What is GPT-5.5?
GPT-5.5 is OpenAI’s flagship model for the most complex professional work, released on 23 April 2026 and described in exactly those words on its own model page. OpenAI writes the name GPT-5.5, while this site’s own label for it — the one the heading and the navigation show — drops the hyphen and reads GPT 5.5; both appear here, and the model code settles which is meant. Its API model code is gpt-5.5, the default snapshot is gpt-5.5-2026-04-23, the documented context window is 1,050,000 tokens and the documented maximum output is 128,000 tokens, with text and image as the documented inputs and text as the documented output. Its knowledge cutoff is 1 December 2025. Three statements from the official pages define how it behaves in practice. The first is a directory fact worth reading before anything else: OpenAI’s Featured models list is now GPT-6 Astra, GPT-5.6 Terra and GPT-5.6 Luna, and the word flagship now sits on GPT-5.6 Sol, so this generation has been succeeded — while the deprecations page lists gpt-5.5 nowhere at all. Succeeded and deprecated are not the same thing, and both official pages are reported here rather than one being chosen. The second is a behaviour change rather than a specification: the prompting guide says the default reasoning effort is now medium, and warns in as many words that a higher effort is not automatically a better answer — with conflicting instructions, thin stopping conditions or an open tool surface it can produce overthinking, unnecessary searching and degraded output. The third is that two official pages disagree on how many effort levels exist: the model page lists five, reading none, low, medium (default), high and xhigh, while the table in the reasoning guide carries a sixth, max. Both counts are given here. Two further statements describe how the release was handled rather than how the model performs: OpenAI rates its biological and cyber capability High under its Preparedness Framework and states that this falls short of Critical, and the announcement credits internal and external red-teaming plus feedback from close to 200 trusted early-access partners. The provider’s own summary of what it is good at is a list rather than a score — writing and debugging code, online research, data analysis, producing documents and spreadsheets, operating software, and carrying work across tools to completion.
What GPT-5.5 is worth using for
Treat the effort dial as a cost, not a quality slider
The official coverage of reasoning effort is unusually direct about its own default being wrong for some inputs. The model page documents five levels — none, low, medium (default), high and xhigh — and the reasoning guide’s table adds a sixth, max, so the two official pages do not agree on the list; both counts are reported here rather than averaged into one. The prompting guide states that the default is medium and calls it the balanced starting point for quality, reliability, latency and cost, then immediately warns that higher effort is not the same as better output: when instructions conflict, when stopping conditions are thin, or when the tool surface is wide open, a higher level can produce overthinking, needless retrieval and lower quality. Two billing details travel with the dial. This model supports interleaved thinking, so it can emit visible tokens before and between its reasoning steps and reason between tool calls; and its reasoning tokens are never shown to you, yet they still occupy the context window and are still billed as output tokens. Because of the second, a level that looks free at the top of a conversation is priced into everything that follows. This chat box has no level selector and no off switch — the catalog entry behind the page marks reasoning as unsupported — so the level that runs here is the one the provider defaults to.
A model defaults its reasoning effort to medium, offers five levels on its model page and six in its reasoning guide, and its own documentation warns that raising it can reduce answer quality. Give me a short procedure for choosing a level per task shape, name the three input shapes its docs call unsafe at high effort, and explain how reasoning tokens change what I actually pay.
Write the stopping condition as carefully as the goal
Three statements in the official prompting guide describe a model that is easier to control and easier to mis-drive at the same time. Instruction following is said to be more literal and more thorough, which the guide treats as both an improvement and a hazard: a prompt that asks for an improvement without naming the invariant to preserve gets exactly the literal instruction it was given. Coding workflows are said to need stronger orchestration than before, which the guide unpacks into reuse, subagent delegation, test expectations, acceptance criteria, and an explicit rule for when to keep going versus when to stop and ask. And the guide is blunt that this should not be adopted as a drop-in replacement for gpt-5.2 or gpt-5.4 — it asks callers to treat it as a new model family, start from a minimal prompt and rebuild their baseline instead. Two smaller instructions are worth copying verbatim because they change the prompt rather than the model: delete the current date from prompts, since the model already knows the current UTC date, and lay a prompt out with its static part first and its dynamic part last so that caching can do its work. Everything above is prompt-writing advice rather than a control this page can give you: the composer here exposes one text field, so there is no temperature dial, no reasoning-level selector and no orchestration setting.
