TL;DR: OpenAI's 2026 lineup splits into three groups. General-purpose GPT-5 models (GPT-5.4 at $2.50/$15, GPT-5.5 at $5/$30). Reasoning o-series models (o4-mini, o3, o3-pro). And specialized models for code, voice, and images. GPT-5.4 is the workhorse. Nano tiers start at $0.05 input per million tokens.
OpenAI ships far more models than Anthropic or Google. In mid-2026 the API lists three groups. The GPT-5 series is the general-purpose line, from the tiny GPT-5.4 nano up to the flagship GPT-5.5. The o-series is built for deep reasoning, with o4-mini, o3, and o3-pro. And a set of specialized models handles coding (GPT-5.3-Codex), voice (gpt-realtime), and images (gpt-image-2). The current default for most teams is GPT-5.4.
OpenAI’s lineup is the largest of any major lab, and the naming is not obvious. GPT-5.4 is cheaper than GPT-5.5, o3 is a reasoning model, and Codex is a coding model. Picking the wrong one wastes money or caps quality. This guide helps your team match the model to the job, the same way we match engineers to roles.
Key takeaways
- Three groups in 2026: GPT-5 general-purpose, o-series reasoning, and specialized code, voice, and image models.
- GPT-5.4 ($2.50/$15) is the workhorse. GPT-5.5 ($5/$30) is the flagship. Nano input starts at $0.05 per million.
- Output is the price driver. GPT-5.5 output ($30) costs 75 times the GPT-5 nano rate ($0.40).
- GPT-5 models take a reasoning effort setting from none to xhigh, so one model can be cheap or deep.
- Most teams should default to GPT-5.4 and only reach for GPT-5.5, o3, or the pro tiers when the task is genuinely hard.
The current OpenAI models at a glance
Here is the core lineup, with the exact model ID your engineers will type into the API. Use this table as your reference. Everything below explains the details. Prices are input then output, per one million tokens.
| Model | Model ID | Context | Price in / out (per 1M) | Best for |
|---|---|---|---|---|
| GPT-5.5 | gpt-5.5 | 1M | $5 / $30 | Flagship general model, hardest agentic work |
| GPT-5.5 Pro | gpt-5.5-pro | 1M | $30 / $180 | Extended reasoning, highest stakes answers |
| GPT-5.4 | gpt-5.4 | 1M | $2.50 / $15 | The default for most production work |
| GPT-5.4 mini | gpt-5.4-mini | 1M | $0.75 / $4.50 | Strong small model, cheaper high volume |
| GPT-5.4 nano | gpt-5.4-nano | 1M | $0.20 / $1.25 | Fast, cheap, simple tasks at scale |
| o3 | o3 | 200K | $2 / $8 | Deep reasoning, math, science, planning |
| o3-pro | o3-pro | 200K | $20 / $80 | The hardest reasoning, maximum effort |
| o4-mini | o4-mini | 200K | $1.10 / $4.40 | Cheap reasoning, good cost per correct answer |
| GPT-5.3-Codex | gpt-5.3-codex | 1M | $1.75 / $14 | Agentic coding, long code sessions |
Which OpenAI model fits your team?
Pick your main use case to see the right model.
For chatbots, tagging, and routing at scale, the nano and mini tiers are fast and start at $0.20/$1.25 per million tokens. You only pay for GPT-5.5 when a task truly needs it. See the cost of building this in-house on our developer rate card.
For everyday coding, reviews, and tool use, Codex is tuned for long code sessions and GPT-5.4 is the balanced default. Pair it with engineers who ship agents daily. Hire AI agent developers who build with it.
For math, science, planning, and hard multi-step problems, the o-series and the flagship earn their price. Strong AI and machine learning engineers get the most from them.
The model is the easy part. The hard part is the team that builds and ships with it. We match you with pre-vetted engineers in about 24 hours. Tell us what you need.
The OpenAI model families in plain English
OpenAI groups its models by job, not by a single naming theme. The GPT-5 line is general-purpose. The o-series is for reasoning. And a handful of specialized models do code, voice, and images. Once you see the three groups, the long list stops feeling random.
The GPT-5 series (general purpose)
The GPT-5 line is where most work happens. GPT-5.5 is the flagship, released in April 2026, at $5 input and $30 output per million tokens. GPT-5.4 is the everyday workhorse at $2.50 input and $15 output, and it is the model we tell most clients to start with. Below it sit GPT-5.4 mini ($0.75/$4.50) and GPT-5.4 nano ($0.20/$1.25) for cheaper, high-volume work. All of them carry a context window around one million tokens.
Above the flagship are the pro tiers. GPT-5.5 Pro and GPT-5.4 Pro both run at $30 input and $180 output. They spend more compute to think longer and check their own work. Reach for a pro tier only when a wrong answer is expensive, because the output price is six times the standard flagship.
The o-series (reasoning)
The o-series is built to think before it answers. These models spend hidden reasoning tokens on hard problems in math, science, code, and planning. o4-mini is the cheap reasoning pick at $1.10 input and $4.40 output. o3 sits in the middle, and after a July 2026 price cut it now costs $2 input and $8 output per million tokens. o3-pro is the top of the range at $20 input and $80 output for the hardest problems. The o-series uses a 200K token context window, smaller than the GPT-5 line.
Specialized models (code, voice, images)
Three families handle jobs that need a purpose-built model. GPT-5.3-Codex ($1.75/$14) is tuned for agentic coding and long code sessions, and it powers the Codex tools. gpt-realtime-2.1 handles live voice, priced separately for audio and text. gpt-image-2 generates and edits images. There are also deep-research variants, o3-deep-research ($5/$20) and o4-mini-deep-research ($1/$4), that run long, tool-heavy research jobs on their own.
