What's New in Claude AI: The 2026 Mid-Year Rundown

What's New in Claude AI: The 2026 Mid-Year Rundown

A practical tour of the model, memory, Skills, Office integrations, and MCP changes that landed through August 2026.

Introduction

Claude has changed shape over the past year. It is no longer just a chat window you visit to ask questions - it now remembers your work, carries reusable instructions between platforms, edits your spreadsheets and decks directly, and plugs into the systems your team already runs on. If you last looked at Claude in 2025, most of what follows will be new to you.

This post walks through the changes layer by layer: the model itself, memory, Skills, the Office and Microsoft 365 surfaces, the connector protocol underneath, and Claude Code. Figures are included to make the structure easier to follow.

1. The model layer: Opus 4.7

The headline model of the moment is Opus 4.7. Beyond raw quality, the most consequential improvement is reliability over long, multi-session work - Opus 4.7 specifically improved the dependability of file-system memory when a task stretches across days rather than minutes. That matters less for one-off questions and a great deal for agent-style work where Claude picks up a project where it left off.

The practical implication is that vague prompting is now the main bottleneck rather than model capability. Ambiguous instructions still produce omissions, so the shift is toward writing durable, structured instructions once and reusing them - which is exactly what Skills are for.

2. Memory: two modes, now on every plan

Memory is probably the change users feel first. It runs in two distinct modes.

Chat memory derives summaries of your past conversations and carries them across sessions, and you can inspect and edit what it has retained under Settings > Capabilities > Memory. It rolled out gradually - Team and Enterprise in September 2025, Pro and Max in October 2025, and the Free plan in March 2026 - so it is now available regardless of which plan you are on.

File-system memory is the agentic counterpart. Claude writes to a /memory folder that is read at the start of each session, with an optional auto-memory mode that lets Claude decide for itself what is worth storing. This is the mode that benefited most from Opus 4.7.

One related governance note worth knowing: an August 2025 policy change extended conversation data retention to five years for users who have not opted out of training. If that matters for your organization, check your settings.

3. Skills became an open standard

Skills are packaged, reusable instructions - your playbook, your house style, your recurring workflow - that Claude loads on demand. The important 2026 development is that Skills are no longer Claude-only. They follow the Agent Skills spec, an open standard, so any platform implementing that spec can run Claude-authored Skills. Team and Enterprise organizations can also deploy Skills across the whole org rather than per-user.

Security tightened alongside that expansion. Skills synced from claude.ai into Claude Code no longer shadow local commands or MCP prompts, their descriptions are sanitized and labeled, and their bodies do not execute shell commands or expand file references on your machine.

4. Claude inside Office and Microsoft 365

This is the biggest surface-area change for non-developers. Claude for Excel launched in beta in October 2025 and was upgraded in February 2026 to run native Excel operations - pivot-table edits, conditional formatting, the things that previously required you to do the clicking. Claude for PowerPoint followed in February 2026.

In March 2026 the Excel and PowerPoint add-ins gained shared conversation context, meaning an action in one application is informed by what happened in the other: build the model in Excel, then ask for the deck in PowerPoint without re-explaining anything. Skills work inside both add-ins too.

Most recently, as of August 2026 the Microsoft 365 connector gained write tools. Claude can now draft, send, and organize email, manage calendar events, update mailbox settings, and create or update files in OneDrive and SharePoint. Read and search continue to work as before - the difference is that Claude can now act, not just look.

5. MCP: the plumbing got serious

The Model Context Protocol is how Claude connects to outside applications, and the 2026-07-28 spec is one of the most significant releases to date. MCP has passed 400 million monthly SDK downloads - roughly four times growth this year - and has effectively become the industry standard for wiring AI agents into software.

Three changes stand out. First, a stateless core: MCP moves from a bidirectional stateful protocol to a plain request/response model, so servers can deploy on serverless and edge infrastructure and scale on standard HTTP. Second, standardized extensions: MCP Apps and Tasks now ship under a versioned extensions framework, giving developers a formal path to add interactive UIs and long-running work without breaking the base protocol. Third, stronger authorization through OAuth and OIDC.

On the connector side, the release also brought embedded UI, enterprise-managed authentication, observability, and private network tunnels - the features that decide whether an integration survives a security review.

6. Claude Code

For engineering teams, the notable addition is self-hosted environments, in public beta on Team and Enterprise plans. Running claude self-hosted-runner turns your own machines or containers into runners; when someone selects your environment from claude.ai, the mobile or desktop apps, or claude --cloud, that session executes inside your network with access to internal services. An Owner or admin enables it first under admin settings, and the environment reports Healthy once a runner registers.

Smaller quality-of-life changes have piled up as well: cross-session messages now show sender and body inline, compaction surfaces retry countdowns and stall hints, Vertex AI credential failures surface in seconds instead of minutes, and newer models can overwrite a file they have not read in the session, matching the Edit tool's rules.

What it adds up to

The through-line across all of these releases is the shift from assistant to operator. Memory gives Claude continuity, Skills give it reusable procedure, the Office and Microsoft 365 integrations give it hands, and MCP gives it reach into everything else. None of these features is dramatic on its own; together they change what you delegate.

If you want a single place to start: turn on memory, write one Skill for the task you repeat most often, and connect one system you actually use daily. That combination delivers most of the benefit described above.

Note on figures: the illustrations in this post are original diagrams summarizing the release information cited above, not screenshots from Anthropic.