Overview
GPT-5.5, released April 23, 2026, is OpenAI's most capable general-purpose model and the first fully retrained base since GPT-4.5. Built on a natively omnimodal architecture and co-designed with NVIDIA's GB200/GB300 NVL72 hardware, it matches GPT-5.4's per-token latency despite substantially higher intelligence. The model leads the Artificial Analysis Intelligence Index (score: 60), achieving state-of-the-art results on agentic coding benchmarks like Terminal-Bench 2.0 (82.7%) and setting a new OpenAI mark on SWE-bench Verified (88.7%). Its configurable reasoning effort dial — spanning `none` through `xhigh` — lets developers trade latency and cost for depth on a per-call basis.
The model's positioning centers on agentic work: multi-step coding, computer use, long-horizon research, and tool-calling agent loops. Codex, OpenAI's coding agent, is the primary deployment surface, with over 85% of OpenAI's own employees using it weekly. GPT-5.5 demonstrates a genuine leap in long-context retention (74.0% at 1M tokens, up from 36.6% on GPT-5.4) and ARC-AGI-2 abstract reasoning (85.0%, up from 73.3%). The architecture reportedly uses mixture-of-experts with community estimates of 100–200B active parameters, though OpenAI has not disclosed specifics.
GPT-5.5 is priced at a premium — $5/$30 per 1M input/output tokens, double GPT-5.4's rates — though OpenAI argues that ~40% token efficiency gains partially offset the increase. The Pro variant ($30/$180) targets the hardest reasoning tasks as an escalation tier. The model's main competitive weaknesses are its high hallucination rate on broad knowledge tasks relative to Claude Opus 4.7, its loss to Opus on SWE-bench Pro (the harder, less-memorized coding benchmark), and the cost premium that makes naive migration expensive. For teams building agentic systems, tool-using agents, or computer-use workflows, GPT-5.5 is currently the strongest available foundation model from OpenAI.
Benchmarks & Performance
## Core Benchmark Scores
| Benchmark | GPT-5.5 | GPT-5.4 | Claude Opus 4.7 | Gemini 3.1 Pro |
|---|---|---|---|---|
| SWE-bench Verified | 88.7% | ~80% | ~87.6% | — |
| SWE-bench Pro | 58.6% | 57.7% | 64.3% | 54.2% |
| Terminal-Bench 2.0 | 82.7% | 75.1% | 69.4% | 68.5% |
| GPQA Diamond | 93.6% | 92.8% | 94.2% | 94.3% |
| ARC-AGI-2 | 85.0% | 73.3% | 75.8% | 77.1% |
| FrontierMath T1–3 | 51.7% | 47.6% | 43.8% | 36.9% |
| FrontierMath T4 | 35.4% | 27.1% | 22.9% | 16.7% |
| HLE (no tools) | 41.4% | 39.8% | 46.9% | 44.4% |
| HLE (with tools) | 52.2% | 52.1% | 54.7% | 51.4% |
| BrowseComp | 84.4% | 82.7% | 79.3% | 85.9% |
| OSWorld-Verified | 78.7% | 75.0% | 78.0% | — |
| GDPval | 84.9% | 83.0% | 80.3% | 67.3% |
| Tau2-bench Telecom | 98.0% | 92.8% | — | — |
| CyberGym | 81.8% | 79.0% | 73.1% | — |
| GeneBench | 25.0% | 19.0% | — | — |
| MMLU-Pro | 92.4% | — | — | — |
| AIME 2026 | 97.50% | — | — | — |
## Long-Context Performance (MRCR v2)
| Context Range | GPT-5.5 | GPT-5.4 | Claude Opus 4.7 |
|---|---|---|---|
| 8K–16K | 93.0% | 91.4% | — |
| 32K–64K | 90.0% | 90.5% | — |
| 128K–256K | 87.5% | 79.3% | 59.2% |
| 256K–512K | 81.5% | 57.5% | — |
| 512K–1M | 74.0% | 36.6% | 32.2% (Opus 4.6) |
## Key Takeaway
GPT-5.5 leads on agentic coding (Terminal-Bench 2.0), abstract reasoning (ARC-AGI-2), and long-context retention at 1M tokens. It trails Claude Opus 4.7 on SWE-bench Pro and Humanity's Last Exam (no tools), and is roughly matched by Gemini 3.1 Pro on GPQA Diamond. Its FrontierMath performance shows the strongest progress on hard mathematics among frontier models.
