Dateline: September 11, 2026 | Next update: September 18, 2026
No new frontier model this week — the work was in the platforms around them. OpenAI released the Agents API in public beta on September 10, exposing the managed Codex harness so that session orchestration, context compaction and recovery are handled server-side, and made GPT-Live-1 generally available in the API at $0.05 per minute billed per second, with backend model and tool usage charged separately. GPT Image 2.5 Flare and Sunburst shipped on September 8. ChatGPT Work gained a Data agent that reads an organisation's own metric definitions, a tailored Financial Services experience built on Astra, and OpenAI paused new sign-ups to the $200 Pro tier. Also on September 8, OpenAI published what it describes as a resolution of the Navier–Stokes Millennium Prize Problem; the Clay Mathematics Institute has not accepted it and a competing priority claim surfaced the same week, so it belongs in the digest as a claim under review rather than a settled result. Anthropic launched smart reports in beta for Claude Enterprise, gave Managed Agents permission policies an auto mode that evaluates each tool call server-side, and shipped maxEffortLevel in Claude Code — the first setting that caps thinking compute organisation-wide across Bedrock, Vertex and Foundry — alongside a fix for a prompt-cache miss that had been firing roughly once an hour in long sessions. Google released the Gemini desktop app for Windows with an Alt+Space shortcut, took Gemini Enterprise Projects to GA, and moved several Workspace admin controls into place. On the Copilot side, GPT-6 Astra reached general availability in GitHub Copilot, enterprise-managed permissions for agent operations went GA, and MAI-Code-1-Flash was deprecated.
Claude / Anthropic
★ Smart reports — an AI-written read on how your organisation actually uses Claude
Smart reports turn Claude Enterprise usage telemetry into a written analysis rather than a dashboard. A report can cover a specific team, a time range of up to 28 days, and one or more products — chat, Claude Code and Claude Cowork — and it answers custom questions you add to it. Reports are generated on demand and typically take a few hours.
What a report contains: the workstreams a team runs and the deliverables produced, cost per session, task outcomes, common friction points and inefficiencies, the sessions where spend concentrates, the most complex autonomous work, and — the part with the clearest payback — repeated patterns that could be packaged as a shared skill so the whole team gets the same result faster. Anthropic's own examples are turning call notes into follow-ups, building account briefs and drafting QBR outlines. Enabling it takes Owner or Primary Owner access: Organisation settings → Capabilities → Analytics → "Smart reports (beta)". After that, Primary Owners, Owners, Admins and custom roles with analytics view access can create and view reports.
Plans: Claude Enterprise | Status: beta, 10 reports per organisation per month | Enablement: Organisation settings → Capabilities → Analytics | Access: Primary Owner, Owner, Admin, or a custom role with analytics view access | Scope: team (via groups or SCIM department and cost-centre fields), up to 28 days, products chat / Claude Code / Cowork, plus custom questions | Generation: on demand, typically a few hours | Not available for: organisations using customer-managed encryption keys (CMEK), HIPAA configurations, or Access Transparency
Best for: Enterprise admins who have the analytics dashboard and still cannot answer "what is this actually being used for". Note Anthropic's own warning: the reports are not designed for, and should not be used for, evaluating individual performance or making employment decisions.
★ Managed Agents — an auto permission mode, and a terminal you can attach to a session
Managed Agents permission policies gained a third option. Instead of allowing or denying a tool up front, auto has the server evaluate each agent or MCP tool call as it happens and then run it, deny it, or pause for your approval.
The `agent.tool_use` and `agent.mcp_tool_use` events now report how each call was evaluated, in an `evaluation` field alongside `evaluated_permission` — so the decision is auditable rather than implicit. Separately, the `ant` CLI added `ant beta:sessions connect`, which attaches your terminal to a running Managed Agents session: you can follow it live, send messages, and allow or deny tool calls that are waiting for approval. Passing `--web` serves the Claude Console's session viewer locally and opens the session there instead.
