OpenAI went from GPT-4.5 to GPT-5.4 in thirteen months, acquired two companies for nearly $10 billion, hit 700 million weekly users, and somehow still found time to partner with Disney. Here is what actually shipped, what quietly flopped, and what it all means for the rest of us.
Q1: The Warm-Up Nobody Recognized as a Warm-Up
January 2025 felt quiet by OpenAI standards. The company launched Operator, an autonomous AI agent designed to perform tasks on the web. Browse a site, fill out a form, book a reservation. It was impressive as a demo, underwhelming in daily use, and important primarily as a signal of where OpenAI’s product strategy was heading. Agents, not chatbots, would be the main event.
February brought two releases that looked unrelated but told a connected story. Deep Research debuted on February 2, leveraging the o3 reasoning model to conduct multi-step web research, synthesize sources, and produce structured reports. It was initially limited to ChatGPT Pro subscribers at $200/month. Then, on February 27, OpenAI released GPT-4.5, its largest model yet, positioned as a “research preview.” GPT-4.5 offered improved reasoning and more natural conversation, but it was always a bridge to what came next. Think of it as the soft opening before the grand opening.
The real fireworks in Q1 came in April. On April 16, OpenAI released the full versions of o3 and o4-mini, the company’s dedicated reasoning models. These were architecturally distinct from the GPT line: they “think” before answering, spending extra compute on chain-of-thought reasoning. The o4-mini became the top-scoring model on the AIME 2024 and 2025 math benchmarks. Alongside these, OpenAI quietly shipped the GPT-4.1 family (4.1, 4.1 mini, 4.1 nano), optimized specifically for software engineering tasks.
By March, looking back, Q1 was not a warm-up. It was OpenAI stress-testing every component (reasoning, agents, multimodal, code) that would be unified in GPT-5. Nobody realized it at the time because the releases were presented as separate products rather than parts of a single strategy.
Q2-Q3: GPT-5 and the Great Unification
In May 2025, OpenAI made two acquisitions that signaled the scale of its ambitions. Windsurf, a coding assistant platform, for $3 billion. And io, a data infrastructure company, for $6.4 billion. Nearly $10 billion spent in a single month. This was not a research lab shopping spree. This was a platform company building its stack.
June brought an aggressive pricing move. OpenAI cut o3 pricing by 80%, making reasoning-heavy applications dramatically more affordable. The signal was clear: standalone reasoning models were being commoditized because reasoning was about to be absorbed into the mainline GPT series.
Then came the main event. GPT-5 launched August 7, 2025, and it was genuinely a step change. For the first time, reasoning (previously the o-series’ domain) and multimodal capabilities (previously GPT-4o’s domain) were unified in a single model. GPT-5 became the default for all ChatGPT users, free and paid.
The benchmarks were not incremental improvements. They were generational leaps:
- 94.6% on AIME 2025 (math, without tools)
- 74.9% on SWE-bench Verified (real-world coding)
- 88% on Aider Polyglot (multilingual code generation)
- 84.2% on MMMU (multimodal understanding)
- 46.2% on HealthBench Hard (medical reasoning)
In the same month, OpenAI released gpt-oss-120b and gpt-oss-20b, its first open-weight models since GPT-2 in 2019. Licensed under Apache 2.0, these were designed for self-hosting and on-premises deployment. The open-source community had spent years criticizing OpenAI’s name as false advertising. This was OpenAI’s answer: good enough to compete with Meta’s Llama and Google’s Gemma, but notably not GPT-5.
Sora 2 arrived in September with meaningfully improved video generation. Longer outputs (15-25 seconds versus Sora 1’s 6-second limit), better temporal coherence, and a “remixing” feature. An iOS app followed with a “cameo” feature for inserting your likeness into generated videos. Then, in early 2026, came the Disney partnership: a three-year licensing deal allowing Sora to generate fan-inspired content featuring over 200 Disney, Marvel, Pixar, and Star Wars characters. That deal alone repositions Sora from “impressive tech demo” to “content production tool with actual commercial licensing.”
Q4 and Into 2026: The Pace Gets Absurd
The fourth quarter of 2025 is where the release cadence crossed from “aggressive” into “difficult to track without a spreadsheet.”
October brought DevDay 2025, where Sam Altman announced AgentKit, an integrated suite of tools for building, deploying, and optimizing AI agents. MCP (Model Context Protocol) support was added to the API, letting ChatGPT connect to external tools and data sources. The enterprise version of Codex became generally available. OpenAI was no longer just shipping models. It was shipping infrastructure.
November saw GPT-5.1. December brought GPT-5.2, claiming “expert-level” performance on business tasks, plus GPT-5.2-Codex for repo-scale code reasoning. By March 2026, GPT-5.3 and GPT-5.4 had arrived, with GPT-5.4 introducing native computer-use capabilities and a 1-million-token context window.