Detailed Comparison
## GPT-5.5 vs Claude Opus 4.7
| Dimension | GPT-5.5 | Claude Opus 4.7 |
|---|---|---|
| API Pricing (in/out per 1M) | $5 / $30 | $5 / $25 |
| Pro Pricing | $30 / $180 | N/A |
| Context Window | 1M tokens | 500K tokens |
| Arena Elo | ~1474 | ~1492 |
| AAII Score | 60 (#1) | 58.1 (#3) |
| SWE-bench Pro | 58.6% | 64.3% |
| Terminal-Bench 2.0 | 82.7% | 69.4% |
| GPQA Diamond | 93.6% | 94.2% |
| AA-Omniscience Hallucination | 86% | 36% |
| Tool-call success (first attempt) | 97.4% | 94.2% |
GPT-5.5 wins on agentic terminal workflows, tool reliability, long-context retention, and the breadth of its tool surface (hosted shell, apply_patch, computer use, MCP). Claude Opus 4.7 wins on SWE-bench Pro, conversational preference, factual accuracy, multi-file refactoring, and hallucination rate. For coding agents in production, GPT-5.5 offers better tool-call reliability; for knowledge-intensive retrieval, Opus 4.7's lower hallucination rate is a meaningful advantage.
## GPT-5.5 vs Gemini 3.1 Pro
| Dimension | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|
| API Pricing (in/out per 1M) | $5 / $30 | $2–3.50 / $12 |
| Context Window | 1M | 2M |
| GPQA Diamond | 93.6% | 94.3% |
| Terminal-Bench 2.0 | 82.7% | 68.5% |
| FrontierMath T1–3 | 51.7% | 36.9% |
| Translation (N5) | 84 | 92 |
| BrowseComp | 84.4% | 85.9% |
Gemini 3.1 Pro offers a larger context window (2M vs 1M), stronger GPQA Diamond performance, superior translation quality, and significantly lower pricing. GPT-5.5 dominates on agentic coding, abstract reasoning (ARC-AGI-2), tool-call reliability, and the maturity of its developer ecosystem (SDK coverage, Azure integration, ChatGPT distribution). Gemini's advantage is most pronounced for teams embedded in the Google Workspace ecosystem or running high-volume workloads where cost sensitivity dominates.
## GPT-5.5 vs DeepSeek V4 Pro
| Dimension | GPT-5.5 | DeepSeek V4 Pro |
|---|---|---|
| API Pricing (in/out per 1M) | $5 / $30 | ~$0.30 / $1.10 |
| AAII Score | 60 | ~54.7 |
DeepSeek V4 Pro offers comparable capability on several benchmarks at roughly one-tenth the API cost. The pricing gap makes it the default choice for cost-sensitive bulk extraction and classification tasks. GPT-5.5 maintains a clear lead on agentic coding, tool reliability, and enterprise compliance (SOC 2, HIPAA BAA, GDPR). For teams that don't need OpenAI's ecosystem advantages, DeepSeek represents the strongest cost-efficiency option in the frontier tier.
Community Feedback
Developer and researcher sentiment on GPT-5.5 is strong but polarized along clear lines.
**Positive reception** centers on agentic coding performance and tool reliability. Dan Shipper, CEO of Every, described GPT-5.5 as "the first coding model I've used that has serious conceptual clarity," after it resolved a post-launch debugging issue that GPT-5.4 and his best engineer had previously failed to fix. Pietro Schirano, CEO of MagicPath, successfully had GPT-5.5 merge a branch with hundreds of frontend and refactor changes into a substantially divergent main branch in a single pass. Michael Truell, CEO of Cursor, reported that GPT-5.5 "stays on task for significantly longer without stopping early." NVIDIA's VP of Enterprise AI, Justin Boitano, stated that 10,000+ NVIDIA employees use GPT-5.5-powered Codex, with Jensen Huang describing it as a "jump to lightspeed" moment. One NVIDIA engineer went further: "Losing access to GPT-5.5 feels like I've had a limb amputated."
**Critical reception** focuses on three areas: pricing (the 2x increase over GPT-5.4 is contentious, especially for high-volume API users), hallucination rates (the AA-Omniscience 86% figure vs Claude's 36% is widely cited as a material limitation), and conversational quality (LMArena Elo trails Claude Opus 4.7, suggesting humans generally prefer Claude's conversational outputs). Sam Altman's description of GPT-5.5 as an "autistic genius" during the press briefing generated significant backlash from disability advocates and was widely discussed on X without an official OpenAI response.
**Adoption patterns** show GPT-5.5 rapidly becoming the default for agentic coding toolchains — Cursor, GitHub Copilot, and Codex all integrated it at or near launch. Enterprise adoption is accelerating through Azure OpenAI, with Microsoft 365 Copilot integration following shortly after. However, teams running cost-sensitive high-volume workloads or requiring high factual accuracy in broad-domain retrieval are splitting traffic between GPT-5.5 (for hard tasks) and cheaper or more reliable alternatives (for routine queries). The "route by task" pattern — using GPT-5.5 for the hardest 20% of queries and cheaper models for the rest — is increasingly common among engineering teams.