Permission policies: `auto` added alongside the existing options | Evaluation: server-side, per call | Events: `agent.tool_use` and `agent.mcp_tool_use` report an `evaluation` field alongside `evaluated_permission` | CLI: `ant beta:sessions connect`, with `--web` to open the Console session viewer locally
Best for: Teams running agents that need a human in the loop for some calls but not all of them — `auto` is the setting that makes that distinction without pre-enumerating every tool.
★ Claude Code — a cap on thinking compute, and a prompt-cache miss that was costing money
The operationally significant addition is maxEffortLevel, set at the top level or per model under modelSettings. It caps the effort level on every provider, Bedrock, Vertex and Foundry included; users can still choose a lower level but cannot exceed the cap. That is the first straightforward lever admins have over thinking spend in a managed deployment.
The most consequential fix is quieter. In long sessions, tool definitions were being re-rendered after an OAuth token refresh — roughly once an hour — which missed the prompt cache and discarded extended-thinking context each time it fired.
Also added: `--system-prompt-snapshot off`, which renders the system prompt fresh on every request instead of reusing the conversation's recorded prompt, so you can iterate on prompt text without restarting the session. In 2.1.268, the Claude apps gateway can set `pricing:` in `gateway.yaml` so signed-in clients receive the same rates through managed settings and `/cost` matches the spend meter; a new `gatewayInternalNetworks` managed setting lets admins allow `/login` on an organisation's own public IPv4 block; and a startup warning now fires when `access_control.allow_cidrs` is empty. A cluster of prompt-cache fixes landed in 2.1.267 — switching models with `/model` no longer re-sends every tool definition, resumed sessions no longer re-render tool descriptions, and subagents started with `--system-prompt` record the system prompt once instead of re-rendering it. Security-adjacent fixes: plugin and marketplace errors no longer show a token or password from a git source URL, `/mcp` and `/plugin` output no longer shows secrets resolved from `${VAR}` placeholders, and deny rules now apply on symlinked directories such as `/etc`, `/tmp` and `/var` when a path is given by its real location.
2.1.267 (Sept 9): `maxEffortLevel` top-level or per model in `modelSettings`, applies on Bedrock, Vertex and Foundry; `--system-prompt-snapshot off`; OAuth-refresh prompt-cache fix; Cowork scheduled tasks fixed for organisations whose managed settings require sandboxing | 2.1.268 (Sept 10): gateway `pricing:` in `gateway.yaml`, `gatewayInternalNetworks`, `claude self-hosted-runner --remove-session-state`, `--json` on `claude plugin install/uninstall/update/enable/disable`; WebFetch now fails after 300 seconds, overridable with `CLAUDE_CODE_WEBFETCH_DEADLINE_MS` | 2.1.265 (Sept 8): 1 GB cap on tool results saved to disk, `user.email` and `user.groups` added to gateway telemetry, Foundry Entra ID upstreams fixed when `ANTHROPIC_FOUNDRY_API_KEY` is set | 2.1.266: reverted a 2.1.265 regression where `CLAUDE_CODE_USE_GATEWAY` forced Cloud-gateway sign-in on its own
Best for: Enterprise admins should deploy `maxEffortLevel` — it is the cost control that was missing. Everyone running long sessions with extended thinking should update: the hourly cache miss was silent, and it was billing you for it.
⚠ Anthropic walks away from the Decart deal
Bloomberg reported that Anthropic has abandoned its roughly $6 billion pursuit of Decart, the Israeli startup whose optimisation stack squeezes more throughput out of NVIDIA, AWS Trainium and Google TPU hardware. The talks were first reported on August 13 and we covered them in the August 21 issue; they ended after due diligence was complete.
This is press reporting rather than an Anthropic announcement, and the reason has not been disclosed. Reporting offers two competing explanations: that Anthropic concluded after diligence that it could build the capability itself, and that a demand to move Decart's operations from Tel Aviv to San Francisco was refused. Both are attributed to unnamed sources, so treat the motive as unsettled — what is reported as settled is that the deal is off.
