GPT-5.5 Developer Guide: API, Pricing & Benchmarks (April 2026)

OpenAI released GPT-5.5 on April 23, 2026 — three days ago at the time of writing — and the model earns its version bump. It scores 88.7% on SWE-bench Verified, delivers a 60% reduction in hallucinations versus GPT-5.4, and introduces a 1-million-token context window alongside built-in computer use, MCP, hosted shell execution, and a new Skills system. It ships as three distinct variants with meaningfully different speed, accuracy, and cost trade-offs. This guide covers every confirmed detail: benchmarks, API pricing, variant selection, code examples, and how GPT-5.5 stacks up against this week’s other frontier model release, DeepSeek V4.

The Three GPT-5.5 Variants

For the first time, OpenAI shipped a frontier model family simultaneously with three distinct inference modes. Understanding the differences determines where each variant belongs in your stack.

GPT-5.5 (Standard)

The baseline model. Available in both the Chat Completions API and the Responses API via the identifier gpt-5.5. Optimized for the best balance of throughput, latency, and accuracy. This is the model powering ChatGPT for Plus and Enterprise users. OpenAI positions it as “the next step toward a new way of getting work done on a computer” — language that deliberately emphasizes end-to-end task completion over conversational assistance.

GPT-5.5 Thinking

Extended-reasoning variant. Uses visible chain-of-thought similar to the o-series models, allocating additional compute tokens to plan before generating the final response. The Thinking variant materially outperforms the standard model on problems requiring multi-step deduction: complex algorithmic design, graduate-level mathematics, adversarial security analysis. Latency is higher and pricing carries a premium, so this variant suits batch workflows and high-stakes automated pipelines more than interactive applications.

GPT-5.5 Pro

Highest-accuracy variant. Available exclusively via the Responses API (not Chat Completions). OpenAI’s description: “for tougher problems that benefit from more compute.” This is the model behind the Pro tier in ChatGPT. If you are building automated systems where accuracy per decision has direct business consequences — legal document analysis, financial modeling, security audit automation — GPT-5.5 Pro is the variant to benchmark first.

Benchmark Results

SWE-bench Verified: 88.7%

SWE-bench Verified is the closest thing the software engineering community has to a practical benchmark. It measures end-to-end autonomous issue resolution: the model reads a real GitHub issue, navigates the repository, writes a fix, and passes the existing tests — no scaffolding, no hints. GPT-5.5 scores 88.7%, resolving nine out of ten real-world software defects without human assistance.

To put that in context: the benchmark was calibrated so that 50% corresponds roughly to a competent human engineer working on an unfamiliar codebase. At 88.7%, GPT-5.5 is not just keeping pace with developers; it routinely outperforms them on this specific task class. For teams building autonomous coding agents, this score justifies using GPT-5.5 as the default model for issue-resolution pipelines.

MMLU: 92.4%

The Massive Multitask Language Understanding benchmark covers 57 domains from elementary school math to graduate-level STEM. GPT-5.5’s 92.4% places it firmly in the frontier tier alongside Claude Opus 4.7 and Gemini 3.1 Ultra. For knowledge-intensive applications — research assistants, domain Q&A, document analysis — this score signals reliable factual recall across a wide subject range without heavy retrieval infrastructure.

60% Fewer Hallucinations

OpenAI claims a 60% reduction in factual hallucinations versus GPT-5.4, measured on their internal evaluation suite. This figure needs independent replication on production workloads before it translates directly to trust, but directionally it reflects a post-training emphasis on groundedness. Teams currently running fact-checking layers on GPT-5.4 outputs should validate whether those layers are still necessary at scale — the error rate may have dropped enough to change the cost calculus.

Token Generation Speed

GPT-5.5 generates tokens approximately 20% faster than GPT-5.4. OpenAI attributes this to an efficiency optimization in the generation pipeline. For streaming applications, real-time coding assistants, and high-frequency tool-calling loops, the throughput improvement is directly visible to end users: faster first-token times, snappier completions, no integration changes required.

API Pricing and Access

Both GPT-5.5 and GPT-5.5 Pro are live in the OpenAI API as of April 24, one day after public launch. No waitlist, no special access request needed.

Model	Input (per 1M tokens)	Output (per 1M tokens)	Context Window	Max Output
gpt-5.5	$5.00	$30.00	1,050,000	128,000 tokens
gpt-5.5-pro	Not yet announced	Not yet announced	1,050,000	128,000 tokens
gpt-5.4 (prior)	$2.50	$15.00	128,000	16,384 tokens

GPT-5.5 standard is priced at exactly double GPT-5.4: $5 versus $2.50 per million input tokens, $30 versus $15 per million output tokens. OpenAI frames this as “a new class of intelligence” — a premium it believes the benchmark improvements justify. Batch and Flex pricing halve the standard API rate. If your pipeline can tolerate asynchronous processing with up to 24-hour turnaround, Batch mode remains the most cost-efficient option for high-volume workloads.

For comparison: Claude Sonnet 4.6 is $3/$15 per million tokens. Claude Opus 4.7 is $15/$75. GPT-5.5 standard slots between them in price but competes at the Opus tier in coding benchmarks. For agentic coding specifically, the cost-per-resolved-issue metric likely favors GPT-5.5 over Opus 4.7 if the 88.7% SWE-bench score translates to your actual workload distribution.

Context Window: 1 Million Tokens in the API

The 1,050,000-token context window is the full limit available in the Chat Completions and Responses APIs. In Codex (OpenAI’s agentic coding environment), the working context is capped at 400,000 tokens for latency reasons. Maximum output length is 128,000 tokens — the highest OpenAI has shipped on any public model, compared to 16,384 tokens on GPT-5.4.

