On this page · 10 sections
Summary. OpenAI's deprecations page, checked on 20 November 2026, lists gpt-5.5-cyber-preview as deprecated with a notice date of 5 November 2026 and a shutdown date of 1 February 2027, 73 days from today. The replacement column names no model: it tells callers to use the most capable cyber model available to them. OpenAI's Daybreak guide is more specific. Requests to the preview will fail after retirement, and teams should change the model field to gpt-5.6-cyber, or to gpt-6-cyber if their organisation is approved for it. GPT-6 Cyber, released on 5 November 2026, costs $20 per million input tokens and $100 per million output tokens for requests up to 272,000 input tokens, with a 1,050,000-token context window. gpt-5.6-cyber costs $12.50 and $75. Neither model is open to everyone. Both sit in Daybreak Red, and GPT-6 Cyber also needs a separate model approval, project enablement and an IP allowlist before the first request. Our verdict: switch to gpt-5.6-cyber in December, apply for GPT-6 Cyber in parallel, and move only once approval and setup are proven on a test project.
About this article. This is for security engineering leads, application security teams and SOC managers who call OpenAI's cyber models from code scanning, triage or validation pipelines. We checked every model ID, date, price and access rule on 20 November 2026 against OpenAI's deprecations page, model pages, pricing page, API changelog, Daybreak and cybersecurity safeguard guides, Trusted Access documentation, IP allowlist and Private Safety Processing guides, plus an archived copy of OpenAI's May 2026 GPT-5.5-Cyber announcement. Rupee figures use European Central Bank reference rates. Vendors change prices and access terms, so re-check before you commit; this is not legal advice, and no vendor paid for inclusion. Researched with AI assistance and checked against the primary sources listed in the references.
What OpenAI announced and when
OpenAI introduced GPT-5.5-Cyber on 7 May 2026 as a limited preview for defenders who secure critical infrastructure, describing it as its most permissive option for authorised red teaming, penetration testing and controlled validation. The same post said individual Trusted Access members using the most permissive models would need Advanced Account Security from 1 June 2026, while organisations could instead attest to phishing-resistant authentication in their single sign-on.
On 7 August 2026 OpenAI split its security programme into two tiers, Daybreak Blue and Daybreak Red, and released gpt-5.6-cyber as a purpose-trained Red model. On 5 November 2026 it added GPT-6 Cyber, and the deprecations page gives the same date as the notice date for the preview's shutdown.
The timeline below comes from OpenAI's announcement, changelog and deprecations page.
| Date | Event | Source type |
|---|---|---|
| 7 May 2026 | GPT-5.5-Cyber limited preview announced | OpenAI announcement |
| 1 June 2026 | Advanced Account Security required for individual members on permissive models | OpenAI announcement |
| 7 Aug 2026 | Daybreak Blue and Red tiers, gpt-5.6-cyber released | API changelog |
| 5 Nov 2026 | GPT-6 Cyber released; gpt-5.5-cyber-preview deprecated | Changelog and deprecations page |
| 1 Feb 2027 | gpt-5.5-cyber-preview shuts down | Deprecations page |
Why the notice is only 88 days
From 5 November 2026 to 1 February 2027 is 88 days. OpenAI's policy promises at least 6 months for generally available models and 3 months for specialised variants, but says preview models, identified by preview in the name, may be retired on much shorter notice, giving 2 weeks as an example. The policy also advises against preview models for business-critical production work unless you can migrate quickly.
The previous cyber model shows how short that can be. gpt-5.4-cyber received notice on 11 September 2026 and shut down on 1 October 2026, 20 days later. Teams on the 5.5 preview have more time than that, but not much more than a quarter.
What shutdown means in practice
OpenAI defines a shut-down model as one that no longer accepts requests through the standard API. The Daybreak guide says requests naming gpt-5.5-cyber-preview will fail after retirement, and that only organisations with existing access can use it now. There is no redirect. A scanner that names the preview in its configuration will start returning errors on 1 February.