Best for: Informational, but it leaves the original signal intact and unresolved: inference cost is still the constraint Anthropic is spending against, and it no longer has an acquisition to solve it with.
Plans and pricing
No pricing changes this week. Claude Fable 5.1 and Mythos 5.1 remain at $10/$50 per MTok with cache reads at $0.25 per MTok — 0.025x the base input price, against 0.1x on other models — which is still the most consequential active pricing change for cache-heavy agentic work. Both models require 30-day data retention and are not available under zero data retention unless Anthropic expressly authorises it.
Fable 5.1 / Mythos 5.1: $10/$50 per MTok, cache reads $0.25 per MTok, 1M context, 128k max output, always-on adaptive thinking | Data retention: 30 days required; ZDR only by express authorisation | Opus 5: $5/$25 per MTok | Sonnet 5: $2/$10 per MTok | Haiku 4.5: $1/$5 per MTok | No model retirements announced this period
Best for: Nothing to action on price. The 30-day retention requirement on Fable 5.1 and Mythos 5.1 is the constraint worth checking if your workspace runs under zero data retention.
ChatGPT / OpenAI
★ Agents API — the Codex harness, operated by OpenAI
The Agents API puts the managed Codex harness behind an API, so OpenAI handles session orchestration, context compaction and recovery rather than each developer building that layer themselves. That is the whole proposition: the parts of a long-running agent that are tedious and easy to get subtly wrong — keeping a session alive across context windows, recovering from a failure mid-task — become server-side concerns.
Status: public beta, all developers | Harness: managed Codex harness | Handled server-side: session orchestration, context compaction, recovery | Fees: no additional charge for the API itself — you pay for the tokens and tools the agents use
Best for: Teams that have already written their own orchestration, compaction and retry logic around repeated model calls, and would rather maintain application code than agent infrastructure.
★ GPT-Live-1 — full-duplex voice, generally available at $0.05 a minute
GPT-Live-1 is now generally available in the API. The voice session costs $0.05 per minute, billed per second, and backend model and tool usage is charged separately — which is the important half of the pricing: the voice layer is a thin, predictable cost on top of whatever reasoning model the agent actually calls.
Status: generally available in the API | Price: $0.05 per minute for voice sessions, billed per second | Backend: model and tool usage billed separately from the voice layer
Best for: Anyone costing out a voice agent. Model the voice minutes and the backend reasoning as two separate lines, because that is how they bill.
★ GPT Image 2.5 — Flare and Sunburst
Two image models shipped: gpt-image-2.5-flare and gpt-image-2.5-sunburst. Both support the new xhigh and max quality settings and use GPT Image 2 token rates, so the quality ceiling rose without the per-token price moving.
Models: `gpt-image-2.5-flare`, `gpt-image-2.5-sunburst` | New quality settings: `xhigh`, `max` | Pricing: GPT Image 2 token rates
Best for: Teams generating images at volume — the new quality tiers are available without a new rate card to model.
⚠ Navier–Stokes — OpenAI publishes a claimed resolution, and it is already contested
OpenAI published what it says is a resolution of the Navier–Stokes existence and smoothness problem, one of the Clay Mathematics Institute's Millennium Prize Problems, together with a formalisation in the Lean proof assistant. The company says an internal system running roughly 10,000 agents produced the result in about 88 hours, and that it will not seek the $1 million prize.
Treat this as a claim under review, not a settled result. The Clay Mathematics Institute still lists Navier–Stokes as unsolved; its president, Martin Bridson, called the announcement exciting but said evaluation would be deliberately unhurried and absolutely rigorous. Two substantive objections are already public. First, the proof relies on a smooth external force applied to the fluid: the written problem permits that, but most working mathematicians exclude it from the core question, and there is no public evidence the construction works without it. Second, priority is disputed — Tristan Buckmaster of New York University, working with Anthropic researcher Levent Alpöge, says they completed a related three-dimensional Euler result on August 15 and finished verification on August 22, ahead of OpenAI's September 5 result.