A model’s own guide says it now follows instructions more literally, needs stronger orchestration for coding work, and should not be treated as a drop-in replacement for the previous version. Rewrite the prompt below so that it states its success criteria and stopping condition explicitly, names the invariant I care about, and puts the static instructions before the dynamic ones.
Image input is kept at full fidelity unless you say otherwise
The default for image input changed with this generation, and the old habit now costs more than it used to. Where image_detail is absent or set to auto, the model page says images are handled as original rather than downscaled, up to a ceiling of 10,240,000 pixels and a 6,000-pixel edge; high lowers that ceiling to 2,500,000 pixels and a 2,048-pixel edge, and low downsamples more aggressively for anything above a 512-pixel edge. The practical reading is that the cheap setting is no longer the default one, and that a screenshot-heavy prompt now consumes image tokens at a different rate than the same prompt did on the previous generation unless the caller sets the detail level explicitly. The documented input modalities are text and image, and this workspace accepts JPEG, PNG and WebP files alongside PDFs at up to 24 MB each, so the attachment path exists here. The detail level itself has no control on this page, which means whatever the API defaults to is what runs. PDF handling is worth stating plainly too, because it is a different mechanism rather than the same one: a PDF uploaded here is converted to extracted text before it is sent, so a scanned document or a complex layout may lose content unless it is attached as an image or pasted as text instead.
A model now treats an unset image detail level as full fidelity up to 10,240,000 pixels, where the previous generation downscaled by default. Tell me how that changes the token cost of a screenshot-heavy prompt, which images I should explicitly downgrade, and what the downgrade costs me in accuracy.
Answer length is a setting, and the default is not the shortest
The official stance on answer length has moved. The prompting guide says this generation is more concise and direct by default, that text.verbosity defaults to medium, and that low is often the better starting point for a shorter answer — which is an official statement that the default is not the terse one. Length and price interact in two documented ways. The announcement claims comparable per-token latency with the previous generation while completing the same Codex tasks in fewer tokens, and the pricing commentary makes the related claim that this model costs more per token than GPT-5.4 while being, in OpenAI’s words, smarter and more token-efficient. Neither claim converts into a number on this page, and both are reported as the provider’s own. The verbosity parameter itself has no control in this composer, so the default applies; if you want shorter answers here, the practical route is to ask for them in the prompt and to ask for a specific shape — a list of a fixed length, a table, or a single paragraph — rather than for a general shortening.
A model’s documentation says its answers are more concise by default, that its verbosity parameter defaults to medium, and that low is often a better starting point. Give me three closing lines I can append to a prompt to control answer length precisely, and tell me which one to use when I need an output of a fixed length.
