GPT-5.5 is OpenAI’s model for complex professional work, agentic coding, online research, and computer-use workflows. It remains technically substantial, but its practical decision is now a lifecycle question: use it as a pinned compatibility baseline only if your route still exposes it and you have a migration plan.
The dated anchor is unusually important. The GPT-5.5 API model page, checked on September 20, 2026, still lists the gpt-5.5-2026-04-23 snapshot, 1,050,000-token context, and $5 input / $30 output per million tokens. But ChatGPT’s official September 16 post says GPT-5.5 will leave ChatGPT, ChatGPT Work, and Codex on October 14, 2026. A live API card and a safe long-term default are not the same thing.
Key takeaways
GPT-5.5 targets complex coding, online research, data analysis, documents, spreadsheets, and software operation through tools.
The API card lists
gpt-5.5, a 1,050,000-token context, 128,000 maximum output, text/image input, and reasoning effort from none through xhigh.Standard API pricing is $5/M input, $0.50/M cached input, and $30/M output. Above 272K input, the full request uses 2x input and 1.5x output pricing.
OpenAI reports 82.7% on Terminal-Bench 2.0, 84.9% on GDPval, 78.7% on OSWorld-Verified, and 84.4% on BrowseComp. Treat each as a vendor row with its own harness.
The official product notice says ChatGPT, ChatGPT Work, and Codex will say goodbye to GPT-5.5 on October 14. API and third-party provider retirement must be checked separately.
GPT-5.5 at a glance
The GPT-5.5 model resources, prompt collection, and review collection keep the model-specific source trail. They do not grant API credits, ChatGPT access, Codex quota, or browser availability.
| Question | Current answer | Boundary |
|---|---|---|
| Model role | Flagship model for the most complex professional work | Positioning is not a success guarantee. |
| Model ID / snapshot | gpt-5.5 / gpt-5.5-2026-04-23 | Pin a snapshot only where your provider supports it. |
| Context / output | 1,050,000 / 128,000 tokens | API ceilings; clients and agents can compact earlier. |
| Input / output | Text and image input; text output | Audio and video are not listed as supported. |
| Reasoning | none, low, medium (default), high, xhigh | A UI or provider may expose fewer choices. |
| API rates | $5 input, $0.50 cached input, $30 output per million | Tool fees and regional processing are separate. |
| Long-input rule | Above 272K input: 2x input and 1.5x output for the full request | API billing rule, not a subscription promise. |
| Lifecycle | ChatGPT/Work/Codex retirement announced for 2026-10-14 | API, Azure, and third-party routes remain separate checks. |
What changed, and what the numbers do not prove
OpenAI’s release describes GPT-5.5 as a model that can define a goal, use tools, inspect its work, and continue across software. Its published rows show 82.7% on Terminal-Bench 2.0, 73.1% on an Expert-SWE internal evaluation, 84.9% on GDPval, 78.7% on OSWorld-Verified, 55.6% on Toolathlon, 84.4% on BrowseComp, 51.7% on FrontierMath T1–3, and 81.8% on CyberGym. The release also says GPT-5.5 generally keeps GPT-5.4-like token latency while using fewer tokens on comparable Codex tasks; that is an OpenAI claim, not a bill for your workload.
Vellum’s comparison is useful precisely because it keeps the rows separate. GPT-5.5 leads its cited Terminal, GDPval, OSWorld, BrowseComp and long-context MRCR rows, while Claude Opus 4.7 leads SWE-Pro, MCP Atlas and no-tool HLE rows. Neither source supplies a single prompt, tool harness, seed, or repeated run that settles every task. Compare accepted results and review time on your own fixed sample.
For the predecessor’s role, read the GPT-5.4 overview. For the current family, the GPT-5.6 Sol overview and GPT-5.6 Terra overview answer current routing questions; the GPT-6 Astra overview covers the newer high-end path. GPT-5.5 should not be presented as the default future route while an official retirement notice is active.
Access and pricing without mixing products
API: use
gpt-5.5only if the organization and provider still return it. Log snapshot, effort, tools, input/output tokens, retries, and accepted-task outcome.ChatGPT and ChatGPT Work: the official notice says these surfaces will remove GPT-5.5 on October 14, 2026. Plan, region, workspace, and app rollout can change the exact date or fallback behavior.
Codex: the same notice recommends GPT-5.6 Sol or GPT-6 Astra. Before switching, compare a pinned repository task with the same tests and tool permissions.
API pricing: $5/$30 is not a ChatGPT or Codex subscription conversion. Above 272K input, the full API request receives the higher multiplier; tool-specific fees and regional processing can add more.
Tabbit: a browser subscription and model-picker availability are separate from OpenAI API rates. This article has no authenticated GPT-5.5 run.
