OpenAI released GPT-5.5 on April 23, 2026. It remains available in the API, but it is no longer the default starting point for a new OpenAI integration.
OpenAI released the GPT-5.6 family in July 2026 and now recommends GPT-5.6 Sol for complex reasoning and coding, Terra for a balance of capability and cost, and Luna for efficient high-volume work. See the current OpenAI model guidance.
If an existing workflow uses GPT-5.5, this guide explains its role and what to evaluate before moving. For a new workflow, test GPT-5.6 and GPT-5.5 against the same representative tasks rather than choosing from launch claims alone.
You can also explore the full model profile on our GPT-5.5 model page.
What is GPT-5.5?
GPT-5.5 is OpenAI’s newest frontier model and the direct successor to GPT-5.4. OpenAI describes it as “a new class of intelligence for real work” - which is a bit of marketing-speak, but the model genuinely earns it in the areas that matter.
Where GPT-5.4 was already excellent at reasoning and language tasks, GPT-5.5 takes the next step toward autonomous, multi-step work. It is designed to understand what you are trying to do, take on the messy, multi-part task, and keep going until it is finished.
That means planning, using tools, checking its own work, navigating ambiguity, and moving across different parts of a system - all without needing you to manage every step.
Three areas show the biggest gains:
- Agentic coding - writing, debugging, refactoring, and testing across large codebases
- Knowledge work - research, document creation, spreadsheets, and computer use
- Scientific research - multi-stage data analysis, hypothesis testing, and co-scientist workflows
What is New in GPT-5.5
Here are the changes that matter most for builders and power users.
1. Agentic Coding at a New Level
GPT-5.5 is OpenAI’s strongest coding model to date. The numbers back that up.
On Terminal-Bench 2.0, which tests complex command-line workflows that require planning, iteration, and tool coordination, GPT-5.5 scores 82.7% - up from 75.1% on GPT-5.4 and significantly ahead of Claude Opus 4.7 at 69.4%.
On SWE-Bench Pro, which tests real-world GitHub issue resolution, it reaches 58.6%, solving more tasks end-to-end in a single pass than any previous model. On Expert-SWE - OpenAI’s internal benchmark for long-horizon coding tasks with a median estimated human completion time of 20 hours - GPT-5.5 also outperforms GPT-5.4.
But numbers only tell part of the story. Early testers describe a qualitative shift in how the model approaches code. GPT-5.5 understands the shape of a system: not just the function you asked it to fix, but why something is failing, where the fix needs to land, and what else in the codebase would be affected.
Dan Shipper, Founder and CEO of Every, called it “the first coding model I’ve used that has serious conceptual clarity.” Pietro Schirano, CEO of MagicPath, watched it merge a branch with hundreds of frontend and refactor changes into a substantially different main branch - resolving everything in one shot in about 20 minutes.
Michael Truell, Co-founder and CEO of Cursor, summed it up well: “GPT-5.5 is noticeably smarter and more persistent than GPT-5.4, with stronger coding performance and more reliable tool use. It stays on task for significantly longer without stopping early, which matters most for the complex, long-running work our users delegate to Cursor.”
2. Better Knowledge Work and Computer Use
The same qualities that make GPT-5.5 great at coding carry over into everyday knowledge work.
The model is better at understanding intent, which lets it move more naturally through the full loop: finding information, figuring out what matters, using tools, checking its own output, and turning raw material into something finished.
In Codex, GPT-5.5 outperforms GPT-5.4 at generating documents, spreadsheets, and slide presentations. Alpha testers said it handled operational research, financial modeling, and turning messy business inputs into structured plans better than anything before it.
Some real examples from inside OpenAI are revealing:
- The Finance team used GPT-5.5 in Codex to review 24,771 K-1 tax forms totaling 71,637 pages - accelerating the task by two weeks compared to the prior year.
- The Comms team used it to build an automated Slack agent for handling speaking requests, keeping low-risk ones automatic and routing high-risk ones to humans.
- A Go-to-Market employee automated generating weekly business reports, saving 5–10 hours a week.
Today, more than 85% of OpenAI’s company uses Codex every week across software engineering, finance, communications, marketing, data science, and product management.
On OSWorld-Verified, which measures whether a model can operate real computer environments on its own, GPT-5.5 scores 78.7% - comparable to Claude Opus 4.7 at 78.0%. On GDPval, which tests agents’ ability to produce knowledge work across 44 occupations, it scores 84.9%.
3. Scientific Research
This might be the most exciting area for the long term.
GPT-5.5 shows meaningful gains on scientific and technical workflows - the kind that require more than answering a question. Researchers need to explore an idea, gather evidence, test assumptions, interpret results, and decide what to try next. GPT-5.5 is better at persisting across that full loop.
On GeneBench, a new benchmark for multi-stage scientific data analysis in genetics and quantitative biology, GPT-5.5 scores 25.0% versus GPT-5.4’s 19.0%. On BixBench, focused on bioinformatics and data analysis, GPT-5.5 achieved 80.5% - leading performance among models with published scores.
Two real examples are worth highlighting:
Derya Unutmaz, an immunology professor at the Jackson Laboratory, used GPT-5.5 Pro to analyze a gene-expression dataset with 62 samples and nearly 28,000 genes. It produced a detailed research report that surfaced key questions and insights - work he said would have taken his team months.
An internal version of GPT-5.5 with a custom harness helped discover a new mathematical proof about Ramsey numbers, later verified in Lean. It is a concrete example of the model contributing not just code or explanation, but a genuinely surprising and useful mathematical argument.
4. Token Efficiency: More Intelligent, Not More Expensive
One of the most important things about GPT-5.5 is what it does not do. Bigger, more capable models are almost always slower and more expensive to serve. GPT-5.5 breaks that pattern.
GPT-5.5 matches GPT-5.4’s per-token latency in real-world serving while operating at a meaningfully higher level of intelligence. And crucially, it uses significantly fewer tokens to complete the same tasks. On OpenAI’s Artificial Analysis Coding Index, GPT-5.5 delivers state-of-the-art intelligence at half the cost of competitive frontier coding models.
For long-running agents - the kind that plan, act, inspect, retry, and refine over many cycles - that token efficiency compounds into real savings.
5. Long-Context Performance
GPT-5.5 comes with a 1M token context window for API use and 400K tokens in Codex.
More importantly, it uses that context much better than GPT-5.4. On the long-context reasoning benchmarks (Graphwalks BFS at 1M tokens), GPT-5.5 scores 45.4% - compared to GPT-5.4’s 9.4%. That is not a small improvement. It means GPT-5.5 can actually reason across million-token contexts in ways that GPT-5.4 simply could not.
For teams building agents that work with large codebases, long documents, or extended multi-session workflows, this is a significant unlock.
GPT-5.5 Variants
OpenAI offers two variants of GPT-5.5.
GPT-5.5 (Standard)
The standard model is available in ChatGPT as GPT-5.5 Thinking - a version that delivers smarter, more concise answers for harder problems. It is available to Plus, Pro, Business, and Enterprise users.
In Codex, GPT-5.5 is available on Plus, Pro, Business, Enterprise, Edu, and Go plans. A Fast mode is also available, generating tokens 1.5x faster for 2.5x the cost - useful when you need speed over economy.
For developers, the API is coming “very soon” with a 1M token context window.
GPT-5.5 Pro
GPT-5.5 Pro uses parallel test-time compute to push accuracy even higher on the most demanding tasks. It is available to Pro, Business, and Enterprise users in ChatGPT.
Early testers described using GPT-5.5 Pro less like a question-answering tool and more like a research partner - critiquing manuscripts over multiple passes, stress-testing technical arguments, and working with code, notes, and PDF context together.
On GDPval (knowledge work across 44 occupations), GPT-5.5 Pro scores 82.3%. On FrontierMath Tier 4 (extremely hard graduate-level math), it reaches 39.6%, compared to 35.4% for standard GPT-5.5 and 27.1% for GPT-5.4. The Pro variant earns its premium for high-stakes, accuracy-critical work.
GPT-5.5 Pricing
Here is the full pricing breakdown once the API launches.
| Model | Input | Output |
|---|---|---|
| GPT-5.5 | $5 / MTok | $30 / MTok |
| GPT-5.5 Pro | $30 / MTok | $180 / MTok |
| GPT-5.4 (for reference) | $2.50 / MTok | $15 / MTok |
A few things worth noting:
- Batch and Flex pricing are available at 50% of the standard rate - good for high-volume, non-urgent workloads.
- Priority processing is 2.5x the standard rate for latency-sensitive tasks.
- GPT-5.5 is priced higher per token than GPT-5.4, but uses significantly fewer tokens per task. OpenAI says that for most Codex users, the net cost is comparable or better.
The practical takeaway: if you are running short, high-volume tasks with simple inputs, GPT-5.4 may still be the smarter economic choice. If you are running complex, multi-step, long-horizon work - the kind where GPT-5.5’s efficiency really kicks in - the upgrade likely pays for itself.
How GPT-5.5 Compares to the Competition
Here is how GPT-5.5 stacks up on the benchmarks that matter.
| Benchmark | GPT-5.5 | GPT-5.4 | Claude Opus 4.7 | Gemini 3.1 Pro |
|---|---|---|---|---|
| Terminal-Bench 2.0 | 82.7% | 75.1% | 69.4% | 68.5% |
| GDPval (wins or ties) | 84.9% | 83.0% | 80.3% | 67.3% |
| OSWorld-Verified | 78.7% | 75.0% | 78.0% | - |
| ARC-AGI-2 | 85.0% | 73.3% | 75.8% | 77.1% |
| FrontierMath Tier 1–3 | 51.7% | 47.6% | 43.8% | 36.9% |
| BrowseComp | 84.4% | 82.7% | 79.3% | 85.9% |
| GPQA Diamond | 93.6% | 92.8% | 94.2% | 94.3% |
The pattern is clear: GPT-5.5 leads on agentic tasks (Terminal-Bench, OSWorld, GDPval) and abstract reasoning (ARC-AGI-2). Claude Opus 4.7 and Gemini 3.1 Pro are still competitive on academic benchmarks, and GPT-5.5 Pro narrows the gap further on the hardest tasks.
For coding specifically, note that Claude Opus 4.7 scores 64.3% on SWE-Bench Pro compared to GPT-5.5’s 58.6%. The two models are genuinely competitive on code - the better choice often comes down to your specific workflow and toolchain rather than a universal winner.
Who Should Use GPT-5.5?
GPT-5.5 is the right call if you fall into any of these categories:
- Software engineers and agent builders doing complex, long-horizon work - large refactors, multi-step automations, debugging hairy issues across a codebase.
- Knowledge workers and analysts who spend time on research synthesis, financial modeling, report generation, or document-heavy tasks.
- Scientific and technical researchers who need a model that can persist across a multi-step investigation rather than answering one question at a time.
- Teams running Codex workflows at scale - the token efficiency gains are most valuable here.
If your work is mostly short, simple, high-volume tasks like classification, extraction, or quick Q&A, GPT-5.4 (or even earlier models) are likely more economical. GPT-5.5 earns its keep on the hard stuff.
Safety and Deployment
OpenAI released GPT-5.5 with what it calls its “strongest set of safeguards to date.”
The model was evaluated across OpenAI’s full suite of safety and preparedness frameworks, with targeted testing for advanced cybersecurity and biology capabilities. Nearly 200 trusted early-access partners provided feedback on real use cases before launch.
For cybersecurity specifically, OpenAI rates GPT-5.5’s capabilities as High under its Preparedness Framework - a significant level, though not Critical. Stricter classifiers are deployed to limit potential misuse, while OpenAI’s new Trusted Access for Cyber program gives verified security professionals expanded access for legitimate defensive work. You can apply at chatgpt.com/cyber.
Use GPT-5.5 in an Internal Operations Workflow
A capable model is only one part of a reliable business system. The workflow still needs approved inputs, a data model, access controls, validation, human review, and a record of what happened.
Appaca is an AI workspace for operators. Teams can build internal apps and AI coworkers that use shared business data, company knowledge, integrations, and scheduled work.
Useful examples include:
- Extracting invoice fields for finance review
- Preparing a weekly operations report from approved records
- Answering SOP questions with source references
- Drafting a sales follow-up from CRM activity
- Turning onboarding information into a checklist
Appaca is not intended for building and selling public customer-facing AI products. Use it when the model needs a defined role inside an internal process.
The Bottom Line
GPT-5.5 remains a capable option for existing professional and coding workloads, but OpenAI’s current guidance starts with the GPT-5.6 family for new work.
Do not migrate only because the model number changed. Build a test set from the real workflow and compare accuracy, tool use, latency, token usage, human edits, and failures. Check current availability and pricing in the official GPT-5.5 model page and OpenAI pricing.
To compare model specifications side by side, use the LLM comparison hub. To put a model inside an internal process, build an AI workflow in Appaca.