GPT-5.4 Goes Native: Enterprise Excel/Sheets Plugins, Tool Search, and a New Era of AI Workflows

The AI news cycle has kicked into another gear as OpenAI unveils GPT-5.4 in two varieties: GPT-5.4 Thinking and GPT-5.4 Pro. This follows the recent GPT-5.3 Instant and marks a shift toward agents that can act across tools and desktop apps, not just generate text. The headline capability is native computer use — the model can operate a computer via libraries like Playwright and even issue mouse and keyboard actions in response to screenshots. OpenAI is also shipping a new suite of Excel and Google Sheets integrations that allow GPT-5.4 to model complex financial data and automate spreadsheet tasks. In addition to the interface changes, the company highlights token efficiency and a massive context window up to a million tokens for API and Codex users, enabling longer, multi-step workflows. Pricing and access are tiered across API, Codex, and ChatGPT plans, with free users getting occasional access when queries are auto-routed to GPT-5.4.

The two model paths reflect different user needs: GPT-5.4 Thinking will be available to all paid ChatGPT subscribers, while GPT-5.4 Pro sits behind enterprise-grade plans for organizations that demand deeper automation and integration. The release emphasizes token efficiency, with claims of substantial reductions in token usage on many tasks, and introduces a tool search capability that reduces context pollution by loading tool definitions only when needed. The result is a more scalable foundation for long-running workflows that weave spreadsheet modeling, data retrieval, and code execution into a single orchestration layer.

Beyond the user interface, OpenAI highlights a broader enterprise strategy: richer context windows, more reliable factuality, and tighter integration with software commonly used in finance and operations. The company confirms that GPT-5.4 supports up to one million tokens of context for API and Codex use, enabling agents to plan and verify actions across long horizons. It also notes pricing differentials, including a step where input tokens beyond a threshold may incur higher costs, a design choice aimed at balancing capability with budget discipline for large-scale deployments.

In tandem with this release, the AI ecosystem is advancing in complementary ways that matter for enterprises. Databricks has rolled out KARL, a Knowledge Agents via Reinforcement Learning system designed to handle six enterprise search behaviors at once. Trained on synthetic data without human labeling, KARL claims to match Claude Opus 4.6 on a purpose-built benchmark at lower cost per query and notably lower latency. The approach hinges on a multi-task RL stack that builds a contextual memory, compresses context when needed, and uses an off-policy learning regime to keep training budgets manageable. The takeaway for teams is clear: a single model can generalize across multiple search behaviors, from constraint-driven entity lookups to cross-document synthesis, reducing the need to stitch together ad hoc pipelines.

As these capabilities scale, the industry is seeing a flurry of activity around inference, shipping, and real-world deployments. Perplexity has teamed with CoreWeave to demonstrate enhanced inference throughput, underscoring the ongoing push to run ever larger models with lower latency and cost. Retailers in Australia are piloting agentic shopping assistants that promise to plan meals, assemble shopping lists, and manage purchases, while regulators in Europe probe privacy aspects of AI wearables like smart glasses from Meta. Into this mix enters a broader infrastructure narrative, with talks of underwater data centers aboard offshore wind installations to meet surging compute demand, highlighting how the physical footprint of AI is expanding as rapidly as its capabilities. Taken together, these threads point to a period where AI moves from chat-like interactions to embedded, auditable workflows that touch finance, operations, and daily life.

In this moment, the trend is not simply faster answers but smarter, more accountable action across the entire workday. GPT-5.4’s emphasis on native computer use, on-demand tool definitions, and longer horizons for planning aligns with the real world need to automate multi-step tasks while maintaining control and traceability. Yet as AI becomes more capable inside spreadsheets, dashboards, and enterprise apps, the conversations around governance, privacy, and workforce impact intensify. Industry observers are watching not just for new benchmarks, but for how organizations implement guardrails, monitor outcomes, and ensure that these powerful agents augment human decision making rather than supplant it. For daily readers, today’s mix of model enhancements, enterprise tooling, and infrastructure innovation signals a future where AI is woven into the fabric of work and everyday life, with far-reaching implications for productivity, security, and policy.

Sources

1. OpenAI launches GPT-5.4 with native computer use mode and financial plugins for Excel/Sheets
2. Databricks built a RAG agent it says can handle every kind of enterprise
3. Gemini’s Canvas in AI Mode Available in Google Search in US
4. Underwater Data Center Project Aboard Offshore Wind Turbine
5. ‘Our consciousness is under siege’: Michael Pollan on chatbots, social media and mental freedom
6. Retailers want ‘delightfully human’ AI to do your shopping, but will the chatbots go rogue?
7. Euro Regulators Question Meta Over AI Glasses Privacy Fears
8. Perplexity, CoreWeave Deal Boosts Inferencing
9. Checking your ex’s socials or overusing Find My Friends? Welcome to the age of interpersonal surveillance