Best for: Anyone tracking what frontier systems can actually do. The honest summary this week is that a lab claims a major mathematical result, published its proof and a Lean formalisation, and the mathematical community has not yet reached a verdict. That is a meaningful event either way — but it is not the same thing as the problem being solved.
★ The Data agent — company data, answered by asking
OpenAI launched a Data agent in ChatGPT Work that connects to company data, investigates what changed, and builds interactive dashboards you can share — the point being that someone without SQL can ask the question themselves.
It connects to approved sources including Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB and Snowflake, and it builds and interacts with dashboards in Omni, Oracle BI, Power BI, Sigma, Tableau and ThoughtSpot. The part that decides whether the answers are trustworthy is the context layer: it reads your organisation's business terms, metric definitions, custom calculations and data relationships from semantic layers such as Databricks Genie Ontology, dbt, GitHub and Snowflake Horizon, rather than inventing its own definition of a metric.
Surface: ChatGPT Work | Data sources: Amazon Redshift, Datadog, Google BigQuery, ClickHouse, Databricks, MongoDB, Snowflake and others | BI tools: Omni, Oracle BI, Power BI, Sigma, Tableau, ThoughtSpot | Semantic context: Databricks Genie Ontology, dbt, GitHub, Snowflake Horizon and BI dashboards | Access: through approved connected data sources
Best for: Finance, operations and leadership teams whose questions currently queue behind an analyst. Check the semantic layer first — an agent reading your real metric definitions is useful, one guessing at them is worse than nothing.
★ ChatGPT for Financial Services — Astra plus licensed market data
A version of ChatGPT Work built for financial institutions, combining GPT-6 Astra with built-in financial data for research, modelling and client-ready materials. OpenAI says it was shaped in design partnership with Morgan Stanley and Evercore.
Model: GPT-6 Astra | Design partners named: Morgan Stanley, Evercore | Bundled premium data providers named: Daloopa, PitchBook, LSEG News, Crunchbase | Purpose: research, modelling and client-ready materials | Availability: eligible financial institutions, through OpenAI
Best for: Investment banking, research and corporate finance teams. The bundled data is the commercial argument — it removes the step where every institution negotiates provider access separately before anyone can start.
★ OpenAI says it has reached the "automated research intern" milestone
OpenAI reports it has met the goal it set of building an "automated research intern" by September — defined as a system that can carry out well-defined research tasks under human direction, including tasks that would take a skilled researcher a few days.
Definition: well-scoped research tasks under human direction, of a few days' human effort | Claimed achievement: by September 2026 | Usage ratio: 3.1 agent-workdays of effort for every workday of human labour across OpenAI's research organisation, on an eight-hour day, as of mid-August 2026 | Inflection: before June 2026, total agent runtime across the research organisation was still below total human labour
Best for: Anyone modelling what agentic work does to headcount and compute. The 3.1 figure is the most concrete disclosure any lab has published about its own internal substitution rate — and note it is OpenAI measuring OpenAI.
⚠ New Pro $200 sign-ups paused
OpenAI temporarily paused new sign-ups and upgrades to the $200 ChatGPT Pro tier on September 10. The pause covers moves up from Free, Go, Plus and Pro $100 alike.
Existing Pro $200 subscriptions are not affected and continue as normal, and new or existing Pro $100 subscriptions are unaffected. No reopening date has been announced. If you were planning to move a heavy user onto the $200 tier this month, that route is closed for now and Pro $100 is the one still open.
Best for: Anyone budgeting a seat upgrade. This is a capacity signal as much as a commercial one — the tier most exposed to heavy agentic use is the one OpenAI stopped selling.
Plans and pricing
The new prices this week are GPT-Live-1's voice layer at $0.05 per minute, billed per second, with backend model and tool usage charged separately, and the Agents API, which carries no orchestration fee at all during public beta — you pay for the tokens and tools your agents consume. GPT Image 2.5 Flare and Sunburst use GPT Image 2 token rates.