The retirement list grew just as fast. GPT-5.1 was removed from ChatGPT on March 11, 2026. GPT-4o, GPT-4.1, and o4-mini have all been deprecated. If you built production systems on any of these models, you have already migrated or you have a problem that is getting worse by the week.
| Date | Release | Key Capability | Status (Mar 2026) |
|---|---|---|---|
| Jan 2025 | Operator | Autonomous web agent | Superseded |
| Feb 2025 | Deep Research | Multi-step research agent | Active |
| Feb 2025 | GPT-4.5 | Largest chat model, bridge release | Retired |
| Apr 2025 | o3 + o4-mini | Frontier reasoning | Retired |
| Apr 2025 | GPT-4.1 family | Code-optimized models | Retired |
| May 2025 | Codex agent | Full AI software engineer | Active (GPT-5.3-Codex) |
| Aug 2025 | GPT-5 | Unified reasoning + multimodal | Retired |
| Aug 2025 | gpt-oss-120b / 20b | Open-weight, Apache 2.0 | Active |
| Sep 2025 | Sora 2 | Video generation, 15-25s | Active |
| Oct 2025 | AgentKit + MCP | Agent dev platform | Active |
| Dec 2025 | GPT-5.2 + Codex | Expert-level, repo-scale code | Legacy |
| Mar 2026 | GPT-5.4 | Computer use, 1M context | Current flagship |
The Business Behind the Launches
Product launches are exciting. Revenue numbers tell you whether they matter.
OpenAI reported $12 billion in annualized revenue as of July 2025, up from $3.7 billion in 2024. By February 2026, estimates put that figure at $25 billion ARR. ChatGPT reached 700 million weekly active users by August 2025, with 20 million paid subscribers as of April 2025, up from 15.5 million at the end of 2024. Five million business users were on enterprise plans.
The subscription restructuring tells its own story. OpenAI introduced ChatGPT Go at $8/month in January 2026, slotting below the existing Plus tier. This was a segmentation play: capture users who hit free-tier rate limits but do not need the full Plus feature set. Plus stayed at $20/month with added Codex access. Pro remained at $200/month for maximum performance and unlimited Deep Research.
On the infrastructure side, the Stargate Project, announced January 21, 2025, as a joint venture between OpenAI, Oracle, SoftBank, and MGX, committed $500 billion to AI infrastructure in the US. The initial $100 billion tranche is funding two data centers in Abilene, Texas, with construction underway. This is not a product launch. It is OpenAI building the physical substrate for everything that comes next.
On the API side, pricing dropped aggressively. GPT-5 starts at $1.25 per million input tokens and $10.00 per million output tokens. GPT-5 Mini runs at just $0.25 per million input tokens, five times cheaper than full GPT-5. The o4-mini offered reasoning at roughly half the cost of o3. The trend is unmistakable: OpenAI is using price cuts to expand the market while keeping premium tiers for maximum-performance use cases.
The Honest Assessment: What It All Means
Strip away the launch announcements and press releases, and a few strategic realities become clear.
The model treadmill is real and it is exhausting. Seven distinct model generations in thirteen months. Deprecation windows measured in weeks, not years. If you are building on OpenAI’s API, you are on a treadmill that does not have a stop button. Model abstraction layers are not a nice-to-have. They are survival infrastructure. Budget engineering time for quarterly migrations and accept that the model your application runs on today will not be the model it runs on in July.
Agents are the product now; chat was the prototype. Operator, Deep Research, Codex, AgentKit, computer use. The through-line of OpenAI’s 2025 strategy is unmistakable: the era of “type a question, get an answer” is being superseded by “describe a goal, watch an AI pursue it.” Codex does not suggest code; it writes, tests, and ships it. GPT-5.4 does not describe what is on your screen; it clicks the buttons. This is a fundamentally different product category than what ChatGPT was eighteen months ago.
OpenAI is a platform company now, not a research lab. The acquisitions ($3B Windsurf, $6.4B io), the AgentKit SDK, MCP integration, the Disney licensing deal, the enterprise Codex deployment, the four-tier subscription model, the Stargate infrastructure buildout. These are not the moves of an organization primarily concerned with publishing papers. They are the moves of a company building an ecosystem that other companies build on top of.
The competition is closer than the product announcements suggest. Anthropic’s Claude, Google’s Gemini, Meta’s open-weight Llama, and a dozen well-funded startups are all credible alternatives for specific use cases. OpenAI’s strategy is to move so fast that switching costs accumulate before competitors match each capability. Whether that pace is sustainable, technically and organizationally, is the question that $25 billion in revenue has not yet answered.
For users and developers, the practical advice is straightforward: use what works, abstract away the model layer, maintain optionality across providers, and do not bet your product on any single model version lasting more than six months. The tools have never been more capable. The ground beneath them has never shifted faster.
Frequently Asked Questions
For most professionals who use ChatGPT daily, Plus at $20/month hits the best balance of cost and capability. It includes Codex access, extended file uploads, and priority access to new features. The free tier is genuinely useful for casual use since it runs on GPT-5.4 Instant. ChatGPT Go at $8/month makes sense if you consistently hit free-tier rate limits but do not need Codex or extended uploads. Pro at $200/month is only justified if you need unlimited Deep Research or maximum model performance for complex analytical work where marginal quality gains matter.
Yes, but defensively. Use model abstraction layers so swapping model IDs requires a configuration change, not a code rewrite. Pin to the latest stable model and monitor deprecation announcements. Budget engineering time for migration every quarter. Seriously consider a multi-provider strategy using both OpenAI and alternatives like Anthropic or Google to reduce single-vendor lock-in. The API interface itself is stable; it is the models behind it that rotate aggressively. The capability is worth the operational overhead, but only if you architect for change from the start.
For supervised, low-stakes automation tasks, yes. GPT-5.4 can navigate interfaces, click buttons, fill forms, and execute multi-step workflows in a browser environment. For anything where errors carry real consequences, such as financial transactions, healthcare systems, or production deployments, maintain human oversight. The Codex agent is more mature and reliable specifically for software engineering tasks. Think of computer use as a capable intern: impressive at following instructions, not yet trustworthy enough to run unsupervised on critical systems.