What the current models can do
The GPT-5 and o-series models share the same core skills. Every current model can read images, call tools, and return strict JSON when you ask for it. That last one, called structured outputs, forces the reply to match a schema you set, which makes the data safe to parse in code. This is what powers agents that do real work, not just chat.
Vision is built in. The model reads screenshots, charts, PDFs, and photos. Tool use lets the model call your functions, run code, search the web, and read files. The GPT-5 line and the o-series both work through the Responses API and the older Chat Completions API, so most existing code keeps working after you switch the model ID.
Two features save real money at scale. Cached input charges a repeated prompt at about one tenth of the normal input price. The Batch API runs non-urgent jobs at half price. A team that uses both can serve heavy workloads for far less than the sticker rate suggests. Knowing which lever to reach for is part of what we test when we vet engineers.
Pricing compared across the lineup
Output tokens are the real cost driver. Most apps read a prompt and write a longer answer, so output volume adds up fast. The gap across the lineup is huge. GPT-5 nano output is $0.40 per million. GPT-5.5 output is $30 per million, which is 75 times more, and the pro tiers reach $180. The chart below shows output price for the main current models.
One way to cut cost is to route by difficulty. Send easy requests to GPT-5.4 nano and only escalate the hard ones to GPT-5.5 or o3. Another lever is cached input, which charges a repeated context at about a tenth of the input price. Teams that build this routing layer well often cut their model bill by half or more. This is the kind of work that separates a strong AI engineer from an average one.
Context windows and reasoning effort
The context window is how much text the model can read at once, counting both your prompt and the conversation so far. The GPT-5 line offers a window around one million tokens, which is roughly 750,000 words. That is enough to drop a whole codebase or a stack of long documents into a single request. The o-series uses a smaller 200K window, which is still about 150,000 words and fine for most reasoning tasks.
The GPT-5 models fold reasoning into one setting. You pass a reasoning effort value from none, to low, medium, high, and xhigh. Low effort means fewer thinking tokens and faster, cheaper replies. High effort means the model thinks longer on hard problems. A separate verbosity setting controls how long the answer is. This means one model can be cheap or deep, depending on how you call it.
This matters for hiring. An engineer who understands reasoning effort and caching will tune a workflow to use far fewer tokens at the same quality. We see this gap often when we vet candidates. The ones who have shipped real OpenAI apps know these levers cold. You can read more in our breakdown of Claude versus ChatGPT for coding.
Which model should you use
Start with GPT-5.4. It is the safe default for most teams and most tasks. Move down to the mini and nano tiers when the task is simple and you run high volume. Move up to GPT-5.5, o3, or a pro tier when the task is genuinely hard or the cost of a wrong answer is high. The table below maps common jobs to a model.
| Job | Recommended model | Why |
|---|---|---|
| Support chatbot, FAQ | GPT-5.4 nano | Fast and cheap at high volume |
| Data tagging, classification | GPT-5.4 nano | Simple, repeated, low risk |
| Daily coding and code review | GPT-5.3-Codex | Tuned for long agentic code sessions |
| Document summaries, extraction | GPT-5.4 | 1M context, good accuracy, fair price |
| Math, science, planning | o3 or o4-mini | Deep reasoning at a fair token cost |
| Long autonomous agents | GPT-5.5 | Strong tool use over long tasks |
| Hardest reasoning, high stakes | GPT-5.5 Pro or o3-pro | Most compute, worth the premium |
A simple rule helps. If a cheaper model gives the right answer most of the time, use it and add a check. Escalate only the cases that fail. This keeps quality high and cost low. We worked with a fintech team that ran every request on GPT-5.5. Moving routine checks to GPT-5.4 and tagging to nano cut their bill by more than half with no drop in quality.
Legacy and retired models
OpenAI keeps older models running for a while, then retires them. If your code still calls an old model ID, plan to move it. The table shows the main older models and what to use instead. Retired models return an error, so those need a fast fix.
| Older model | Status | Move to |
|---|---|---|
| GPT-5.0, 5.1, 5.2, 5.3 | Superseded | GPT-5.4 or GPT-5.5 |
| GPT-4o, GPT-4o mini | Deprecated | GPT-5.4, GPT-5.4 mini |
| GPT-4.1, 4.1 mini, 4.1 nano | Active or deprecated | GPT-5.4 family |
| o1, o1-mini, o3-mini | Retired | o4-mini, o3 |
| GPT-4 Turbo, GPT-3.5 Turbo | Retired | GPT-5.4, GPT-5.4 nano |
A move is usually small. Change the model ID, switch older sampling settings to the reasoning effort value the new models expect, and update your endpoint to the Responses API if you want the newest features. The current OpenAI models page lists every active and retired ID in one place, so check it before you migrate.
What this means for hiring
Knowing the model lineup is step one. Getting value from it is the harder step. The teams that win are not the ones on the biggest model. They are the ones who route work to the right model, cache context, set reasoning effort well, and ship a clean product around it. That is engineering, not just a model choice.
Strong AI engineers are scarce and costly in the United States. A senior AI engineer there runs $12,000 to $18,000 per month all in. The same skill from a pre-vetted engineer in Vietnam or the Philippines often costs $3,000 to $6,000 per month. The work is the same. The bill is not. Many of our clients build their whole OpenAI stack with engineers from our network.
If you want to go deeper on the wider picture, our ChatGPT statistics for 2026 covers adoption, revenue, and enterprise reach. Read that next to see how fast this market is moving and why the right team matters more each quarter.
Compare other model families
Build your OpenAI stack with the right team
OpenAI gives you a deep lineup and clear prices. The model is the easy part. The team that builds with it is what ships value. We match you with pre-vetted senior engineers across Asia in about 24 hours, with no upfront cost and payroll handled.