GPT-Live-1: $0.05 per minute, billed per second; backend model and tools billed separately | Agents API: no additional fee in public beta | GPT Image 2.5 Flare and Sunburst: GPT Image 2 token rates, with `xhigh` and `max` quality settings | Earlier in the window: mutual TLS and X.509 workload identity federation reached GA on August 29, and the Assistants API shut down on August 26 with migration to the Responses and Conversations APIs
Best for: If you are still calling the Assistants API, it is gone — that shutdown completed on August 26 and the migration path is the Responses and Conversations APIs.
Gemini / Google
★ The Gemini desktop app arrives on Windows
Google released a native Gemini app for Windows, available globally. Alt + Space brings it up over whatever you are working on, and it runs in the background rather than as a browser tab.
Beyond the shortcut, the app handles multi-step tasks through Gemini Spark, drafts summaries from Gmail and Drive, and generates images and video natively. Google says it runs in the background without slowing performance. It is available to Workspace customers and personal account holders, with image generation via Nano Banana; some features require a Google AI subscription, availability varies, and the app is 18+.
Platforms: Windows 10 and 11 | Shortcut: Alt + Space | Availability: globally, Workspace customers and personal accounts | Features: Gemini Spark multi-step tasks, summaries from Gmail and Drive, image generation via Nano Banana, video creation | Requirements: some features need a Google AI subscription; availability varies; 18+ | Download: gemini.google/desktop
Best for: Windows users who have been keeping a browser tab open for Gemini, and Workspace teams that want the assistant reachable from a keyboard shortcut rather than a tab.
★ Workspace — Gemini becomes a cross-app orchestrator, and admins get the controls
The headline Workspace change is orchestration: Gemini can now create documents, sheets and slides, run research, draft emails, schedule meetings and create tasks from wherever you already are — Gmail, Drive, Docs, Slides or Chat — without switching apps. Human-in-the-loop confirmation is required for external communications, and it is English-only at launch.
Gemini in Google Sheets reached Android, with suggested prompts for summarising tables and generating insights (gradual rollout from September 9). Admins gained external sharing controls for Gemini Notebook — four options, Off by default, then Trusted Domains, On, and On with public sharing, configurable at domain, OU or group level (from September 10). Context-aware access controls arrived for Gemini Enterprise in the Admin console, applying device-security and location policies at OU or group level (rollout September 8, expected to complete September 15). Google Sheets also doubled its cell limit from 10 million to 20 million.
Cross-app orchestration: Business Standard/Plus, Enterprise Standard/Plus, Education add-ons, consumer AI Pro (excluding scheduling) and AI Ultra, Frontline Plus (scheduling only); gradual rollout from September 2; English only; subject to usage limits | Gemini in Sheets on Android: Business Standard/Plus, Enterprise Standard/Plus, Education AI Pro, consumer AI Pro/Ultra; rollout from September 9 | Gemini Notebook external sharing: all Workspace customers, default Off | Context-aware access for Gemini Enterprise: Enterprise Standard/Plus, Education Standard/Plus, Frontline Standard/Plus, Enterprise Essentials Plus, Cloud Identity Premium; requires a Gemini Enterprise purchase
Best for: Workspace admins — three of these four are admin controls, and the Gemini Notebook sharing default being Off means external sharing is a decision you make rather than one you inherit.
★ Gemini Enterprise — Projects reach GA, and pay-as-you-go opens up
Gemini Enterprise Projects went generally available on September 4, letting users create dedicated knowledge bases and collaborate in them, alongside GA for the Monday federated data store covering boards, items, updates and docs. New data stores for CoinDesk, Relativity, Streak and Superhuman Mail arrived in preview.
On September 10, the pay-as-you-go edition and AI developer tools became available to all invoiced Cloud Billing accounts, lifting a previous restriction, and Slack app channel mentions and multi-turn conversations reached GA. Semantic search for agents in Google Cloud Marketplace entered preview on September 8. Earlier in the window, agent observability added latency and error-rate views (GA) and Workflow Builder — formerly Agent Designer — reached GA with multi-step automation.
