OpenAI released GPT-5.4 today. It’s available in ChatGPT, the API, and Codex. If you’re running enterprise workloads on OpenAI models — or evaluating whether you should be — this is a significant release that deserves your attention.
Here’s what shipped, what it means for your organization, and what you should do next.
What Actually Shipped
GPT-5.4 is OpenAI’s new flagship model, unifying their best capabilities across reasoning, coding, agentic workflows, and computer use into a single model. There’s also a GPT-5.4 Pro variant for maximum performance on complex tasks.
The headline capabilities:
- Native computer-use: GPT-5.4 is OpenAI’s first general-purpose model that can operate computers directly — navigating desktops, clicking through applications, and executing multi-step workflows across software systems. It scores 75% on OSWorld-Verified, surpassing human performance (72.4%).
- 1M token context window: Up to one million tokens of context in the API, enabling agents to plan, execute, and verify tasks across long-running workflows.
- Tool search: Instead of stuffing every tool definition into the prompt, GPT-5.4 can dynamically look up tools when needed. This reduced token usage by 47% in testing while maintaining accuracy.
- Improved visual perception: New “original” image input detail level supports full-fidelity images up to 10.24M pixels. Better document parsing, better click accuracy, better image understanding.
- Significantly more token-efficient: Uses fewer reasoning tokens than GPT-5.2 to solve the same problems, translating directly to lower costs and faster responses.
What’s New vs. GPT-5.2
If your organization standardized on GPT-5.2 (released late last year), here’s what the upgrade gets you:
Professional work quality jumped dramatically. On GDPval — which tests AI performance across 44 occupations on real work products like presentations, spreadsheets, and schedules — GPT-5.4 matches or exceeds human professionals 83% of the time, up from 70.9% for GPT-5.2. That’s not an incremental improvement. That’s a different class of capability.
Spreadsheets and presentations got a specific focus. On investment banking-style spreadsheet modeling tasks, GPT-5.4 scores 87.3% vs. 68.4% for GPT-5.2. Human raters preferred GPT-5.4’s presentations 68% of the time over GPT-5.2’s output. If your teams use AI for document generation, this matters.
Hallucinations dropped meaningfully. Individual claims are 33% less likely to be false, and full responses are 18% less likely to contain any errors compared to GPT-5.2. For enterprise use cases where accuracy is non-negotiable — legal, financial, compliance — this is the metric that matters most.
Computer use went from experimental to production-ready. GPT-5.2 scored 47.3% on OSWorld desktop navigation tasks. GPT-5.4 scores 75%. That’s not an incremental step — it’s the difference between a demo and a deployable capability.
Agentic web search improved significantly. BrowseComp scores jumped from 65.8% (GPT-5.2) to 82.7% (GPT-5.4). Agents can now more persistently search across multiple rounds to find needle-in-a-haystack information.
Enterprise Pricing: What You’ll Pay
OpenAI’s pricing structure for the GPT-5 family remains competitive. Based on current published rates:
| Model | Input (per 1M tokens) | Cached Input | Output (per 1M tokens) |
|---|---|---|---|
| GPT-5.2 | $1.75 | $0.175 | $14.00 |
| GPT-5.2 Pro | $21.00 | — | $168.00 |
| GPT-5 / 5.1 | $1.25 | $0.125 | $10.00 |
| GPT-5 Mini | $0.25 | $0.025 | $2.00 |
| GPT-5 Nano | $0.05 | $0.005 | $0.40 |
GPT-5.4 pricing hasn’t been separately broken out yet, but expect it to land at or near GPT-5.2 rates. The key efficiency story here is that GPT-5.4 uses fewer reasoning tokens to reach the same answers, so your effective cost per task should decrease even if per-token rates stay flat.
For ChatGPT subscriptions: Plus remains $20/month, Team is $25-30/user/month, and Enterprise is custom pricing with unlimited GPT-5 usage, SOC 2 compliance, SSO, zero data retention, and up to 128K token context windows per session.
The Batch API continues to offer 50% savings for async workloads, and the new priority processing option gives you reliable high-speed performance on a pay-as-you-go basis.
Azure OpenAI Availability
This is where it gets important for regulated enterprises. Azure OpenAI already lists GPT-5.2 and GPT-5.2 Codex across Global and Data Zone deployments. The full GPT-5 family (including GPT-5, 5.1, mini, nano, and Pro variants) is available on Azure with matching token pricing to OpenAI direct.
