Claude Latest Features and Updates: September 2026

September 2026 changed Claude at three levels at once: model capability, work surfaces and safeguards. Anthropic introduced Fable/Mythos 5.1, Opus 5.5 and Sonnet 5.5 while Projects, connectors, Microsoft 365 work and shared-agent patterns continued to expand. The result is a product update story that cannot be reduced to a benchmark table.

That makes the September story larger than a model leaderboard. Sonnet 5.5 changes the everyday cost-performance tier, Opus 5.5 targets harder judgement-heavy work, Fable/Mythos push further into advanced coding and research, while the surrounding product increasingly lets those models operate across real organisational context.

For the evergreen product overview, use our complete Claude AI guide for 2027.

Claude Sonnet 5.5 launched on 28 September

Anthropic introduced Sonnet 5.5 as a faster, lower-cost complement to Opus 5.5. The company says it runs more than 30% faster than Sonnet 5 and can cost up to 30% less per task because it uses fewer tokens.

Anthropic positions it for well-scoped everyday tasks, coding and polished document, slide and spreadsheet work. The model also launched with stronger cybersecurity safeguards because its cyber capability moved closer to Anthropic’s higher-end models.

Claude Opus 5.5 launched on 22 September

Opus 5.5 is Anthropic’s deeper model for complex work requiring sustained judgement. Anthropic says it reaches a similar level to Fable 5.1 on many tasks at lower cost.

The distinction between model tiers is becoming more about workflow economics: routine or high-volume jobs can use faster models while open-ended high-stakes work can justify deeper reasoning.

Fable 5.1 and Mythos 5.1 arrived on 1 September

Anthropic introduced Fable and Mythos as higher-end models for advanced coding, knowledge work and research. The company framed their scientific capabilities as an early view of how AI could contribute to research.

These models also matter for safety because Anthropic applies additional controls to particularly capable systems and certain high-risk domains. TechnologyBlog’s Fable/Mythos launch analysis explains why Anthropic made the same underlying model broadly available as Fable while keeping Mythos behind trusted-access programmes.

Chat and Cowork began merging on 16 September

Anthropic began rolling out a unified Claude experience to Pro and Max users on 16 September. In the new interface, quick questions and longer agentic tasks start from the same conversation, so users no longer have to choose chat or Cowork before describing the job.

The rollout is staged, and Team and Enterprise organisations keep the existing Cowork presentation for now. That makes account type and rollout stage part of the current Claude story: two users can legitimately see different controls while using the same underlying capability.

Interactive connectors expanded the app experience

Anthropic’s August 2026 connector update lets supported apps render live interfaces inside Claude conversations. Users can interact with dashboards, task boards and other controls without leaving the conversation.

The current list includes tools such as Asana, Canva, Figma, Slack, Box and analytics products, with more support being added.

Projects are being redesigned around parallel work

Anthropic’s current Projects documentation describes a new beta experience starting with Claude Code. In the new model, a project becomes one larger conversation that can split work into parallel cloud threads sharing the same files, repositories, instructions and memory.

The rollout is staged, so existing Projects continue to work while the newer experience expands.

Claude Tag brought shared agents into Slack

Claude Tag launched in June for Team and Enterprise customers. Teams can tag Claude inside a Slack channel and delegate work using the context and tools available to that shared agent.

Anthropic says its own teams use the internal version for tasks ranging from product metrics to support tickets and code.

Claude Code research showed more end-to-end agent use

Anthropic’s June study of roughly 400,000 Claude Code sessions found usage shifting away from pure debugging and toward more end-to-end work such as deploying code, analysing data and writing documents.

The study also found that human domain expertise still improved success, which is an important counterweight to claims that agentic coding removes the need for experienced developers.

Claude expanded further into Microsoft 365

Anthropic continued expanding Claude across Excel, PowerPoint, Word and Outlook. In financial-services workflows, the company now describes context moving between applications so users do not have to restate the same problem each time.

These integrations show Claude moving closer to the work surface rather than asking employees to move every task into a separate AI app.

Cowork continued evolving into a multi-surface agent

Anthropic’s current Cowork documentation supports web, desktop and mobile task control, plus connectors, Skills, Projects, file previews and scheduled tasks.

Some local file, browser and computer-use features still depend on the desktop application, which means availability can differ by surface.

Anthropic published a major misuse report in September

Anthropic’s September threat-intelligence report described attempts to use Claude for influence operations, surveillance, cyber activity, scams, biological misuse and other harmful work.

The company says it banned accounts, improved detections and shared information with relevant partners. The report is important because it shows how Claude’s growing capabilities create both useful and harmful applications. Separate TechnologyBlog reporting on Anthropic’s cyber-evaluation incidents examines the operational containment problem from another angle.

Safety controls are becoming more model-specific

Sonnet 5.5 launched with cyber safeguards closer to those used on Anthropic’s most capable models. Anthropic also operates verification programmes for legitimate advanced cybersecurity and life-sciences users.

This is likely to become a recurring product pattern: capability and access controls moving together rather than every model exposing the same behaviour to every user.

Why September’s changes belong together

The model launches matter because they arrived alongside product changes that determine how the models are used. A faster Sonnet is one story; that model operating inside connected apps, projects or longer-running work is a different operational change.

Readers should separate three facts when following Claude: which model is available, which product surface exposes it, and which plan or safeguard controls the capability. A launch announcement does not prove every account received every surrounding feature on the same day.

For feature selection, see Best Claude Tools and Use Cases. For competitive context, see Claude vs the Leading Alternatives. For what these late-2026 changes could imply next, continue to The Future of Claude in 2027.

Sonnet 5.5 also changes the cost-performance discussion

Anthropic says Sonnet 5.5 keeps Sonnet 5’s token prices while using fewer tokens per task and generating output more than 30% faster. The company positions it as the everyday model for well-scoped work, while Opus 5.5 remains stronger for complex, open-ended tasks requiring sustained judgement.

That is important for enterprise deployments because the practical question is often not “which model is smartest?” but “which model is accurate enough for this workflow at a sustainable cost and latency?”

Interactive app experiences are becoming part of Claude itself

The connector system is no longer limited to retrieving text from external services. Anthropic’s interactive connectors can render live app surfaces inside the conversation, allowing users to inspect or change a task board, dashboard or design while still talking to Claude.

This blurs the boundary between assistant and application shell. It also increases the importance of showing users exactly which connected system is being changed.

Enterprise safeguards are becoming a separate product layer

Anthropic introduced Enterprise Frontier Safeguards work with customers in September 2026 and continued expanding verification programmes for high-capability cyber and life-sciences use. The direction suggests that access to the most capable workflows will increasingly depend on identity, organisation and use case.

That matters for publishers covering Claude because a feature can be technically part of the model but unavailable under ordinary account conditions.