A breaking change landed on September 9: VPC Service Controls now block website URL ingestion in Gemini Notebook Enterprise. If your notebooks pull in web sources and your perimeter uses VPC-SC, that path stops working rather than degrading.
Best for: Google Cloud teams standardising on Gemini Enterprise. Check the VPC Service Controls change first — it is the one that breaks something that was working.
Plans and pricing
Nothing was added to the Gemini API changelog between September 4 and September 11. The most recent entries sit just outside this window: Lyria 3.5 entered public preview on September 3 for full-length song generation with 44.1 kHz stereo output, and Gemini 3.8 Flash reached general availability on September 2 as Google's most intelligent Flash model, built for long-horizon software engineering and autonomous agents.
Gemini API changelog entries dated September 4–11: none | September 3: `lyria-3.5` public preview, text and image inputs, 44.1 kHz stereo | September 2: `gemini-3.8-flash` GA | Gemini 3.7 Flash remains at its introductory price through December 31, 2026
Best for: Developers on Gemini 3.7 Flash should note the introductory pricing has an end date — December 31, 2026 — and model the rate that follows it now rather than in December.
Microsoft Copilot
★ GPT-6 Astra is generally available in GitHub Copilot
GPT-6 Astra became generally available in GitHub Copilot on September 4, selectable in Visual Studio Code, Visual Studio, the Copilot CLI, the coding agent, the GitHub Copilot app, github.com, GitHub Mobile, JetBrains IDEs, Xcode and Eclipse. It is billed at provider list pricing under usage-based billing, with no premium request multiplier stated.
Plans: Copilot Pro+, Max, Business, Enterprise | Surfaces: VS Code, Visual Studio, Copilot CLI, coding agent, Copilot app, github.com, GitHub Mobile, JetBrains, Xcode, Eclipse | Billing: provider list pricing, usage-based | Admin control: model policy in Copilot settings — new models are enabled automatically unless an admin has turned off the global default or explicitly disabled the model
Best for: Copilot admins who care about spend should know the default: new models turn themselves on unless the global default is off. That is a policy decision worth making deliberately rather than by omission.
★ Enterprise controls — managed permissions for agent operations, and a managed sandbox in JetBrains
Enterprise managed permissions for agent operations reached general availability on September 9. Administrators can centrally control which agent operations are blocked, which require human approval and which can proceed without a prompt, across shell commands, file reads and edits, and network domains. The constraint that makes it worth deploying: managed restrictions cannot be weakened by user or workspace settings, by auto-approval, or by previously saved approvals.
On September 8, an enterprise-managed sandbox arrived in Copilot for JetBrains in public preview. Admins can centrally configure sandbox enablement, filesystem and network access, proxy settings, developer-tool access and macOS Keychain access; managed policies override user settings and the IDE locks the affected controls. It ships with an enterprise policy diagnostics tool so developers can verify that policies are being detected and enforced locally.
Managed permissions: GA in the GitHub Copilot app, Copilot CLI and VS Code sessions using Agent Host; covers shell commands, file reads and edits, network domains; Copilot Business and Enterprise | JetBrains sandbox: public preview; surfaces under GitHub Copilot → Sandbox when the `Editor Preview` feature flag or a managed setting is enabled; includes enterprise policy diagnostics
Best for: Security teams rolling out coding agents. Until now, a determined developer could widen their own permissions; managed restrictions close that.
★ The September 7 weekly release — Jira, recurring agent tasks, voice mode
The weekly bundle brought Jira issues into the GitHub Copilot app, where they land on a shared canvas for investigation and implementation. VS Code 1.137 added recurring agent tasks in public preview — automations scheduled hourly, daily, weekly or on demand.
Also experimental in VS Code 1.137: a voice mode that lets you talk to, interrupt or redirect Copilot while it works, and agent issue review, which shows issue and pull-request detail in the Agents window without opening the repository. In the Copilot CLI, Project HydraFusion is an experimental semantic router that picks between local, cloud and compound models to balance performance, cost and latency.
