What is GPT 5.6 Sol?
GPT-5.6 Sol is the flagship of OpenAI’s GPT-5.6 family, and the official model page describes it in exactly those words as a flagship model for complex professional work. OpenAI writes the name GPT-5.6 Sol, while this site’s own label for it — the one the heading and the navigation show — drops the hyphen and reads GPT 5.6 Sol; both appear here, and the model code settles which is meant. Its API model code and its default snapshot are both gpt-5.6-sol, the family alias gpt-5.6 routes to it, the documented context window is 1,050,000 tokens, the documented maximum input is 922,000 tokens, the documented maximum output is 128,000 tokens, the documented inputs are text and image, the documented output is text, and the knowledge cutoff is 16 February 2026. The naming scheme is new with this generation: the number identifies the generation, while Sol, Terra and Luna are durable capability tiers that can advance on their own cadence, and OpenAI says Sol roughly corresponds to the unsuffixed tier used in earlier GPT-5 families. Three statements from the official pages matter more than the rest. The first is that this model has been succeeded but not deprecated, and the two official pages that would tell you are both reported here rather than one of them being picked. OpenAI’s model catalog now features GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, and the word flagship sits on GPT-6 Astra, described as the flagship model for complex reasoning and coding; the deprecations page, meanwhile, lists gpt-5.6-sol in neither its upcoming nor its past table, and instead prints it repeatedly in the replacement column for older models being retired. Succeeded and deprecated are different states. The second is that the price moved after launch. The family page’s own availability section still prints the launch figures, $5 per million input tokens and $30 per million output tokens, while the model page and the pricing page both now read $4, $0.40 cached and $20, and the model page calls that a 20% reduction in input pricing and a 33% reduction in output pricing available at least through 21 November 2026. The same family page carries the update that dates the cut to 21 August 2026 for the following three months. The third is that reasoning changed shape rather than simply getting longer. This model documents six effort levels on its own model page — none, low, medium, high, xhigh and max, with medium as the default — where the previous generation’s model page listed five and left max to a separate guide. Pro is a mode here rather than a separate model slug, persisted reasoning defaults to keeping earlier turns, and reasoning tokens are never shown to you while still occupying the context window and still being billed as output tokens.
What GPT-5.6 Sol is worth using for
Six effort levels, and this box has no dial for any of them
The model page for this generation prints the full list itself: none, low, medium, high, xhigh and max, with medium as the default when you omit the setting. That is a change from the generation before it, whose model page listed five and pushed the sixth into the reasoning guide. Two of the additions are worth naming separately. Max is described as giving the model more time than xhigh to reason, explore alternatives, run checks and revise its approach, and the official advice is to compare max against xhigh rather than assume one wins. Pro is not a model: it is a second execution mode in the Responses API that applies more model work before returning a single answer, it works with any GPT-5.6 model, its supported efforts begin at medium, and the tokens it aggregates are billed at the selected model’s standard rates. Three billing facts travel with all of this. Reasoning tokens are never visible to the caller, they still occupy space in the context window, and they are billed as output tokens. Because of the third, a level that looks free at the top of a conversation is priced into everything that follows it. This chat box has no level selector, no mode selector and no off switch, because the catalog entry behind the page marks reasoning as unsupported, so whatever the provider defaults to is what runs here.
A model documents six reasoning levels with medium as the default, offers a separate pro execution mode whose efforts begin at medium, and bills its reasoning tokens as output tokens without ever showing them. Give me a short rule for choosing between the six levels per task shape, tell me when pro is worth its extra model work, and explain how the hidden reasoning tokens change what a long conversation costs.
The documented migration hazards are not about speed
The official upgrade guide lists the ways this move can quietly break an integration, and most of them are behavioural rather than performance-related. The first: previous generations commonly defaulted to no reasoning at all, while this one defaults to medium when the setting is omitted, so a caller who never set the field can suddenly be slower, more expensive and incompatible with Chat Completions function tools. The second: Chat Completions with function tools now needs the effective reasoning set to none explicitly, which is a request-shape change rather than a prompt change. The third: prompt caching became more explicit, cache writes are billed at 1.25 times the uncached input rate, and a stable prefix followed by a changing suffix can lose the hits the caller thought they had. The guide’s own instruction is blunt about the shape of the work — do not perform a blind model-string replacement, preserve the behaviour and the contracts first, then make the smallest safe migration, and do not adopt reasoning fields, new schemas or changed tests just because a model upgrade happened. One list is easy to forget because it is invisible in the request: model pickers, registries, allowlists, pricing metadata, capability flags and tests all have to move together with the model string.