Official GPT-5.5 specifications
- Developer
- OpenAI
- Model ID
- gpt-5.5
- Official context window
- 1,050,000 tokens
- Official maximum output
- 128,000 tokens
- Official inputs
- Text and image
- Output
- Text
GPT-5.5 and GPT-5.4 side by side
Every value below comes from a single page — the Evaluations summary table in OpenAI’s announcement post, checked on 29 September 2026 — so one official page carries both columns. The left column is this page’s model and the right column is GPT-5.4, the generation it succeeds. Rows are matched by evaluation name, never by position. The table prints six model columns in the order GPT-5.5, GPT-5.4, GPT-5.5 Pro, GPT-5.4 Pro, Claude Opus 4.7 and Gemini 3.1 Pro; only the first two are reproduced here, because those are the only two this page has a reason to compare. All nine rows carry a GPT-5.4 value, so all nine are available for this pair, and all nine move in this model’s favour. The narrow wins are stated rather than dropped: Toolathlon moves from 54.6 to 55.6, GDPval from 83.0 to 84.9, and Expert-SWE from 68.5 to 73.1, against much wider gaps on FrontierMath Tier 4 at 35.4 against 27.1 and Terminal-Bench 2.0 at 82.7 against 75.1. That is a claim about this pair and nothing wider. A second official table points the other way and belongs with these numbers rather than apart from them. On MRCR v2 eight-needle, the long-context recall benchmark the same announcement publishes band by band, this model leads at 4K–8K and 8K–16K and again from 128K upward, but it scores lower than GPT-5.4 in three bands: 96.5 against 97.2 at 16K–32K, 90.0 against 90.5 at 32K–64K and 83.1 against 86.0 at 64K–128K. The advantage opens sharply only in the top two bands, where it reaches 81.5 against 57.5 and 74.0 against 36.6. One further caution is recorded here because it is a conflict between two official sources rather than between an official page and a reseller: for at least one evaluation, a third party’s official model card publishes a different figure for GPT-5.5 than OpenAI does. This page reports only OpenAI’s value, does not present the two as a range, and does not carry the other figure into the table.
| Provider benchmark | GPT 5.5 | GPT-5.4 |
|---|---|---|
| Terminal-Bench 2.0 | 82.7% | 75.1% |
| Expert-SWE (Internal) | 73.1% | 68.5% |
| GDPval (wins or ties) | 84.9% | 83.0% |
| OSWorld-Verified | 78.7% | 75.0% |
| Toolathlon | 55.6% | 54.6% |
| BrowseComp | 84.4% | 82.7% |
| FrontierMath Tier 1–3 | 51.7% | 47.6% |
| FrontierMath Tier 4 | 35.4% | 27.1% |
| CyberGym | 81.8% | 79.0% |
How to use GPT-5.5 online
Choose the model
This page opens with GPT-5.5 selected. The model navigation on the left and the selector inside the chat box stay in sync, and switching models keeps the current conversation, so start a new chat when you want a clean context. Two spellings of the name appear on this page and both are correct in their own place: the official model code is gpt-5.5, and the site’s own display label, which is what the navigation and the heading show, omits the hyphen.
Write the task the way you would brief a person
The documented inputs are text and image, and this composer accepts both: paste text directly, or attach JPEG, PNG and WebP files or a PDF up to 24 MB each. Three habits from the official prompting guide are worth adopting before you start. State the invariant you want preserved as well as the outcome you want, because instruction following is described as more literal than before. Write the stopping condition explicitly, since the guide says coding work in particular now needs stronger orchestration. And leave the current date out of the prompt, because the model already knows it; putting the static part of the prompt first and the changing part last also helps caching. Any prompt in this guide can be sent as it stands or edited first.
Sign in, send and check the cost
Sending a message needs an account with available points. The composer reads a reference price from the live model catalog and a finished request reports what it actually cost. Two ceilings are set by this site rather than by the model, so plan around them: replies are requested with at most 4,096 output tokens, and there is no way to raise or inspect the reasoning effort from here. Follow up with corrections rather than restarting, and re-check any price against the official pages linked below before relying on it.
