March's first major AI release arrived sooner than expected.
Overnight, OpenAIGPT-5.4introduced GPT-5.4, a genuinely unified model that combines reasoning, coding, computer use, deep research, and a million-token context window.
OpenAI describes it asa model designed for professional work.That positioning is deliberate.
The model launched simultaneously across ChatGPT, the API, and Codex.
In ChatGPT, it appears as GPT-5.4 Thinking, with GPT-5.4 Pro available for higher-performance workloads.
For the past several years, the dominant way of using large language models has remainedquestion-and-answer collaboration.。
You ask a question and receive an answer. You provide a prompt and the model tries to explain or produce what you intended.
Even many more sophisticated applications simply place a process aroundthe same conversational capability.The wrapper changes, but the underlying interaction does not.
Real work is different.
A typical professional assignment may require a much longer chain:
review a large body of material, research several rounds of web sources, call tools, interpret spreadsheets, documents, and screenshots, then operate websites or software directly.At the end, it still needs to produce a usable deliverable.
The central change in GPT-5.4 is its ability to cover more of that chain within one model.
OpenAI's launch materials emphasize three categories:knowledge work, native computer use, and advanced programming and debugging.
Together, they clarify the model's real product position.
This is more than a routine model upgrade.
It is OpenAI's latest answer to the question of how AI should enter real workflows and contribute across the full path to delivery.
Closer to a system that gets work done
One of OpenAI's most important published metricsGDPval。
comes from an evaluation designed around realistic professional assignments rather than conventional academic questions.
It spans knowledge work across 44 occupations.
On that benchmark,GPT-5.4 scored 83.0%.
GDPval requires models to complete clearly specified knowledge-work tasks.
The assignments reflect deliverables from 44 occupations across the nine industries that contribute most to US GDP, including sales presentations, accounting spreadsheets, emergency-department schedules, manufacturing diagrams, and short-form video.
OpenAI also published internal evaluations closer to everyday office work.
GPT-5.4 scored 87.3% on investment-banking modeling, compared with 68.4% for GPT-5.2.
In a human-reviewed presentation-generation test, reviewers preferred GPT-5.4's output 68% of the time.
The important point behind these figures is straightforward:
models are becoming more reliable contributors to real deliverables.
Previous claims that a model could write a report, build a spreadsheet, or create a plan often carried the feel ofa polished demonstration.。
The output looked impressive at first glance, but when it came time touse it in an actual business setting, something was still missing.。
Sometimes the logic failed, sometimes the formatting did, and sometimes the last step from 'presentable' to 'deliverable' was absent.
GPT-5.4 is clearly targeting that final gap.
For content teams, this meansbetter support for research and synthesis across long documents.
For strategy, finance, and consulting teams, it meansstronger spreadsheet, presentation, and document understanding and creation.
For business teams, it moves the model one step closer fromoffering recommendations to producing usable work.That is a meaningful product shift.
Computer use becomes a core capability
The most widely discussed improvement is that GPT-5.4 is OpenAI's first mainstream general-purpose model with native computer-use capabilities.
OpenAI's description is direct: it can interpret screenshots, understand interfaces, click with the mouse, and enter text with the keyboard.
Why does that matter?
Because many digital tasks break down inthe last mile of execution.。
- A model that explains how to complete a form and one that actually completes it belong to different product categories.
- A model that explains how to schedule a meeting is not equivalent to one that creates the event and sends the invitations.
- Likewise, describing how to navigate a website is fundamentally different from completing the workflow directly.
By bringing this capability into the core model, GPT-5.4 makes the agent concept feel more like a complete execution system.
GPT-5.4 can interpret screenshots of a browser interface and interact with UI elements through coordinate-based clicks, enabling tasks such as sending email and scheduling calendar events. The demonstration video runs at normal speed.
On OpenAI's published benchmarks, the model scored 75.0% on OSWorld-Verified, 67.3% on WebArena-Verified, and 92.8% on Online-Mind2Web.
Those figures point to a practical shift:the model can increasingly move through software interfaces and complete tasks directly.
That is why many observers evaluate GPT-5.4 and OpenClaw alongside other agent frameworks of this kind.
The reason is simple.When task understanding, coding, tool use, visual interface interpretation, and UI interaction converge in one model,AI becomes substantially more useful as an execution system.
Coding and world knowledge finally converge
Another important part of the release is easy to overlook:
GPT-5.4 incorporates the frontier coding capabilities of GPT-5.3-Codex into the mainstream model.
That may sound like an incremental upgrade, but it has significant implications.
Many developers have experienced the same trade-off in recent models.