One million tokens holds approximately 750,000 words of text. In software engineering terms, that is the full source code of a mid-size application — every file, every function, every test — plus conversation history and tool outputs, all in a single context window. This eliminates the need for retrieval-augmented generation in many agentic coding scenarios. Instead of building a RAG pipeline to surface relevant code, you load the full repository and reason over it directly. The tradeoff is cost: a single 1M-token request at $5/1M input costs $5. Use prompt caching aggressively on repeated large-context requests.

Six Built-In Capabilities

GPT-5.5 ships six integrated capabilities that previously required separate integrations or workarounds in the GPT-5.4 era:

• Computer Use: Native ability to navigate UIs, click elements, and operate desktop and web software. No separate computer-use endpoint — it is part of the core model’s tool suite in the Responses API.
• Hosted Shell: OpenAI provides a sandboxed execution environment. The model can run shell commands, execute test suites, and return stdout/stderr within a managed container. Directly useful for coding agents that need to verify changes actually compile and pass tests before returning results.
• Apply Patch: Produces and applies code diffs in patch format rather than requiring full-file rewrites. For large files, this cuts output token cost significantly and eliminates a common failure mode where a full-file rewrite inadvertently breaks unchanged sections.
• Skills: OpenAI’s new capability packaging system. Skills are reusable instruction modules that load into a session to give the model persistent behavioral patterns. Think structured system-prompt modules with optimized activation for specific task classes — similar in concept to what Claude Code skills offer on the Anthropic side.
• MCP (Model Context Protocol): Full native support for MCP servers. GPT-5.5 is now a first-class citizen in MCP-enabled agent infrastructure. The same MCP servers and tool integrations used with Claude Code and other MCP-compatible systems work natively with GPT-5.5.
• Web Search: Built-in web search as a native tool. No Bing API key, no separate integration, no workaround required.

API Integration Examples

Chat Completions — Standard Variant

from openai import OpenAI

client = OpenAI(api_key="your-api-key")

response = client.chat.completions.create(
    model="gpt-5.5",
    messages=[
        {"role": "system", "content": "You are a senior software engineer."},
        {"role": "user", "content": "Review this diff and identify security vulnerabilities."}
    ],
    max_tokens=8192,
    temperature=0.1)

print(response.choices[0].message.content)

Responses API — GPT-5.5 Pro with Tools

from openai import OpenAI

client = OpenAI(api_key="your-api-key")

response = client.responses.create(
    model="gpt-5.5-pro",
    input="Analyze this contract and flag clauses with unlimited liability exposure.",
    tools=[
        {"type": "web_search"},
        {"type": "apply_patch"}
    ])

print(response.output)

Streaming with Hosted Shell

stream = client.chat.completions.create(
    model="gpt-5.5",
    messages=[{"role": "user", "content": "Fix the failing unit tests in this repository."}],
    tools=[
        {"type": "hosted_shell"},
        {"type": "apply_patch"}
    ],
    stream=True
)

for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

GPT-5.5 vs. the Frontier

The week of April 21–26 was unusually active: two frontier-tier models shipped within 48 hours of each other. GPT-5.5 on April 23, then DeepSeek V4-Flash and V4-Pro on April 24. The comparison is directly relevant for teams evaluating both.

Model	SWE-bench	Context	Input Price	License
GPT-5.5	88.7%	1M tokens	$5.00/1M	Proprietary
Claude Opus 4.7	~85%	200K tokens	$15.00/1M	Proprietary
DeepSeek V4-Pro	Claimed SOTA (open)	1M tokens	$1.74/1M	MIT
Gemini 3.1 Ultra	—	2M tokens	Variable	Proprietary

DeepSeek V4-Pro at $1.74/1M input is the most obvious cost alternative. The practical questions are: ecosystem lock-in (OpenAI’s Codex, computer use, and Skills ecosystem has no direct V4 equivalent), tooling maturity, and whether V4-Pro’s claimed open-weight SOTA matches GPT-5.5’s confirmed 88.7% on your specific task distribution. Both models support 1M context and OpenAI-compatible APIs. Both are worth benchmarking on your own evaluation set before committing at scale.

Which Variant Should You Use?

• GPT-5.5 Standard — Default for most applications. Agentic coding, document analysis, content generation, interactive tools. Best balance of speed, quality, and cost.
• GPT-5.5 Thinking — Complex reasoning, multi-constraint planning, graduate-level problem-solving, adversarial security analysis. Use where latency budget allows and correctness matters more than throughput.
• GPT-5.5 Pro — High-stakes automated workflows where accuracy per decision has direct business consequence: legal review automation, financial analysis pipelines, security audit systems. Responses API only.

Conclusion

GPT-5.5 is a step-function improvement over GPT-5.4: 88.7% SWE-bench, 1M context, built-in computer use and MCP, a three-variant architecture, and the highest maximum output length OpenAI has shipped. The $5/$30 pricing is twice GPT-5.4, which is a real cost consideration for high-volume pipelines, but the cost-per-resolved-issue metric likely justifies it for agentic coding workloads.

The model is live today on any active OpenAI API account and already powering Codex on NVIDIA infrastructure. If you are running GPT-5.4 in production, the upgrade path is a one-line model identifier change. Benchmark on your own task distribution first — especially comparing against DeepSeek V4-Pro if cost is a constraint — then flip it in production. The 60% hallucination reduction alone may justify the price differential for applications where factual accuracy drives downstream trust.