. The write-access requirement in Teams is the control worth checking before you turn it on.
★ The August 10 release — Kimi K3, CLI rewind, JetBrains memory
The weekly bundle rolled Kimi K3 out to Copilot Pro, Pro+, Max, Business and Enterprise, and added plugin version management and batch updates to the Copilot app, along with a side chat for discussing an agent's questions separately from the main thread.
Copilot CLI gained `/tasks` for managing subagents, queued prompts and commands during agent execution, `--plan` with `--mode autopilot` in headless mode, `/rewind` to undo changes without git, and `/app` to preserve session and folder context. JetBrains gained Copilot memory across chat sessions and local Ollama models as a BYOK provider. VS Code 1.133 added model switching between Claude BYOK and Copilot models, pinned prompts in long chats, and live HTML refresh. Gemini 3.7 Flash arrived in Copilot on August 13 and Grok 4.6 on August 14. Earlier in the window, on August 7, code review effort levels reached GA — renamed Lite and Balanced, with org admins setting the default and developers choosing per review — and on August 18 enterprise managed settings arrived in Copilot for JetBrains.
Kimi K3: Pro, Pro+, Max, Business, Enterprise | Copilot app: plugin version management, batch updates, side chat | CLI: `/tasks`, queued prompts, `--plan` with `--mode autopilot`, `/rewind`, `/app` | JetBrains: Copilot memory, Ollama as BYOK provider, enterprise managed settings (Aug 18) | VS Code 1.133: Claude BYOK model switching, pinned prompts, live HTML refresh | Models added: Gemini 3.7 Flash (Aug 13), Grok 4.6 (Aug 14) | Code review effort levels GA (Aug 7): Lite and Balanced, renamed from Low and Medium; Pro, Pro+, Max, Business, Enterprise
Best for: CLI users — `/rewind` undoes an agent's changes without git, which is the safety net that makes autopilot mode reasonable to try.
★ Microsoft 365 Copilot — authoritative SharePoint sites, Anthropic models in Word
Administrators can now designate specific SharePoint sites as authoritative, so company news and policies are prioritised in Copilot Search over whatever else matches. Outlook gained coaching feedback on drafts in chat, applying only the suggestions you choose, and meeting preparation that pulls the relevant context, tasks and documents together beforehand.
Word can now use Anthropic models in addition to OpenAI models when editing with Copilot. PowerPoint gained three additions: creating a presentation from the web app home screen, referencing web sources while creating one, and using enterprise image assets hosted on Adobe Experience Manager. The Planner Agent became available in all group-based Planner plans, including basic plans, for licensed users. A Consumption Dashboard in Viva Insights now tracks Copilot credit usage across Cowork and the Work IQ API, with budget information for managers and admins. Copilot connectors also began running content and identity crawls in parallel, so ingested content becomes available faster, and the ServiceNow Knowledge and Catalog connectors now enforce role-based permissions.
SharePoint Authoritative Sites: Roadmap 561323, Windows and web, GA | Outlook coaching feedback: Roadmap 559418, all platforms, GA | Outlook meeting preparation: Roadmap 542186, Windows, Copilot licence required | Word with Anthropic models: Roadmap 558440, web, GA | PowerPoint: web app creation (Roadmap 560537), web sources (555898), Adobe Experience Manager assets (516038/516039) | Planner Agent: Roadmap 511820, Copilot licence required | Consumption Dashboard: Roadmap 566302, Viva Insights and admin centre | Connectors: parallel content and identity crawl; ServiceNow role-based permissions
Best for: Admins who have watched Copilot Search surface a stale deck instead of the actual policy — Authoritative Sites is the setting that fixes that, and it is a decision nobody makes for you.
⚠ The consumer Copilot app merges — three features retired August 18
Microsoft is consolidating Copilot into a single app that combines Copilot chat and image creation with Microsoft 365 for work. Web users are redirected automatically, mobile apps update in place, and installed Windows and Mac apps get a new icon.
Three features retired on August 18 and the data does not all survive: Group Chat goes, and its threads, messages and images do not transfer; Podcasts become inaccessible; and Deep Research retires for consumer app users. Chats, images and created content migrate to the updated app, and files move to OneDrive. Work and personal experiences stay separated by design — you switch between them with the account switcher, and Microsoft states that data does not flow between them. Microsoft also warns that some features may be temporarily unavailable during the rollout.
Best for: Anyone who used Copilot Group Chat or Podcasts. That content was not carried over, so if it mattered, it needed exporting before August 18.
Plans and pricing
No Copilot seat-price changes were published in this window. The commercial change that did land is a billing mechanic: annual subscribers are charged a 0.25× premium request multiplier on MAI-Code-1.1-Flash, while usage-based billing runs at provider list pricing.
Copilot seat prices: unchanged this period | MAI-Code-1.1-Flash: 0.25× premium request multiplier for annual subscribers; provider list pricing under usage-based billing; Microsoft states a 73% lower list price than the model it replaces | Agent Plugins 1.0: generally available on all Copilot plans at no additional cost
Best for: Nothing to action on seat pricing. The multiplier on the new Flash model is the number to put into any cost model built around high-volume completions.