What Is GPT-5.5? Features, Specs, Pricing, and Use Cases

GPT-5.5 is an OpenAI reasoning model released on April 23, 2026 for coding, professional work, computer use, research, and long-running tasks that involve tools. OpenAI made GPT-5.5 and GPT-5.5 Pro available in the API on April 24.

The company introduced it as “a new class of intelligence for real work,” emphasizing a change in how the model operates rather than only how it answers. GPT-5.5 was designed to take a messy goal, form a plan, inspect context, use tools, check the result, recover from ambiguity, and continue across several steps.

There is one essential update for readers arriving later: GPT-5.5 is not OpenAI’s current recommended starting point in August 2026. The current reasoning guide says to start most new reasoning workloads with GPT-5.6. GPT-5.5 remains available and important, especially for systems already tested around its behavior, but “latest flagship” descriptions from its launch are now historical.

GPT-5.5 at a Glance

The standard model supports streaming, structured outputs, function calling, file search, file uploads, image input, web search, and prompt caching. Its documented tool set includes function calling, web search, file search, tool search, image generation, code interpreter, hosted shell, patch application, skills, computer use, and MCP.

Those specifications explain why the model was positioned for complete workflows instead of short question answering.

What Makes GPT-5.5 Different from a Basic Chat Model

A basic chat model receives a prompt and produces a response. GPT-5.5 can do that, but its more important use case is a reasoning loop:

1. Understand the user’s intended outcome.
2. Decide what information or actions are needed.
3. Call tools and inspect their results.
4. Re-plan when the environment does not match the assumption.
5. Verify the final artifact or answer.
6. Stop when the task’s completion conditions are met.

This is why OpenAI emphasized agentic coding, knowledge work, and computer use. The model’s value grows when the task crosses multiple files, applications, data sources, or production stages.

Examples include fixing a bug and running the relevant tests, researching a market and building a spreadsheet, reviewing a group of documents and drafting a report, or converting a creative brief into a storyboard and asset plan.

The model does not gain unlimited autonomy. It can act only through the tools, permissions, and orchestration a product gives it. A safe workflow still needs narrow permissions, approval boundaries, logs, and verification.

The 1,050,000-Token Context Window

GPT-5.5’s 1.05-million-token context window allows it to work with unusually large inputs: a substantial codebase, long document collection, complex conversation history, or a project brief with many supporting artifacts.

That does not mean every task should fill the window. Large inputs cost more, take longer to process, and can make important instructions harder to find. Retrieval and context selection remain useful. Feed the model the authoritative files and constraints it needs rather than every artifact available.

Context also includes reasoning tokens, which are not shown as raw chain-of-thought but still occupy space and are billed as output tokens. Leave headroom for the model to reason and generate the requested result.

OpenAI applies special pricing to very large GPT-5.5 sessions. When prompts exceed 272,000 input tokens, the full session is priced at twice the normal input rate and 1.5 times the normal output rate for standard, Batch, and Flex processing. That threshold makes deliberate context management an economic requirement, not just a technical preference.

Reasoning Effort Explained

The reasoning.effort parameter controls how much model work GPT-5.5 applies. It is a latency, cost, and quality tradeoff.

None

Useful for latency-critical tasks that do not need multi-step reasoning or chained tools. OpenAI recommends trying low first and moving to none only when the workload demands it.

Low

Designed for efficient planning, search, tool use, data analysis, execution-oriented coding, classification, and support flows. It is a good first test when latency and cost are important.

Medium

The default and recommended general starting point for GPT-5.5. It balances quality, reliability, cost, and performance for agentic coding, research, documents, spreadsheets, slides, and long-horizon delegation.

High

Better suited to hard debugging, deep planning, high-value analysis, long research, and complex agent tasks where quality matters more than latency.

Xhigh

Intended for difficult, asynchronous, or long-running work. OpenAI advises using it only when evaluations show a clear improvement worth the extra time and cost.

Do not treat reasoning effort as a substitute for a good task definition. If the prompt does not identify the target file or the acceptance criteria, higher effort can produce a more detailed wrong answer. Fix goals, constraints, tool schemas, and verification before spending more compute.

