On April 23, 2026, OpenAI released GPT-5.5, which it called its smartest and most intuitive model yet and its strongest agentic coding model to date. The release targeted long, messy, multi-tool work: writing and debugging code, operating software, researching and producing documents with less step-by-step supervision. GPT-5.5 reached ChatGPT and Codex first and arrived in the API a day later, at double GPT-5.4's per-token price, with OpenAI arguing that its efficiency makes the real cost per task competitive. It also shipped with OpenAI's strictest cyber safeguards so far.

Key Facts#

  • Release dates: April 23, 2026, in ChatGPT and Codex; April 24 in the API as gpt-5.5 and gpt-5.5-pro, with the snapshot gpt-5.5-2026-04-23.
  • API pricing per million tokens: gpt-5.5 at $5 input and $30 output, gpt-5.5-pro at $30 and $180. Batch and Flex run at half price, and Priority costs 2.5 times the standard rate.
  • Context: a one-million-token window in the API and 400,000 tokens in Codex. Microsoft's Azure documentation lists 1,050,000 tokens in total, with a combined prompt and output budget of about 922,000 tokens in the Responses API.
  • Vendor-reported results: 82.7% on Terminal-Bench 2.0 (GPT-5.4: 75.1%), 58.6% on SWE-Bench Pro, 78.7% on OSWorld-Verified and 84.9% on GDPval.
  • Speed: OpenAI says GPT-5.5 matches GPT-5.4's per-token latency in real-world serving while using significantly fewer tokens on Codex tasks.
  • Safety classification: "High" for biological and chemical and for cybersecurity capabilities under OpenAI's Preparedness Framework, below the "Critical" level.
  • Early access: nearly 200 trusted partners tested the model before release.

What Happened#

OpenAI's pitch was that GPT-5.5 can take "a messy, multi-part task" and plan, use tools, check its work and keep going with less hand-holding. Agentic coding was the centerpiece. Beyond the Terminal-Bench 2.0 result, OpenAI cited an internal Expert-SWE evaluation of long-horizon coding tasks with a median estimated human completion time of 20 hours, where GPT-5.5 scored 73.1% against 68.5% for GPT-5.4. In OpenAI's comparison table, Claude Opus 4.7 scored 69.4% and Gemini 3.1 Pro 68.5% on Terminal-Bench 2.0.

Early testers quoted in the announcement focused on judgment rather than raw output. Dan Shipper, founder and CEO of Every, called it "the first coding model I've used that has serious conceptual clarity." Cursor said the model stays on task significantly longer without stopping early.

The infrastructure story was unusual. OpenAI said GPT-5.5 was co-designed for, trained with and served on NVIDIA GB200 and GB300 NVL72 systems, and that Codex and GPT-5.5 itself helped optimize the serving stack. In one example, Codex analyzed weeks of production traffic and wrote new load-balancing heuristics that increased token generation speed by more than 20%.

The API launch was staggered for safety reasons. On April 23, OpenAI said API deployments need different safeguards and promised API access "very soon"; it followed through on April 24. OpenAI also deployed stricter classifiers for potentially harmful cyber requests, warning that some users might find them annoying at first, and expanded its Trusted Access for Cyber program so verified defenders can use more permissive capabilities.

Background#

GPT-5.5 arrived seven weeks after GPT-5.4, which had introduced native computer use and tool search, and one week after Anthropic released Claude Opus 4.7. The pace reflected how quickly agentic coding had become the main battleground among frontier labs, with agent-oriented benchmarks such as Terminal-Bench and SWE-Bench Pro taking center stage in launch announcements.

Pricing had also started to move upward. GPT-5.4 had already raised per-token prices over GPT-5.2, and GPT-5.5 doubled them again, while OpenAI argued that fewer tokens and fewer retries would offset the increase.

Why It Matters for Developers#

Start with economics. A doubling of per-token prices can still lower costs if a model finishes tasks in fewer tokens and fewer attempts, but you only know by measuring. Track cost per successful task, not per request, and compare models on your own workloads; our guides to LLM observability and cost control and evaluating AI applications in .NET show how.

Next, mind the context budget. Microsoft documents that, in the current Responses API implementation, prompt and output tokens share a combined budget of about 922,000 tokens, and that a request can return HTTP 200 with an incomplete response when the budget runs out. That failure mode is easy to miss, so check the finish reason in code:

C#
using Microsoft.Extensions.AI;
using OpenAI.Chat;

string apiKey = Environment.GetEnvironmentVariable("OPENAI_API_KEY")
    ?? throw new InvalidOperationException("Set OPENAI_API_KEY.");

IChatClient client = new ChatClient("gpt-5.5", apiKey).AsIChatClient();

string bundle = await File.ReadAllTextAsync("incident-bundle.txt");
var options = new ChatOptions
{
    MaxOutputTokens = 16_000,
    Reasoning = new ReasoningOptions { Effort = ReasoningEffort.Medium },
};

var response = await client.GetResponseAsync($"Find the root cause:\n{bundle}", options);

if (response.FinishReason == ChatFinishReason.Length)
{
    Console.Error.WriteLine("Output was truncated: trim the prompt or raise the budget.");
}

The cyber safeguards matter if you build security tooling. Stricter classifiers can refuse legitimate requests from vulnerability scanners, code auditors or incident-response assistants. Design for refusals as a normal outcome, and if you work on defensive security, look at OpenAI's trusted-access route rather than prompt workarounds. Our guide to responsible AI and LLM security covers handling refusals gracefully.

Finally, the agentic-coding gains are most visible inside tools such as Codex, Cursor and GitHub Copilot. Our guide to AI-assisted .NET development covers how to evaluate a new model against your own repositories before changing team defaults.

What's Next#

OpenAI moved on quickly. A GPT-5.6 family with Sol, Terra and Luna models followed on July 9, 2026, and the GPT-6 generation began on September 3, 2026. GPT-5.5 remains available in the API and in Azure, and Microsoft's documentation notes that OpenAI also refers to its continuously updated chat model as GPT-5.5 Instant.

The open question is where pricing goes next. OpenAI's later releases have not moved prices in a single direction, which is one more reason to treat model routing, sending easy work to cheaper models, as a core architectural skill rather than a late optimization.

Sources#