- Specialized coding models can be excellent at implementation but struggle to align with people, business context, and complex multi-part objectives. They oftenexecute well but communicate poorly.；
- General-purpose models understand context more naturally but can fall short on demanding engineering work.
GPT-5.4 attempts to connect those two strengths.
In OpenAI's published results,
it scored 57.7% on SWE-Bench Pro, slightly above GPT-5.3-Codex at 56.8%.
It also adds stronger frontend construction and visual debugging workflows, including experimental capabilities such as Playwright Interactive. The objective is clear: the model should generate code, test it, inspect the result, and improve it in the same loop.
For developers, the practical benefit is that
model selection may no longer require a constant choice betweenthe strongest coding model and the strongest model for complex general tasks.GPT-5.4 may not lead every competitor on every individual metric, but it combines
strong coding performance with more complete task understandinginside a single workflow.That matters for agents, automated development, and complex product delivery.
A million-token context window and tool search
A million- token context window and tool search
GPT-5.4's unified capabilities are the obvious headline, but two additional improvements
deserve particular attention from developers and enterprise teams.
1.A one-million-token context window
GPT-5.4 supports up to 1.05 million tokens of context, allowing it to process larger codebases, broader document collections, and longer agent trajectories in one run.
What changes as a result?
Work that previously required constant chunking, compression, and loss of context can now proceed with greater continuity.
The improvement should be tangible in code review, long-document analysis, and complex workflow tracking.
2. Tool Search
OpenAI has also introduced a tool-selection mechanism that behaves more like looking up a definition when it is needed.The model no longer requires every detail upfront.Previously, systems with many tools often had to place every tool definition in the context window.
That approach was both expensive and slow.Now the model can inspect a tool list and retrieve a detailed definition only when a specific tool is required.；
OpenAI reports that this mechanism reduced total token use by 47% while maintaining the same accuracy in its evaluation.
It may look like a minor optimization, but it addresses a critical production concern.
Many agent projects become costly, slow, and unstable not because
the model lacks the capability, but because the tool context has become too large and unwieldy.Teams that solve this layer effectively will have a much easier path from agent prototype to business deployment.
Greater control during execution
In ChatGPT, GPT-5.4 Thinking now begins longer, more complex requests with a short preamble that outlines its working plan.Users can add instructions or redirect the work while the response is being generated.
This makes it easier to steer the model toward the intended result without restarting or adding several corrective turns.
More capable,and more expensive
GPT-5.4 API pricing is $2.50 per million input tokens and $15 per million output tokens.
GPT-5.4 Pro rises to $30 for input and $180 for output.
The unit price of GPT-5.4 is higher than GPT-5.2.
OpenAI repeatedly emphasizes, however, that capability and efficiency have improved together.
The practical point is that cost cannot be judged solely by the price per million tokens. Teams must measure the total tokens required to complete the same outcome.A model with a higher unit price may still cost less overall if it uses a shorter reasoning path, calls tools more efficiently, compresses context more effectively, and requires fewer retries.
In that case,the total cost may not be higher at all.
This is another way GPT-5.4 resembles a production model: it treats capability, speed, and cost as parts of the same product equation.
The Pro price remains striking, of course.
That version appears designed for exceptionally complex, high-value work,not as a default model for everyday conversation.
Teams adopting the family at scale will need a more deliberate model-routing strategy.
They should define which workloads justify GPT-5.4, which can use a lighter model, and where Pro should appear only at critical decision points.
The rise of digital workers
GPT-5.4 is positioned asan agent foundation capable of completing real work.
That reflects a broader shift.
Models were previously compared by academic domain.
Which one was strongest in mathematics, coding, vision, or search?
Now they are increasingly competing at the level of complete units of work.
The model that can carry an assignment from start to finish is closer to creating real value.
AI product competition is moving from isolated capability benchmarks to workflow integration.
Final thoughts
As a conventional model review, the GPT-5.4 story would be straightforward:stronger knowledge work, native computer use, more complete coding, a million-token context window, tool search, and fewer hallucinations.
For teams responsible for products, systems, and implementation, a more important question follows:what kind of software will these capabilities create next?
AI is evolving from a tool that provides recommendations into a system that participates in execution.
Once a model acts directly, intelligence is only one part of the problem. Teams must also control permissions, connect tools, govern data flows, preserve audit trails, decide which tasks can be delegated, and keep people accountable at critical points.
That is why GeekOnUp focuses onthe interaction between business, data, and AI,rather than evaluating the model in isolation.
Business requirements define what the system should achieve. Data determines what it can see. Product and engineering controls determine what it is allowed to do. The model is the intelligence connecting those layers.
For the past several years, the industry has trained models to answer questions. The next challenge may belearning how to bring models into real work responsibly.
GPT-5.4 is a clear signal that this next phase has begun.