GPT-5.5’s Strongest Use Cases

Agentic coding

GPT-5.5 was a major coding release because it improved the behaviors needed around code generation: understanding a whole system, diagnosing ambiguous failures, editing several related files, testing changes, and carrying a task through to completion.

At launch, OpenAI reported 82.7% on Terminal-Bench 2.0, 58.6% on SWE-Bench Pro, and stronger performance than GPT-5.4 on its internal Expert-SWE long-horizon benchmark. OpenAI also reported that it used fewer tokens across the tested coding tasks.

Benchmarks do not predict every repository, but the evaluated behaviors are relevant: planning, command-line work, issue resolution, and tool coordination.

Knowledge work

GPT-5.5 was designed to move through a complete office workflow: locate information, decide what matters, use software, check outputs, and turn raw material into a document, spreadsheet, presentation, or decision memo.

OpenAI reported 84.9% wins or ties on GDPval, 78.7% on OSWorld-Verified, and 98.0% on Tau2-bench Telecom with the original prompts. The release also described internal uses in finance, communications, reporting, and data analysis.

The model is useful when the deliverable has structure and can be verified. A spreadsheet with formulas, a research table with sources, or a formatted report provides clearer checks than an open-ended request to “analyze this.”

Scientific and technical research

The release highlighted multi-stage scientific workflows: exploring a question, collecting evidence, analyzing data, testing assumptions, interpreting results, and deciding what to try next. GPT-5.5 showed gains on genetics, bioinformatics, mathematical reasoning, and research-tool construction evaluations.

Scientific use still requires expert review. The model can accelerate analysis and implementation; it does not validate a discovery merely by producing a confident explanation.

Computer use

With the right harness, GPT-5.5 can interpret a screen, click, type, navigate applications, and move data between tools. Computer use expands what an agent can do but also increases risk. Products should limit accessible applications, protect sensitive data, require approval for consequential actions, and verify state before writing.

Creative planning and production

GPT-5.5 can turn a broad creative goal into a shot list, prompt set, review rubric, asset inventory, or publishing plan. The actual media can then be produced with a dedicated visual system.

Creators can compare engines in the DeepFake AI model library, develop the brief through text-to-video, or animate an approved concept frame with image-to-video.

This separation is useful: the language model coordinates intent and constraints; the media model renders the visual output.

Responses API vs. Chat Completions

GPT-5.5 supports both the Responses API and Chat Completions API, but OpenAI recommends Responses for reasoning models. Responses is designed for multi-step state, tool calling, and reasoning-aware workflows.

Using previous_response_id is often the simplest way to continue a task because OpenAI preserves the prior response state. If an application manually replays the history, it must preserve every relevant field.

For long-running and tool-heavy GPT-5.5 workflows, OpenAI specifically recommends retaining the assistant message phase value:

• phase: "commentary" for intermediate progress updates or preambles before tool calls.
• phase: "final_answer" for the completed result.

Dropping this distinction can make an intermediate message look final and cause premature stopping.

Tools and Structured Outputs

GPT-5.5’s function calling and structured-output support help turn model decisions into reliable software actions. A tool should expose a narrow operation with explicit parameters and return observable results.

For example, update_invoice_status(invoice_id, status) is safer and easier to validate than a universal shell function. A file-editing tool should return the changed path and whether the patch applied. A publishing tool should distinguish draft creation from public release.

Structured outputs are useful when another system consumes the response. Validate the schema on the application side even if the model promises to follow it. Tool calls can still fail because of permissions, unavailable resources, stale IDs, or network errors.

MCP support allows external systems to expose tools and resources through a common protocol. It reduces custom integration work but does not remove the need for access control.

GPT-5.5 Pricing

At the time of this guide, OpenAI’s model page lists standard API text pricing per one million tokens as:

• Input: $5.00
• Cached input: $0.50
• Output: $30.00

Batch pricing, Flex processing, tool fees, regional processing, and long-context multipliers can change the real cost. Regional data-residency endpoints carry a documented 10% uplift for GPT-5.5.

Estimate cost with a representative workload. Include retries, reasoning output, tool-generated context, and failed runs. The lowest price per token does not guarantee the lowest cost per completed task; a model that needs fewer attempts may be cheaper in practice.