Use Cases
### 1. Agentic Coding and Autonomous Development Workflows
GPT-5.5's Terminal-Bench 2.0 score of 82.7% (vs 69.4% for Claude Opus 4.7) makes it the strongest model for complex command-line coding workflows. Use it for: repository-level refactors, automated debugging loops, multi-step CI/CD pipeline work, and any coding agent that chains tool calls across sessions. Its 97.4% first-attempt tool-call success rate (vs 94.2% for Opus) reduces retry loops in production agent systems. Choose GPT-5.5 over Claude when your agent operates in terminal/shell environments or needs to coordinate multiple tools (file editing, testing, deployment) in a single autonomous run. Note: for multi-file refactoring specifically, Claude Opus 4.7 still leads on cross-file consistency.
### 2. Long-Document Analysis and Knowledge Work Under 1M Tokens
GPT-5.5's 74.0% score on MRCR v2 at 512K–1M tokens (vs 36.6% for GPT-5.4) makes it uniquely suited for workflows processing entire codebases, legal corpora, or large document sets in a single pass. Combined with GDPval's 84.9% across 44 professional occupations, the model excels at: investment banking modeling (88.5% on internal evals), financial document review (OpenAI's finance team processed 24,771 K-1 tax forms using GPT-5.5 in Codex), operational research synthesis, and multi-source competitive analysis. Choose over Gemini 3.1 Pro when tool reliability and structured output quality matter more than raw context window size (Gemini offers 2M).
### 3. Scientific Research and Domain-Specific Data Analysis
GPT-5.5 shows meaningful gains on GeneBench (25.0%, up from 19.0%) and BixBench (80.5%, up from 74.0%) for bioinformatics and genetics workflows. It contributed a novel proof about Ramsey numbers in combinatorics, verified in Lean. Researchers report using GPT-5.5 Pro for: multi-pass manuscript critique, stress-testing technical arguments, building custom mathematical visualization tools (an algebraic-geometry app from a single prompt in 11 minutes), and analyzing large gene-expression datasets. Choose GPT-5.5 over alternatives when the task requires persisting across the full research loop — exploring, gathering evidence, testing assumptions, and iterating — rather than one-shot question answering.
### 4. Enterprise Agent Orchestration and Tool-Use Pipelines
For teams building production agent systems that need to operate across multiple software tools — browsing, code execution, document creation, data analysis — GPT-5.5's tool surface (MCP, hosted shell, apply_patch, computer use, web search, file search) is the most mature in the market. Tau2-bench Telecom's 98.0% score demonstrates reliable multi-turn customer service workflows. BrowseComp's 84.4% confirms strong multi-step web research. The model's structured output compliance (98.1% schema-valid JSON) and lower tool-call error rate make it the safest choice for agent loops where a failed tool call breaks the entire chain. Choose over alternatives when tool-call reliability and schema compliance are gating requirements for production deployment.
Latest News
## Timeline of Key Developments
- **April 23, 2026**: GPT-5.5 and GPT-5.5 Pro released to API. Model launches with 1M context window, $5/$30 standard pricing, and strongest OpenAI safety safeguards to date. Available in ChatGPT for Plus, Pro, Business, and Enterprise users.
- **April 24, 2026**: API deployments go live. System card updated with additional safeguards. GPT-5.5 and GPT-5.5 Pro available in Responses and Chat Completions APIs.
- **May 5, 2026**: GPT-5.5 Instant becomes the default ChatGPT model, replacing GPT-5.3 Instant. Adds Gmail integration, memory source transparency panel, and conversation history search.
- **July 9, 2026**: GPT-5.6 released as successor, with GPT-5.5 now classified as superseded on BenchLM. GPT-5.6 becomes the preferred model in Microsoft 365 Copilot.
## Pricing Changes
Standard API pricing ($5/$30) represented a 2x increase over GPT-5.4 ($2.50/$15). Key cost tiers:
- Cached input: $0.50/1M (90% discount)
- Batch/Flex: $2.50/$15 (50% off)
- Priority: $12.50/$75 (2.5x)
- Overage (>272K input): 2x input, 1.5x output for full session
- Pro tier: $30/$180 (6x standard)
## Key Partnerships
- NVIDIA GB200/GB300 NVL72 co-design: GPT-5.5 was trained and served on NVIDIA's latest rack-scale systems, achieving GPT-5.4 latency parity despite higher capability. NVIDIA reports 10,000+ employees using Codex with GPT-5.5.
- Microsoft Azure: Full availability through Azure OpenAI Service with enterprise compliance (SOC 2, HIPAA, GDPR).
- Trusted Access for Cyber: New program giving verified cybersecurity defenders expanded access to GPT-5.5's advanced capabilities with fewer restrictions.
## Safety Developments
OpenAI deployed stricter classifiers for cyber risk, introduced tiered safeguard levels based on domain (biological/chemical and cybersecurity capabilities classified as "High" under Preparedness Framework), and engaged nearly 200 trusted early-access partners for pre-release testing. The Trusted Access program at chatgpt.com/cyber allows verified security professionals to access cyber-permissive model variants.