Choose a scenario and run one migration check
| Situation | Start with | Check before committing |
|---|---|---|
| Existing API integration pinned to 5.5 | Re-run the current fixed regression sample | Snapshot, endpoint, rate card, tool calls, tests and migration date |
| Codex repository with high-cost failures | Compare 5.5 with GPT-5.6 Sol and GPT-6 Astra | Same repo, effort, tools, tests, retries, elapsed time and review edits |
| Long research or spreadsheet workflow | Use a reversible sample with explicit evidence requirements | Effective context, source quality, compaction, output tokens and human review |
| High-volume routine work | Trial a current lower-cost family route | Cost per accepted result, not token price alone |
| Browser-first work | Confirm the live picker, then start with a public reversible page | Account access, action confirmation, screenshot/logging, and fallback |
The practical next step is a migration worksheet: freeze the current model ID and effort, select 10–20 representative tasks, record accepted output, retries, latency, input/output tokens and reviewer edits, then repeat with GPT-5.6 Sol or GPT-6 Astra. Keep the AI browser guide and Tabbit Browser overview separate from model entitlement. The Tabbit pricing page is not an OpenAI API invoice.
What users and reviewers actually establish
The media evidence is directional, not a controlled benchmark. Developers Digest’s search result exposes chapters on token efficiency, real-world benchmarks, spreadsheets/browser use, and pricing. Every’s three-week review exposes chapters on real-world builds, where 5.5 falls short, language strengths, and a daily-driver verdict. Those labels justify checking coding and browser workflows; they do not provide reproducible pass rates.
X adds a harder operational fact: the official retirement notice makes migration part of the product decision. A separate community short review frames GPT-5.5 against GPT-5.6 Luna as an intelligence-versus-value comparison, but gives no task set, bill, effort, or repeated run. Reddit was searched during research, but the current results were noisy and unrelated, so no Reddit claim is used as evidence.
Tabbit’s role and limits
If a task begins with pages, tabs, files, screenshots, or a browser-only workflow, Tabbit Browser can provide a separate browser-first runtime. It cannot extend an OpenAI model’s context window, preserve a retiring product route, or prove that GPT-5.5 appears in your picker. This draft did not run an authenticated GPT-5.5 task, so verify the live label and effective limits with a small reversible workflow.
Tabbit BrowserThe CTA downloads Tabbit Browser; it is not a GPT-5.5 entitlement. Keep browser costs and account limits separate from the API card.
Unknown risks
Lifecycle scope: the October 14 notice names ChatGPT, ChatGPT Work, and Codex. Do not assume it also retires API, Azure, or every provider route without a fresh notice.
Effective context: 1.05M is an API ceiling. Compaction, retrieval quality, distraction, and client limits can dominate long-context work.
Benchmark transfer: vendor and secondary benchmark rows use different tasks, tools, effort, and dates. They do not predict your acceptance rate.
Migration drift: model aliases, fallbacks, quotas, price cards, and tool permissions can change while a project is being migrated.
Security and review: stronger coding or computer-use scores do not remove tests, permission boundaries, data handling controls, or human approval for irreversible actions.
Verdict
GPT-5.5 is still a serious professional-work and agentic-coding model, and its API card remains concrete: 1.05M context, 128K output, $5/$30 standard token rates, and a pinned April 23 snapshot. But the official October 14 ChatGPT/Work/Codex retirement announcement changes the recommendation. New projects should start by evaluating GPT-5.6 Sol or GPT-6 Astra; existing 5.5 projects should pin what they have, measure accepted work, and rehearse the migration before the deadline.
The model page answers what GPT-5.5 can expose. Your fixed regression sample answers whether moving is safe. Treat those as separate decisions.
FAQ
What is GPT-5.5?
GPT-5.5 is OpenAI’s model for complex professional work, agentic coding, online research, data analysis, documents, spreadsheets, and computer use. The API model ID is gpt-5.5 and the listed snapshot is gpt-5.5-2026-04-23.
What are GPT-5.5’s context and output limits?
The checked API page lists a 1,050,000-token context window and 128,000 maximum output tokens. Clients, plans, providers, and agent harnesses can expose less, and a large API ceiling is not a guarantee of useful long-context reasoning.
How much does GPT-5.5 cost?
The API page lists $5 per million input tokens, $0.50 per million cached input tokens, and $30 per million output tokens. Above 272K input tokens, the full request uses 2x input and 1.5x output pricing; tools and regional processing can add costs.
What did GPT-5.5 change from GPT-5.4?
OpenAI reports stronger agentic coding, computer use, research, and professional-work results, including 82.7% on Terminal-Bench 2.0 and 78.7% on OSWorld-Verified. These are vendor-published rows, not a universal ranking.
Can I still use GPT-5.5?
The API model page was live when checked, but an official ChatGPT post dated September 16, 2026 says GPT-5.5 will leave ChatGPT, ChatGPT Work, and Codex on October 14 and recommends GPT-5.6 Sol or GPT-6 Astra. Confirm whether your API or provider route is covered before migrating.
Can I use GPT-5.5 in Tabbit Browser?
This draft did not run an authenticated Tabbit GPT-5.5 task and makes no claim about picker visibility, effective context, latency, or availability. Check the live picker and test a reversible task; Tabbit subscription cost is separate from OpenAI API pricing.