What this review covers
This review starts from our editorial baseline of 6–8 September and closes on 23 September 2026. It also fills earlier gaps, including Gemini Spark and Claude’s integrated browser. Each detailed guide lists its sources and review date; it does not assume that a future rollout has already happened. The examples are suggested procedures, not claimed customer outcomes. We focus on businesses with roughly 5–250 employees. Entertainment features and region-specific consumer finance products are outside the main scope.
OpenAI: new models and changes to daily work
Sol and Luna joined Work and Codex on 22 September, separately from the model selection in Chat. Word, business data, Library and voice interaction also become more relevant. Start with existing dependencies: which employees rely on custom GPTs, which tasks require plugins and which files come from connected accounts? Custom GPT migration has been announced, but the available migration path needs checking in the actual workspace. GPT-5.5 retirement in ChatGPT, Work and Codex must not be interpreted as an API retirement.
Claude: connected work across documents and projects
Anthropic introduced Opus 5.5 and is bringing Claude and Cowork together. Docs, Slides, redesigned Projects and business integrations extend the workspace. The useful question for a small business is whether existing sources can become an output that is straightforward to verify. Choose a recurring report with a fixed structure and check its sources, figures, formatting and editing effort. Some capabilities start in beta or roll out gradually. Assess Smart Reports and Marketplace separately from everyday employee features because they involve different administrative roles and procurement decisions.
Google: separate product names, accounts and regions
Gemini 3.8 Flash, Live and TTS serve different purposes. Notebooks inside the Gemini app and the standalone Gemini Notebook, formerly NotebookLM, are also distinct products. Our guides keep these names separate. For a business in Germany, an announcement for personal accounts does not establish Workspace or EEA availability. We therefore do not recommend Spark as generally available locally. For Meet, we distinguish the announcement date from a default setting that takes effect later. Developers should prioritise Antigravity migration according to the APIs they actually use.
Microsoft: model choice does not replace access review
Copilot expands model choice and the tools for creating apps, presentations and team assistants. A Microsoft 365 business can start with a well-maintained SharePoint area or a recurring presentation. Check licensing, administrator approval, region and data access separately. A model being named in an announcement does not prove it is enabled in your tenant. A Brand Kit can help with visual consistency, while factual accuracy and a suitable source template remain the team’s responsibility.
How to choose your next pilot
Start by listing the models, assistants and connections used in production. Upcoming retirements take priority over optional new features. Then choose a frequent task whose result can be checked, such as a weekly report, proposal preparation or research in approved documents. Define inputs, expected output, handling time and failure cases before the trial. Compare the result with the existing process, including human rework. Adopt the change once access, quality, ownership and a fallback procedure are clear.
Common questions about cost and adoption
Do we need new licences immediately? No. First establish what your current plan and administrators already allow. API usage, additional subscriptions and connected business systems may be billed separately. A low model price says little about total cost once setup, review and operation are included. Can this overview establish access to every feature? The guides support selection, but availability and permissions still need checking in your own account. The AI System Check helps identify a useful next task based on how your business already works.
Decision matrix
| Work area | What to assess | A useful first test |
|---|---|---|
| OpenAI | Sol/Luna, Word, plugins and APIs; distinguish Chat from Work | Inventory an existing GPT and test its successor on identical tasks |
| Anthropic | Opus 5.5, the unified interface, Docs/Slides and Projects | Create a report from approved sources and trace each central claim |
| Gemini 3.8, notebooks, Workspace and regional restrictions | Evaluate a defined knowledge collection with a known expected result | |
| Microsoft | Copilot models, Cowork, Studio, Brand Kit and Agent Builder | Test a team assistant with two accounts that have different permissions |
Keep it verifiable
Primary sources
- OpenAI: ChatGPT release notesSource checked:
- Anthropic: Claude Opus 5.5Source checked:
- Claude: Cowork is now ClaudeSource checked:
- Google: Gemini API changelogSource checked:
- Microsoft: More models, one CopilotSource checked:
