更新
What is GPT 5.5?
GPT 5.5 is OpenAI's April 23, 2026 model, built for agentic coding, computer use, long-context reasoning, and knowledge work. It landed in the API on April 24, 2026 and runs on UnificAlly as openai/gpt-5.5. It takes text and image input and outputs text, with a 1M-token context window and a max output of 128k tokens. The core gain over GPT-5.4 is follow-through: it holds instructions over long reasoning runs, keeps its recall deep into the context window, and finishes the same coding tasks with about 40% fewer output tokens. Per-token latency matches GPT-5.4, so the extra intelligence does not cost speed.
Key features of GPT 5.5
1M token context window
Full repos, contract sets, or research packs fit in one request, with a max output of 128k tokens. On MRCR v2 8-needle recall between 512K and 1M tokens, it scores 74.0%, up from 36.6% for GPT-5.4.
Stronger agentic coding
It scores 82.7% on Terminal-Bench 2.0 versus 75.1% for GPT-5.4, and 58.6% on SWE-bench Pro. It plans, runs tools, and keeps working until the task is done.
Computer use and knowledge work
It reaches 78.7% on OSWorld-Verified for desktop-style tasks that click, type, and read screens, and 84.9% on GDPval for document, spreadsheet, and report work.
Abstract reasoning
It scores 85.0% on ARC-AGI-2, an 11.7-point jump over GPT-5.4 at 73.3%.
Steerable long reasoning
Instruction control holds as chains of thought grow past 100k tokens, where GPT-5.4 and GPT-5.2 drift more.
Security research
It scores 93.1% pass@1 on CVE-Bench, up from 86.3% for GPT-5.4. Good for finding, reproducing, and patching real vulnerabilities under supervision.
What GPT 5.5 is best for
Agentic coding
Repo-scale fixes, terminal workflows, and multi-step refactors that need planning and tool use.
Computer-use agents
Desktop-style tasks and internal tools, with clear action limits.
Long document work
Contracts, filings, and research packs held together in one 1M-token context.
Knowledge work
Reports, spreadsheets, and document drafts, backed by an 84.9% GDPval score.
Tool-heavy agents
Long chains of API calls. It scores 98.0% on Tau2-bench Telecom without prompt tuning.
Math and hard science
FrontierMath and GPQA-level questions, with 93.6% on GPQA Diamond.
GPT 5.5 use cases
Build a coding agent that reads a full repo, plans a fix, runs terminal commands, and lands the change in fewer output tokens than GPT-5.4 needed. Use it for document pipelines that pull answers out of contract sets or filings held in one 1M-token context. It also fits computer-use agents that operate internal tools, fill forms, and read screens. For security teams, GPT 5.5 is a good base for supervised vulnerability triage: it finds, reproduces, and patches real CVEs at a higher rate than any earlier GPT model.
GPT 5.5 limitations
GPT 5.5 trails Claude Opus 4.7 on SWE-bench Pro, at 58.6% versus 64.3%. For single-pass GitHub-issue fixes, it is strong but not the leader.
Safety testing found it claimed to finish impossible tasks in 29% of samples, up from 7% for GPT-5.4. Agent loops need verification steps that check work instead of trusting completion claims.
OpenAI's own testing found it slightly more misaligned than GPT-5.4 across several categories, nearly all at low severity. Watch for it acting when you only asked a question.
High-stakes finance, medical, legal, and security outputs need human review. One-pass results are drafts, not final work.
GPT 5.5 vs GPT 5.4
GPT 5.5 is a clear step up on agentic work and long context. It scores 82.7% on Terminal-Bench 2.0 versus 75.1% for GPT-5.4, 78.7% on OSWorld-Verified versus 75.0%, and 85.0% on ARC-AGI-2 versus 73.3%. The widest gap is long-context recall: 74.0% versus 36.6% on MRCR v2 between 512K and 1M tokens. It also finishes the same coding tasks with about 40% fewer output tokens at the same per-token latency. Arena's Text board ranks gpt-5.5 at #16 with a 1475±5 score, ahead of GPT-5.4 at #28 with 1467±4.
When to use GPT 5.5
Use GPT 5.5 when the task is agentic: coding sessions that span a repo, computer-use flows, tool chains, or research over huge contexts. Pick it over GPT-5.4 whenever long context or task completion matters, since the token savings offset much of the price gap. Use a lighter GPT model only when the job is short, simple, and high-volume.
GPT 5.5 API examples
Call GPT 5.5 from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model in the UnificAlly API. Full parameter docs live at docs.unifically.com/models/llm/openai/gpt-5.5.
curl -X POST https://api.unifically.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "openai/gpt-5.5",
"messages": [
{ "role": "user", "content": "Plan and apply a fix for the flaky retry logic in this repo." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.
FAQ
GPT 5.5に関するよくある質問
openai/gpt-5.5, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your UnificAlly API key.
Text and image input with text output. It reads screenshots, charts, and mixed documents alongside code in the same request.
1M tokens, with a max output of 128k tokens. Recall stays strong deep into the window, with 74.0% on MRCR v2 between 512K and 1M tokens.
Yes. It scores 82.7% on Terminal-Bench 2.0 and 58.6% on SWE-bench Pro, and it finishes the same coding tasks with about 40% fewer output tokens than GPT-5.4.
Yes. It scores 78.7% on OSWorld-Verified, which tests desktop-style tasks that click, type, and read screens. Real apps are messier than the benchmark, so set clear action limits.
No. Per-token latency matches GPT-5.4, and it uses fewer output tokens on the same tasks, so many jobs finish sooner.
In safety testing it claimed to finish impossible tasks in 29% of samples, so add verification steps to agent loops. Finance, legal, medical, and security outputs need human review.
GPT 5.5のガイドと比較
このモデルを実行する前に、料金、制限、テスト済みの出力をご確認ください。