I am moving a production integration to a new model generation whose default reasoning effort changed, whose Chat Completions path now needs reasoning explicitly disabled for function tools, and whose cache writes are newly billed. Give me an inventory checklist and the smallest safe sequence of changes, with the compatibility gates I should run before and after.
Images now keep their original size unless you say otherwise
This generation changed the default for image input, and the old habit now costs more than it used to. The official guidance says that with the detail level omitted or set to auto, image dimensions are preserved, except that an image larger than 65,535 pixels on either side is scaled down to fit that limit, and that the API rejects images still exceeding the 30,000-patch limit rather than resizing them to comply. The practical reading is that the cheap setting is no longer the default one, so a screenshot-heavy prompt consumes image tokens at a different rate than the same prompt did a generation earlier unless the caller sets the level deliberately, and the guidance suggests choosing original detail only where the extra input cost and latency are justified, such as dense, coordinate-sensitive, OCR or localisation work. Files changed in the same direction: in the Responses API, an omitted detail level or auto can use high page-image detail, while Chat Completions file inputs do not expose the same control. This workspace accepts JPEG, PNG and WebP images alongside PDFs at up to 24 MB each, so the attachment path exists here, but a PDF is converted to extracted text before it is sent, which is a different mechanism from sending a page image, and neither the detail level nor the file handling has a control on this page.
Images keep their original dimensions by default unless the detail level is set, while the previous generation downscaled automatically. Tell me how that changes the token cost of a screenshot-heavy prompt, which images I should downgrade deliberately, what accuracy I lose by doing so, and how a PDF attachment differs from an image attachment.
The official advice is to write less, not more
The prompting guidance for this model opens with a section titled simplify prompts first, and its own numbers are worth copying. In a sample of internal coding-agent evaluation runs, configurations with leaner system prompts improved evaluation scores by roughly 10 to 15 percent while reducing total tokens by 41 to 66 percent and cost by 33 to 67 percent; the documentation labels those ranges directional and asks callers to validate them on their own tasks. What to trim is listed as well: repeated statements of the same rule, style or process instructions that do not change behaviour, examples that do not change behaviour, and tools unrelated to the task. What to keep is the user-visible outcome, success criteria and stopping conditions, safety and permission constraints, tool-routing rules, and the required output shape. Two smaller instructions follow from the same idea. Answer length is now a documented parameter, text.verbosity, taking low, medium or high, and the guidance notes that this generation is more concise than the one before it, so a broad instruction to be brief may now be unnecessary or even harmful. And the guide warns that conflicting rules create more instability than missing detail on models that follow prompt contracts closely, which is the opposite of the usual instinct to add one more clarification. None of these controls exist in this composer, so the practical route here is to write the shorter prompt.
A model’s own guidance says leaner system prompts improved evaluation scores by roughly 10 to 15 percent while cutting tokens substantially, and that redundant instructions and unused tools should be removed. Take the prompt below and cut it to the shortest form that still states the outcome, the success criteria, the constraints and the stopping condition.
Official GPT-5.6 Sol specifications
- Developer
- OpenAI
- Model ID
- gpt-5.6-sol
- Official context window
- 1,050,000 tokens
- Official maximum output
- 128,000 tokens
- Official inputs
- Text and image
- Output
- Text
GPT-5.6 Sol and GPT-5.5 side by side
Every value below comes from a single page — the GPT-5.6 launch page on openai.com, checked on 30 September 2026 — and from a single table family within it. That page carries eleven HTML tables, one per evaluation group, and the columns for GPT-5.6 Sol and GPT-5.5 are present in all eleven. Thirty-nine rows report a value for both models, and all thirty-nine are reproduced here, in the order the official page prints them, so nothing is dropped for going the wrong way and nothing is selected for going the right way. The left column is this page’s model and the right column is GPT-5.5, the generation it succeeds. Two groups are labelled internal by OpenAI itself, and those rows are the provider’s own evaluations rather than public benchmarks: Management Consulting Tasks and MedChemBench. Thirty-eight of the thirty-nine rows move in this model’s favour, by margins that run from near-ties such as Toolathlon at 58 against 55.6 and GDPval-AA v2 at 1,747.8 against 1,493.7, up to GeneBench Pro at 28.7 against 12 and KernelGen 1P at 61.1 against 29.3. The thirty-ninth is the one worth reading twice: on the long-context recall test MRCR v2 eight-needle at 512K–1M, GPT-5.6 Sol scores 73.8 against GPT-5.5’s 74. That is the only row in the set that goes the other way, and it sits in the same group as the row where the advantage is largest, 91.5 against 81.5 at 256K–512K. One conflict inside that same official page is recorded rather than smoothed over: its prose says the model sets a new high of 53.6 on Agents’ Last Exam, while its own table prints 52.7%. This page uses the table value and reports the discrepancy. One further figure travels with the ExploitGym row, where the table prints the six-hour result at 33.7% and the prose adds that the two-hour figure rose from 15.1% to 24.9%. These are provider evaluations of the models, not measurements of this website, and no score here predicts the quality or speed of your conversation.