What GPT-5.5 costs
Two kinds of price are published for this model, and neither of them is dated. The official pricing and model pages, checked on 29 September 2026, list two standard tiers for gpt-5.5. At or under 272,000 input tokens the rate is $5.00 per million input tokens, $0.50 per million cached input tokens and $30.00 per million output tokens; above that threshold it becomes $10.00, $1.00 and $45.00. The model page’s own wording for the rule is that prompts with more than 272K input tokens are priced at 2x input and 1.5x output for the full session, and AWS Bedrock’s model card states the same rule from the other side — that long-context rates apply to all input and output tokens, not just the tokens above 272,000. The two sentences are two halves of one rule, and the consequence is worth spelling out: crossing the threshold reprices the whole session rather than only the excess. Three standing modifiers sit on top of those rates. Batch and Flex are half the standard rate, Priority is 2.5 times it, and regional processing for data residency adds 10%. These are permanent tiers rather than promotions, and that distinction is the point of this paragraph: OpenAI publishes no dated or limited-time offer for this model on any page checked — the model page and the pricing page both carry none — so there is no time-limited price to record here and nothing on this page that will expire. Two adjacent figures are easy to mix into this table and are not part of it. gpt-5.5-pro is a separate model ID, priced at $30 per million input and $180 per million output with no cached-input discount, and it does not support Chat Completions at all. And the quick-comparison row the same model page prints for GPT-5.4 reads $2.50, $0.25 and $15.00, which is the previous generation rather than a discount on this one. This site charges none of those figures. It bills points, and the catalog reference price behind this page is 1,220 points per million input tokens and 7,200 points per million output tokens, with cached reads at 120 points per million and web search at 1 point per search. That price is not a conversion of any official number — no published rate relates the two currencies. Resale prices quoted by third-party marketplaces differ from the official pages and are not used anywhere on this page.
View points and plans ↗GPT-5.5 FAQ
Where does GPT-5.5 finish in OpenAI’s own nine-row table?
It finishes ahead of GPT-5.4 on all nine rows, and the margin varies a great deal between them. The widest gaps are FrontierMath Tier 4 at 35.4 against 27.1 and Terminal-Bench 2.0 at 82.7 against 75.1, followed by OSWorld-Verified at 78.7 against 75.0 and CyberGym at 81.8 against 79.0. The narrowest are Toolathlon at 55.6 against 54.6 and GDPval at 84.9 against 83.0. Because every row moves the same way, the table supports a statement about this pair and nothing wider. A second official table published alongside it complicates the picture in one specific way: on MRCR v2 eight-needle, this model scores below GPT-5.4 in three of the eight length bands, 16K–32K, 32K–64K and 64K–128K, and only opens a large lead in the top two bands.
Is GPT-5.5 deprecated now that newer generations exist?
No. The official deprecations page lists gpt-5.5 in neither its upcoming nor its past deprecation table, so nothing about this model is scheduled for removal on the pages checked on 29 September 2026. It has, however, been succeeded: OpenAI’s model catalog now features GPT-6 Astra, GPT-5.6 Terra and GPT-5.6 Luna, and the word flagship sits on GPT-5.6 Sol rather than on this model, even though the GPT-5.5 model page still describes it as a flagship model for the most complex professional work. Succeeded and deprecated are different states, and this page reports both official pages rather than picking one. The honest sentence is that this generation has been superseded, not that it has been retired.
What is GPT-5.5’s context window and maximum output?
The model page documents a context window of 1,050,000 tokens and a maximum output of 128,000 tokens, with the default snapshot gpt-5.5-2026-04-23 and a knowledge cutoff of 1 December 2025. The announcement post refers to the same window as 1M in its API section, which is a rounded figure rather than a second measurement, so this page uses 1,050,000. Three other numbers get mistaken for these and are all wrong here. Codex runs this model with a 400,000-token window, which is a product setting for Codex rather than the API’s specification. chat-latest, which is a different model ID that happens to share the $5 and $30 pricing shape, declares 400,000 tokens of context and 272,000 of maximum input. And GPT-5.5 Pro is a separate model ID with its own documented limits. None of the three describes the model this page runs.
How much does the API cost, and is there a dated offer?
