GPT-5.4: Office Benchmarks, Excel Add-In and API Pricing

The benchmarks are about spreadsheets, not frontier math

OpenAI's own announcement describes GPT-5.4 as built for professional tasks involving spreadsheets, presentations and documents. It reports a mean score of 87.3 percent on an internal benchmark of spreadsheet modeling tasks styled on the kind of work a junior investment banking analyst might do, up from 68.4 percent for GPT-5.2. [1]

On GDPval, a benchmark spanning well specified knowledge work across 44 occupations drawn from the industries that contribute most to US GDP, OpenAI says GPT-5.4 matched or exceeded professional quality work in 83.0 percent of comparisons, up from 70.9 percent for GPT-5.2. Human raters also preferred GPT-5.4's slide decks over GPT-5.2's 68.0 percent of the time, which OpenAI attributes to stronger aesthetics, more visual variety and better use of image generation. [1]

Atlas interpretation: Model releases usually lead with reasoning or math benchmarks. This one leads with a spreadsheet task modeled on junior banker work and a preference study on slide design. That choice of yardstick is itself the story: it is OpenAI measuring itself against office work rather than against a harder research problem, which is exactly the shift the summary above points to. [1]

The coding work came back into the general model

OpenAI says GPT-5.4 combines the coding strengths of GPT-5.3-Codex, a coding focused model it shipped exactly a month earlier, with knowledge work and computer use in one model, and that it matches or outperforms GPT-5.3-Codex on SWE-Bench Pro: 57.7 percent for GPT-5.4 against 56.8 percent for the dedicated coding model. [1]

Atlas interpretation: A month earlier OpenAI had split its coding capability out into its own model. Folding that same capability back into the general release, and matching the specialist model's own benchmark while doing it, reads as a company still undecided about whether coding belongs in a dedicated product or the flagship, rather than a settled roadmap. [1]

A shipped feature, not only a benchmark chart

eWeek reported that OpenAI paired the release with a beta ChatGPT for Excel add-in, powered by GPT-5.4, that builds and edits spreadsheet models from natural language instructions, asks for user approval before applying a change, and links its outputs back to the specific cells used in each calculation. The beta launched limited to Plus, Pro, Team, Business, Enterprise and Edu subscribers in the United States, Canada and Australia, with Google Sheets support described as coming later. [2]

Atlas interpretation: The spreadsheet benchmark is not just a number on a chart here. OpenAI built a specific product surface to sell against it on the same day, which ties the office work framing to something a buyer can actually open, rather than leaving it as a capability claim with no shipped counterpart. [1][2]

Computer use arrives with a price increase

OpenAI describes GPT-5.4 as its first general purpose model with native computer use: it can write code that drives tools such as Playwright and can also issue mouse and keyboard commands directly from screenshots. OpenAI reports a 75.0 percent success rate on OSWorld-Verified, a benchmark of navigating a desktop environment through screenshots and keyboard and mouse actions. [1]

API pricing rose alongside those claims. Input tokens for gpt-5.4 list at $2.50 per million against $1.75 for gpt-5.2, and output at $15 per million against $14. GPT-5.2 Thinking stays available to paid ChatGPT users under a Legacy Models label for three months, retiring on June 5, 2026. [1]

Atlas interpretation: The price increase lands on the release OpenAI is pitching hardest at cost conscious back office buyers, spreadsheets and slide decks rather than research labs. That is worth noting on its own terms, without assuming it was set to recoup any particular cost. [1]