| Official evaluation | GPT 5.6 Sol | GPT-5.5 |
|---|---|---|
| Professional · Agents’ Last Exam | 52.7% | 46.9% |
| Professional · GDPval-AA v2 | 1,747.8 Elo | 1,493.7 Elo |
| Professional · Management Consulting Tasks (Internal) | 43.2% | 31.3% |
| Professional · Big Finance Bench | 53% | 49% |
| Professional · Artificial Analysis Intelligence Index v4.1 | 58.9 | 54.8 |
| Coding · Artificial Analysis Coding Agent Index v1.1 | 80 | 76.4 |
| Coding · SWE-Bench Pro | 64.6% | 59.4% |
| Coding · DeepSWE v1.1 | 72.7% | 67% |
| Coding · Terminal-Bench 2.1 | 88.8% | 85.6% |
| Science and health · GeneBench Pro | 28.7% | 12% |
| Science and health · LifeSciBench | 59.9% | 50.4% |
| Science and health · MedChemBench (Internal) | 48.3% | 35.5% |
| Science and health · HealthBench Professional | 60.5% | 49.5% |
| Computer use · OSWorld 2.0 | 62.6% | 47.5% |
| Computer use · BrowseComp | 90.4% | 84.4% |
| Computer use · BenchCAD | 70.6% | 44.4% |
| Computer use · BenchCAD (python tool) | 83.4% | 55.8% |
| Cybersecurity · Capture-the-Flag Challenges | 96.7% | 88.1% |
| Cybersecurity · SEC-Bench Pro | 71.2% | 45.8% |
| Cybersecurity · ExploitBench | 73.5% | 47.9% |
| Cybersecurity · ExploitGym | 33.7% | 15.1% |
| Self-improvement · Internal Research Debugging Evaluation | 68.3% | 50% |
| Self-improvement · KernelGen 1P | 61.1% | 29.3% |
| Self-improvement · NanoGPT | 9.69% | 2.65% |
| Self-improvement · PostTrainBench Lite | 50.3% | 38.8% |
| Self-improvement · RSI Index | 57.9% | 41.7% |
| Multimodal · MMMU Pro (no tools) | 83% | 81.2% |
| Multimodal · MMMU Pro (with tools) | 84.6% | 83.2% |
| Multimodal · gdp.pdf | 30.7% | 26% |
| Academic · GPQA Diamond | 94.6% | 93.6% |
| Academic · FrontierMath Tier 1-3 (v2) | 89% | 85.3% |
| Academic · FrontierMath Tier 4 (v2) | 83% | 72.5% |
| Tool use · AutomationBench | 18.1% | 12.9% |
| Tool use · Toolathlon | 58% | 55.6% |
| Long context · OpenAI MRCR v2 8-needle 256K-512K | 91.5% | 81.5% |
| Long context · OpenAI MRCR v2 8-needle 512K-1M | 73.8% | 74% |
| Long context · GraphWalks BFS 256k f1 | 90.7% | 73.7% |
| Long context · GraphWalks BFS 1mil f1 | 77.1% | 45.4% |
| Abstract reasoning · ARC-AGI-3 | 7.78% | 0.43% |
How to use GPT 5.6 Sol online
Choose the model
This page opens with GPT 5.6 Sol 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-5.6-sol, and the site’s own display label, which is what the navigation and the heading show, omits the hyphen. Note also that this is the flagship of the GPT-5.6 generation rather than the newest flagship OpenAI ships, and that both statements come from official pages linked at the bottom.