Jira integration: GitHub Copilot app | Recurring agent tasks: VS Code 1.137, public preview, hourly/daily/weekly/on demand | Voice mode: VS Code 1.137, experimental | Agent issue review: VS Code 1.137, experimental | Project HydraFusion: Copilot CLI, experimental, semantic routing across local, cloud and compound models
Best for: Teams whose work starts in Jira, and anyone who wants scheduled agent runs without wiring up their own cron.
⚠ MAI-Code-1-Flash deprecated
MAI-Code-1-Flash was removed from all GitHub Copilot experiences on September 10, including Copilot Chat, inline edits, ask and agent modes, and code completions. The replacement is MAI-Code-1.1-Flash, which has native vision support and, on Microsoft's own figures, a 73% lower list price than the model it replaces.
Update any workflow or integration that names the retired model. Copilot Enterprise administrators have to enable access to the replacement through their model policies in Copilot settings — the migration is not automatic if the policy does not permit the new model. A further deprecation is already scheduled for October 2, 2026: Gemini 3.5 Flash and Gemini 3.6 Flash give way to Gemini 3.8 Flash, Kimi K2.7 Code to Kimi K3, and Claude Opus 4.7 to Claude Opus 5.
Best for: Anyone with a pinned model in a script or CI job. Check the October 2 list now — four more models go the same way.
★ Code review gets an ensemble, and usage metrics cover the Agents window
Copilot code review now resolves its own comments: when you push a commit addressing review feedback, Copilot resolves that comment during its re-review instead of leaving an outdated thread open. It also suggests contextual commit messages when you apply a fix, rather than generic auto-fill text.
The analysis side changed more than the interface. The review agent now uses the full Copilot SDK shell toolset to validate code — build commands, test execution, script operations — and Lite reviews now run several agents in an ensemble instead of one agent alone. Microsoft's reported result for the ensemble: high-severity findings up 47%, medium up 31%, low up 11%, at roughly 8% lower cost. Separately, Copilot usage metrics now include the VS Code Agents window, adding `daily_active_vscode_agent_users` and `totals_by_vscode_agent` aggregates and a `used_vscode_agent` indicator on user-level reports.
Code review: auto-resolution on re-review, contextual commit-message suggestions, full Copilot SDK shell toolset for validation, multi-agent ensemble on Lite effort | Reported ensemble effect: +47% high-severity, +31% medium, +11% low findings, about 8% cheaper | Usage metrics: new aggregates `daily_active_vscode_agent_users` and `totals_by_vscode_agent`, user-level `used_vscode_agent`; visible to enterprise owners, billing managers, organisation owners and custom roles with View Copilot Metrics, where the usage-metrics policy is enabled
Best for: Engineering leads measuring whether agents are actually being used, and teams who left code review on Lite because Balanced cost more — the ensemble narrows that gap.
Plans and pricing
No new Copilot price changes landed this week, but two dated changes announced earlier now sit close. From October 1, all new Copilot Business or Enterprise seat assignments require payment for each seat before users gain access, with upfront charges for all assigned seats at the start of the billing cycle and no prorated refunds when a seat is revoked; prices themselves are unchanged. And from September 28, the default code review effort level moves from Lite to Balanced unless an admin explicitly selects Lite before that date.
October 1, 2026: per-seat payment required before access on new Business/Enterprise seat assignments; upfront charges at billing-cycle start; no prorated refunds on revocation; list prices unchanged | September 28, 2026: code review default effort moves Lite → Balanced unless an admin selects Lite first | No earlier than September 28: Copilot Cloud Agent, github.com Chat and GitHub Mobile Chat consolidate under a single policy, and chat data retention extends from 28 days to account lifetime — admins who opt out lose Copilot access on github.com and Mobile | Microsoft's own Microsoft 365 Copilot release notes have published no new block since August 25
Best for: Copilot admins have two dates to put in the calendar this month, and the September 28 one has a default that changes underneath you if you do nothing.