Two standard tiers, and no dated offer. At or under 272,000 input tokens the official pages list $5.00 per million input tokens, $0.50 per million cached input tokens and $30.00 per million output tokens; above 272,000 the same three figures become $10.00, $1.00 and $45.00. The rule is that prompts over 272K are priced at 2x input and 1.5x output for the full session, so crossing the line reprices the entire session rather than the excess. Batch and Flex are half the standard rate, Priority is 2.5 times it, and data residency adds 10%, all of which are standing tiers. The distinction that matters is that no page checked publishes a limited-time or dated promotion for this model, so there is nothing here to re-check after an expiry date — unlike the previous two pages in this series, which each carried one. This site bills points rather than dollars, at 1,220 points per million input tokens and 7,200 per million output tokens.
Which reasoning effort levels are documented, and which is the default?
The default is medium, and the two official pages do not agree on the full list. The model page documents five levels — none, low, medium (default), high and xhigh — while the table in the reasoning guide carries a sixth, max. The prompting guide states that medium is the balanced starting point for quality, reliability, latency and cost, and states equally clearly that higher effort is not automatically better output, naming conflicting instructions, thin stopping conditions and an open tool surface as the cases where a higher level can degrade the answer. Reasoning tokens are never shown to the caller but still occupy the context window and are billed as output tokens. This chat box has no level selector at all, so the provider’s default is what runs here.
Are images kept at full resolution, and what does that change?
They are, by default, and that is new in this generation. Where image_detail is absent or set to auto, the model page says images are treated as original rather than downscaled, up to 10,240,000 pixels and a 6,000-pixel edge. Asking for high lowers the ceiling to 2,500,000 pixels and a 2,048-pixel edge, and low downsamples harder above a 512-pixel edge. The documented input modalities are text and image, so image input is an official capability as well as an accepted attachment here. Two practical consequences follow. A screenshot-heavy prompt can now cost more than it did on the previous generation unless the detail level is set explicitly, and because the detail level has no control on this page, that choice is not available in this composer. For documents, note that a PDF attached here is converted to extracted text before sending, so a scan or a complex layout is better attached as an image or pasted as text.
How is GPT-5.5 different from GPT-5.5 Pro, GPT-5.5 Instant and chat-latest?
They are four different things and only one of them is this page’s model. gpt-5.5-pro is a separate API model ID priced at $30 per million input and $180 per million output with no cached-input discount, and it does not support Chat Completions, so it cannot be called the way this page calls its model. GPT-5.5 Instant is a ChatGPT consumer product announced separately in May 2026, with its own published figures; those figures describe the consumer product and not the API model. chat-latest is another API model ID that shares the $5 and $30 pricing shape with gpt-5.5 but declares a different context window. And gpt-5.5-2026-04-23 is a snapshot of this same model rather than a different one. Nothing from the first three is used anywhere on this page.
What does this page actually expose, compared with the official specification?
Less than the official pages document, and the difference is listed rather than glossed over. Available here: text chat, streaming replies, image and PDF attachments up to 24 MB each, and a web-search toggle billed at one point per search. Not available here: any context-window figure in the interface, because the catalog entry publishes none; any reasoning-effort selector or off switch; the eleven tools the official Responses API supports, including code interpreter, computer use and MCP; any sampling or verbosity control; and any endpoint selector. Two ceilings are set by this site rather than by the model: replies are requested with at most 4,096 output tokens, and the catalog entry declares 16,384, so the smaller number governs. Pricing is in points, at 1,220 points per million input tokens and 7,200 per million output tokens, which is not a conversion of the official dollar rates.
Sources and model information
Information checked
Start a GPT-5.5 conversation
The chat box above runs this model, and the guide on this page is written to be checked rather than trusted: the specifications come from the official model page, the comparison comes from the single official table that prints both columns, and the places where official pages disagree with each other or with this workspace are named rather than smoothed over. Send the task the way you would brief a person, state the constraint you want preserved as well as the outcome, and re-check any price against the pages linked below before relying on it — none of the figures on this page carry an expiry date, so nothing here needs re-checking on a particular day.
Start a chat with GPT-5.5 ↑