Write the outcome and the stopping condition
The official prompting guidance for this model asks for the destination rather than every step: state the outcome, the constraints, the evidence available and the bar for finishing, then let the model choose an efficient path. Three habits from that guidance are worth adopting before you start. State what must be preserved as well as what must change, because an instruction to improve something without naming the invariant produces exactly the literal instruction given. Name the stopping condition explicitly, so that minimising tool loops never outranks correctness. And cut the prompt rather than pad it, since repeated rules and unused detail are what the same guidance says to remove first. Any prompt in this guide can be sent as it stands or edited first.
Sign in, send and check the cost against the current window
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. Because this model’s official price is currently a promotional one with a stated end date, re-check the figures against the official pages linked below before relying on them, and treat any third-party summary that still quotes the launch price as out of date.
What GPT 5.6 Sol costs, and how long that price lasts
This model carries the first dated price in this series, and the date matters as much as the number. The official model page and pricing page, both checked on 30 September 2026, list two standard tiers for gpt-5.6-sol. At or under 272,000 input tokens the rate is $4.00 per million input tokens, $0.40 per million cached input tokens, $5.00 per million cache-write tokens and $20.00 per million output tokens; above that threshold the same four figures become $8.00, $0.80, $10.00 and $30.00. The stated rule is that prompts with more than 272K input tokens are priced at 2x input and 1.5x output for the full request, so crossing the line reprices the entire request rather than only the excess. Three details sit alongside those numbers. Cache writes are billed at 1.25 times the uncached input rate, which is new for this generation and is why cache-write tokens get their own column. Batch and Flex are half the standard rate and Fast is twice it, with Fast being the renamed form of what used to be called Priority. And regional processing for data residency adds 10% for models released on or after 5 March 2026, as do FedRAMP endpoints. The dated part is this: the model page states that GPT-5.6 Sol’s promotional pricing is available at least through 21 November 2026, and the launch page carries the update that dates the cut to 21 August 2026 and describes it as more than 20% off for the following three months. So that is the window — from 21 August 2026 to at least 21 November 2026 — and this page will need re-checking after it closes. The model page characterises the change as a 20% reduction in input pricing and a 33% reduction in output pricing. One more official inconsistency belongs here rather than being tidied away: the launch page’s own availability section still prints the original $5 and $30, which is the price before the August cut and not a second, simultaneous rate. This site charges none of those figures. It bills points, and the catalog reference price behind this page is 306 points per million input tokens, 1,836 points per million output tokens and 30.6 points per million cached input tokens, which is not a conversion of the official dollar rates.
View points and plans ↗GPT 5.6 Sol FAQ
What does GPT 5.6 Sol cost, and is the price temporary?
Both, and the second half is the part most summaries miss. The official pages list $4.00 per million input tokens, $0.40 per million cached input tokens and $20.00 per million output tokens at or under 272,000 input tokens, with a long-context tier above that at $8.00, $0.80 and $30.00 plus $10.00 for cache writes. The model page states that this promotional pricing is available at least through 21 November 2026, and the launch page dates the reduction to 21 August 2026 and describes it as more than 20% off for three months, so the window is 21 August 2026 to at least 21 November 2026. Batch and Flex take half the standard rate, Fast takes twice, and data residency adds 10%. Nothing here is this site’s checkout price: this workspace bills points, at 306 points per million input tokens and 1,836 per million output tokens.
How does GPT 5.6 Sol compare with GPT-5.5 in OpenAI’s own tables?
Across thirty-nine rows that report both models, thirty-eight favour GPT-5.6 Sol. The margins range widely. GeneBench Pro goes from 12% to 28.7%, KernelGen 1P from 29.3% to 61.1%, SEC-Bench Pro from 45.8% to 71.2%, ExploitBench from 47.9% to 73.5% and Terminal-Bench 2.1 from 85.6% to 88.8%. The narrowest wins are Toolathlon at 58% against 55.6% and GDPval-AA v2 at 1,747.8 against 1,493.7. The one row that goes the other way is worth stating plainly: on MRCR v2 eight-needle at 512K–1M, this model scores 73.8% against GPT-5.5’s 74%. All of these are provider evaluations of the models, not measurements of this website.