Always confirm current pricing on the official page before budgeting.

GPT-5.5 Pro

gpt-5.5-pro uses more compute to produce more precise answers on difficult problems. It has the same 1,050,000-token context window, 128,000-token maximum output, and December 1, 2025 knowledge cutoff, but its endpoint and effort options differ.

GPT-5.5 Pro is available through Responses and Batch. It supports medium, high by default, and xhigh reasoning effort. Some requests may take several minutes, so OpenAI recommends background mode to avoid timeouts.

Pro does not receive a cached-input discount. It is appropriate only when evaluation shows a quality gain that justifies higher latency and cost. A routine interactive assistant usually benefits more from faster feedback than maximum compute.

Safety and Cybersecurity

OpenAI launched GPT-5.5 with stronger safeguards, including tighter controls around higher-risk cybersecurity activity and protections against repeated misuse. The company also offered expanded access for verified defensive work through Trusted Access for Cyber.

Application developers remain responsible for their own controls. Use least-privilege tools, isolate execution environments, protect secrets, limit network destinations, and require human approval for destructive or external actions. Model-level safeguards and system-level security solve different parts of the problem.

Limitations to Keep in Mind

GPT-5.5 is more capable, not infallible.

• It can still hallucinate facts, sources, files, and completed actions.
• Its knowledge cutoff is December 2025, so current facts require search or connected data.
• Long context can contain contradictions or irrelevant material.
• Tool use can magnify a reasoning error into a real side effect.
• Image input does not guarantee perfect reading of small text or complex diagrams.
• High reasoning effort increases latency and cost without guaranteeing a better result.
• Benchmark gains may not transfer to a specialized internal workload.

Build verification into the task. For code, run tests and a build. For research, open sources and check dates. For documents, validate formulas and totals. For external writes, confirm the exact target and final state.

GPT-5.5 vs. GPT-5.6 in August 2026

OpenAI’s current reasoning guide says to start with gpt-5.6 for most reasoning workloads. It points to GPT-5.6 Sol for the highest-intelligence problems that can tolerate more latency, Terra for lower cost, and Luna for the lowest cost and latency.

That makes GPT-5.6 the natural evaluation target for a new project. GPT-5.5 may still be the right choice when:

• An existing production system has passed evaluations on GPT-5.5.
• A regulated change process makes an immediate model migration expensive.
• The team depends on a specific snapshot behavior.
• Price, latency, or feature testing favors it for a particular task.
• The migration is staged and needs a stable baseline.

Do not upgrade on the model name alone. Compare completion rate, quality, latency, token usage, tool errors, and human intervention on real tasks.

Prompting GPT-5.5 Effectively

Reasoning models generally respond well to clear goals and constraints without a long script for every hidden thought.

A good task includes:

• The concrete outcome.
• Authoritative context and target resources.
• Boundaries and prohibited actions.
• Required tools or evidence.
• Output format.
• Completion and verification criteria.

For example:

Review the attached launch plan. Identify schedule conflicts, missing owners, and dependencies. Return a table with issue, evidence, impact, owner, and recommended action. Do not invent dates or owners. Cite the exact section supporting each finding. Finish by listing unresolved questions.

This prompt gives the model a job, a source of truth, a schema, and a definition of done.

Final Takeaway

GPT-5.5 was an important OpenAI release because it combined long context, reasoning, broad tool support, coding strength, computer use, and professional-work capability in one model family. It helped move the conversation from “What can the model say?” toward “What task can the system reliably complete?”

Its key specifications remain substantial: a 1.05-million-token context window, 128,000-token output limit, text and image input, structured output, prompt caching, and a broad tool set. It is available through Responses, Chat Completions, and Batch, with medium reasoning as the default.

As of August 2026, however, GPT-5.6 is the official starting recommendation for new reasoning workloads. GPT-5.5 is best understood as a capable available model, a stable baseline for existing systems, and the release that established many of the agentic design patterns builders now carry into the next generation.

Official references

• Introducing GPT-5.5
• GPT-5.5 API model page
• GPT-5.5 Pro API model page
• OpenAI reasoning guide