Data Zone deployments are the key differentiator for enterprises with data residency requirements. Azure lets you pin your model deployment to specific geographic regions — something you can’t do with OpenAI’s API directly.
If you’re already on Azure OpenAI, expect GPT-5.4 availability to follow the same pattern as previous releases: Global deployment first, Data Zone shortly after. Watch the Azure OpenAI model deprecation and availability pages for exact dates.
For organizations not yet on Azure OpenAI: the enterprise controls (RBAC, private networking, content filtering, abuse monitoring) remain the primary reason to pay the Azure premium over going direct to OpenAI.
The Computer-Use Elephant in the Room
Let’s talk about what matters most in this release: native computer use.
GPT-5.4 can operate desktop environments, navigate browsers, click through applications, and execute multi-step workflows — all through screenshots and keyboard/mouse commands. It can also write Playwright code to automate browser interactions programmatically.
For enterprise IT, this creates both enormous opportunity and significant risk:
The opportunity: Automating complex workflows that span multiple applications — the kind of tasks that RPA tools have struggled with because they’re brittle and break when UIs change. An AI model that can see the screen and adapt is fundamentally different from scripted automation.
The risk: An AI agent with the ability to click through your applications has the same access as the user account it’s running under. Think carefully about:
- What credentials and permissions these agents will run with
- How you’ll audit what they do
- What confirmation policies you’ll enforce (OpenAI now supports custom confirmation policies in the API)
- Whether your DLP and security tooling can see and control AI-driven computer interactions
OpenAI explicitly designed GPT-5.4’s computer-use behavior to be steerable via developer messages, with configurable safety policies. Use them. Don’t deploy computer-use agents with broad permissions and no guardrails.
Tool Search Changes the MCP Conversation
If your organization has been building or evaluating MCP (Model Context Protocol) server integrations, GPT-5.4’s tool search capability is directly relevant.
Previously, every tool definition had to be stuffed into the model’s context upfront. With dozens of MCP servers, that could mean tens of thousands of tokens per request — expensive, slow, and context-polluting.
Tool search lets the model dynamically discover and load tool definitions only when needed. In OpenAI’s testing across 36 MCP servers, this reduced total token usage by 47% with no accuracy loss.
What this means for your architecture: You can now expose much larger tool ecosystems to your AI agents without worrying about context window bloat. The practical limit on how many integrations an agent can access just expanded significantly.
What IT Teams Should Do Now
Here’s your action list:
1. Evaluate GPT-5.4 Against Your Current Workloads
If you’re on GPT-5.2, run your existing prompts and workflows against GPT-5.4. Focus on accuracy improvements and token efficiency. The hallucination reduction alone may justify the switch for high-stakes use cases.
2. Assess Computer-Use Readiness
If you have workflows that involve navigating multiple applications — data entry, report generation, cross-system reconciliation — GPT-5.4’s computer-use capabilities could be transformative. But start with a security review, not a POC. Define your permission model and audit requirements first.
3. Revisit Your Tool Integration Architecture
If you’ve been conservative about how many tools you expose to AI agents because of context window costs, tool search changes that calculus. Re-evaluate your MCP server strategy with the 47% token reduction in mind.
4. Update Your Model Lifecycle Plan
OpenAI is shipping new model versions roughly quarterly now. If you don’t have a model evaluation and migration process, build one. Pin to specific model versions in production, test new versions in staging, and plan for deprecation cycles.
5. Review Azure OpenAI Deployment Options
If you’re using OpenAI’s API directly and you have data residency, compliance, or network security requirements, this is a good time to evaluate Azure OpenAI. The model availability gap has narrowed significantly, and the enterprise controls are worth the operational overhead.
6. Budget for the Shift
GPT-5.4 is more efficient per task, but the capabilities it unlocks — computer use, massive tool ecosystems, million-token context — will drive more usage. Plan your token budgets accordingly. The Batch API and cached input discounts are your best friends for cost management.
The Bottom Line
GPT-5.4 isn’t just another point release. The combination of native computer use, tool search, 1M token context, and dramatically improved professional work quality makes this a capability inflection point.
The models are now good enough to do real professional work across real applications with real accuracy. The question for enterprise IT is no longer “can AI do this?” — it’s “do we have the governance, security, and architecture to let it?”
Start there. The technology is ready. Make sure your organization is too.