OpenAI has outlined new measures to help businesses manage the cost of using its artificial intelligence tools. The changes focus on model efficiency, pricing options and administrative controls for enterprise customers.
Australian companies are among those under growing pressure to expand AI use across their organisations without losing sight of operating costs. The shift reflects a broader market move as AI budgets are treated less as experimental technology spending and more as part of day-to-day operations.
At the centre of OpenAI's approach is GPT-5.5, which it says is more efficient than GPT-5.4 on its GDPVal benchmark. OpenAI said GPT-5.5 scored 84.9% on the benchmark, which covers real-world knowledge work across 44 occupations, and can achieve similar results using fewer tokens.
That matters because token consumption remains one of the main drivers of enterprise AI costs. If a model completes the same task with fewer tokens, a company may cut overall operating costs even if per-token prices are higher.
OpenAI also cited results from Datacurve's DeepSWE benchmark, which measures long-horizon software engineering tasks. According to the company, GPT-5.5 recorded the highest pass rate at 70% and was the most token-efficient model tested, while GPT-5.5 and GPT-5.4 were identified as the most cost-efficient configurations.
Pricing options
OpenAI is also leaning on pricing structures tailored to different kinds of usage. Subscription seats are aimed at organisations that want predictable access and steadier monthly spending, while token-based pricing is positioned for customers with heavier or less predictable workloads.
Some businesses are expected to combine both approaches, particularly where software engineering, customer support and research teams have different demand patterns. That hybrid model mirrors how many large organisations already manage software spending across fixed licences and consumption-based services.
OpenAI recently set GPT-5.5 API pricing at USD $5 per million input tokens and USD $30 per million output tokens. It also offers Batch and Flex pricing at lower rates, alongside priority processing for customers seeking faster performance.
Codex focus
Another part of the strategy is Codex, OpenAI's coding tool, which it says improves efficiency by giving the model the context, tools and review paths needed to reduce retries and avoid unnecessary processing. OpenAI said those workflow improvements can compound gains made at the model level.
The company cited internal examples to show how that can affect time and cost. Its finance team used Codex to process 24,771 K-1 tax forms covering 71,637 pages, which OpenAI said cut completion time by about two weeks compared with the previous year.
In another example, the communications team used Codex to analyse speaking-request data and automate the handling of low-risk requests. OpenAI did not provide financial figures for those projects, but the examples illustrate the kinds of administrative and back-office tasks where companies are increasingly testing AI tools at scale.
Beyond the product itself, OpenAI said its AI Deployment Engineers work directly with customers on evaluations, architecture, latency, reliability and workflow design. That work is intended to improve both performance and cost efficiency as businesses move from pilot deployments to wider internal use.
Governance demand
The most immediate issue for many enterprise customers is control. As AI tools spread across departments, finance and technology leaders want clearer visibility into where spending is happening, how budgets are being used and when usage needs approval or intervention.
That has led to rising demand for governance features that resemble controls already common in software-as-a-service management platforms. Department-level visibility, user-level spending controls, budget monitoring and approval workflows are becoming more important as AI deployments expand beyond small trial groups.
OpenAI said it is developing more in-product controls designed to prevent cost overruns before they happen. These include more granular administrator visibility, easier user-level caps and overrides, clearer budget tracking for users and a way for users to request additional access when needed.
Early testing with alpha customers is already under way, with the aim of giving organisations more visibility into AI consumption before unexpected spending occurs.
Kaylin Voss, Head of Americas Enterprise and Industries at OpenAI, outlined the company's thinking in a LinkedIn post. "Businesses need a better way to manage AI spend with confidence. As companies deploy AI more broadly, leaders need visibility, predictability, and controls that enable broad access while maintaining a clear view of spend and keeping investment within intended limits," Voss said.
She said the company is trying to address the issue through several levers. "A few ways we're thinking about this: Efficiency: GPT-5.5 packs more intelligence into each token and is our most efficient frontier model: over 60% more efficient than GPT-5.4 on GDPVal. We're continuing to improve the amount of useful work customers can accomplish per dollar spent," Voss said.
Voss also pointed to workflow design and deployment support. "But everyone's workflow is different. Our AI Deployment Engineers work directly with customers on evals, architecture, latency, reliability, and workflow design to improve both performance and cost efficiency," she said.
On pricing and administrative controls, Voss said: "Subscription seats provide predictable access, including usage, governance, and centralised management. Customers with heavier or more variable workloads can choose token-based pricing options, giving them greater flexibility to scale usage based on their needs. Customers can choose the structures that best meet their needs."
She added: "We're building more in-product controls to help customers prevent cost overruns before they happen, including more granular admin visibility, easier user-level caps and overrides, clearer budget tracking for users, and the ability to request more access when needed. We've already received strong feedback from alpha partners, and we look forward to rolling these capabilities out more broadly soon."