Is GPT 5.6 Sol still the flagship, and is it deprecated?
Two different answers, both official. It has been succeeded: OpenAI’s model catalog now features GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, and describes GPT-6 Astra as the flagship model for complex reasoning and coding, while the GPT-5.6 launch page had described Sol as the new flagship at the time it shipped. It has not been deprecated: the deprecations page lists gpt-5.6-sol in neither the upcoming nor the past table, and instead prints it in the replacement column for a long list of older models being retired, including snapshots of GPT-5, o1, o3 and several GPT-4 variants. Succeeded and deprecated are different states, and this page reports both official pages rather than choosing one.
What are the context window and the maximum output?
The model page documents a 1,050,000-token context window, a maximum input of 922,000 tokens and a maximum output of 128,000 tokens, with the default snapshot gpt-5.6-sol and a knowledge cutoff of 16 February 2026. Two of those three numbers are easy to confuse. The maximum input is smaller than the whole context window, which is usual, but the figure is printed on the model page rather than inferred. And this page is not running against that output ceiling: replies here are requested with at most 4,096 output tokens, the smaller of the catalog’s declared 16,384 and the composer’s hard ceiling, so plan long answers as a sequence rather than one long reply.
Which reasoning levels exist, and which one runs here?
Six exist and none of them is selectable on this page. The model page lists none, low, medium, high, xhigh and max, with medium as the default when the setting is omitted, which is one more than the previous generation’s model page printed. On top of the levels there is a separate execution mode, pro, available in the Responses API and working with any GPT-5.6 model, whose supported efforts begin at medium and whose aggregated tokens are billed at the model’s standard rates. Persisted reasoning defaults to carrying earlier turns forward for this generation. Reasoning tokens are never shown to the caller but still occupy the context window and are still billed as output tokens. This chat box has no level selector and no off switch, because the catalog entry behind the page marks reasoning as unsupported, so the provider’s default applies.
Do images and PDFs work, and does the detail level change the cost?
Images work and the detail level does change the cost. With the detail level omitted or set to auto, the official guidance says image dimensions are preserved, except that an image larger than 65,535 pixels on either side is scaled down to fit, and images exceeding the 30,000-patch limit are rejected rather than resized. That makes the cheap setting no longer the default, so a screenshot-heavy prompt can cost more than it did on the previous generation unless the level is set deliberately, and there is no detail control on this page. Attachments here accept JPEG, PNG and WebP files and PDFs at up to 24 MB each. A PDF, however, is converted to extracted text before it is sent rather than sent as a page image, so a scanned document or a complex layout may lose content and is better attached as an image or pasted as text.
What does this page expose compared with the official API?
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 per search. Not available here: any context-window figure in the interface, because the catalog entry publishes none; any reasoning-effort selector, off switch or pro mode; the ten tools the official Responses API supports, including code interpreter, computer use, MCP, apply patch, skills and hosted shell; any image detail control; any 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 and is not a conversion of the official dollar rates.
Is NanoBanana Pro the official OpenAI website?
No. NanoBanana Pro is an independent platform providing a shared workspace for multiple AI models, and this page is a guide to one of them rather than a product page from its developer. OpenAI develops and publishes GPT-5.6 Sol; the official model page, the launch page, the pricing page, the two migration and prompting guides, the reasoning guide and the deprecation list are all linked at the bottom of this page, and every specification, price and evaluation figure quoted here comes from one of them. If you need the developer’s own contract, rate limits or support, follow those links rather than relying on this page.
Sources and model information
Information checked
- OpenAI API documentation · GPT-5.6 Sol ↗
- OpenAI · GPT-5.6: frontier intelligence that scales with your ambition ↗
- OpenAI API documentation · Pricing ↗
- OpenAI API documentation · Using GPT-5.6 ↗
- OpenAI API documentation · Upgrading to GPT-5.6 Sol ↗
- OpenAI API documentation · Prompting guidance for GPT-5.6 Sol ↗
- OpenAI API documentation · Reasoning models ↗
- OpenAI API documentation · Model catalog ↗
- OpenAI API documentation · Deprecations ↗
Start a GPT 5.6 Sol 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 page whose eleven tables print 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 — outcome, constraints, evidence, stopping condition — and re-check the price against the pages linked below, because this is the one model in the series whose official price carries an end date.
Start a chat with GPT 5.6 Sol ↑