Which model replaces it
The answer depends on what your organisation is approved for, not on which model looks newest. Prices and context windows below are from each model page and OpenAI's pricing page on 20 November 2026; tier requirements are from the Daybreak guide.
| Model | Daybreak tier needed | Input per 1M | Output per 1M | Context window |
|---|---|---|---|---|
| gpt-5.5-cyber-preview | Red, existing access only | Not published | Not published | Not published |
| gpt-5.6-cyber | Red | $12.50 | $75.00 | 400,000 |
| gpt-6-cyber | Red plus model approval | $20.00 | $100.00 | 1,050,000 |
| gpt-6-sol | Blue or Red | $2.00 | $10.00 | 1,050,000 |
We could not open a model page or price for gpt-5.5-cyber-preview: its documentation URL now returns not found.
gpt-5.6-cyber is the default path. OpenAI's model page says it is available to all Daybreak Red customers once Red is enabled for the project sending requests. Any organisation already using the 5.5 preview holds Red access, so for most teams this is a model string change plus a test run.
GPT-6 Cyber needs a second approval. The model page says an organisation needs Daybreak Red access and approval to use it, and the Daybreak guide lists it as Red plus additional model approval. OpenAI positions it for investigating a vulnerability, reproducing it in an authorised test environment and verifying a fix. OpenAI classifies GPT-6 Cyber as Critical for cybersecurity under its Preparedness Framework, against High for gpt-5.6-cyber, and applies extra automated safeguards to both.
Many teams may not need a Red model at all. OpenAI's Trusted Access documentation recommends starting with Daybreak Blue for most authorised defensive work: vulnerability discovery and triage, secure code review, threat modelling, detection engineering, incident response, malware analysis and patch validation. Red is for controlled vulnerability reproduction, exploit validation, penetration testing and red teaming. If you used the preview for scanning and triage, a Blue model such as gpt-6-sol may cover it at a tenth of gpt-6-cyber's token rates.
One catch: the newer mainline models (gpt-6.1-sol, gpt-6.2-sol, gpt-6.2-astra and others) only get reduced refusals when the organisation holds Red approval, even though they take the Blue value. With Blue approval alone, gpt-6-sol and gpt-6-luna are the eligible choices.
The old aliases are also on the way out. The gpt-daybreak-red-latest and gpt-daybreak-blue-latest model pages are marked deprecated and point callers to the latest cyber model available to them. They have no shutdown date in the deprecations table, but code that relies on them should move to explicit model IDs now. Our earlier analysis of the Daybreak Red and Blue tiers covers how those aliases were priced and where they lacked regional processing.
How Trusted Access approval works
Trusted Access for Cyber is a reviewed access programme, not a model. Approval is specific to a person or service, a workspace or API organisation and project, a model and a product surface. Four rules matter for planning.
- Blue does not lead to Red automatically. Applying, verifying an identity or being approved for Daybreak Blue does not grant Red access, which needs separate approval and provisioning.
- Individuals and organisations apply differently. Individuals use the individual Trusted Access application; organisations submit the enterprise request form and work through their OpenAI representative. An application does not guarantee approval.
- Enterprise access is internal only. OpenAI says enterprise approval covers the organisation's own authorised work and must not be extended to external users, third-party customers or downstream product features. A managed security provider serving clients through its own product should raise this with its account team first.
- Zero Data Retention is separate. Trusted Access does not grant it automatically; approved retention controls have to be confirmed for the exact organisation and endpoint.
GPT-6 Cyber approval has no published turnaround time, so treat it as unknown. Start the request now and plan the February cut-over on gpt-5.6-cyber, so the deadline does not depend on a decision you cannot schedule.
Setting up GPT-6 Cyber in the API
Two requirements apply to every Daybreak Red organisation before any model choice. Red organisations must require single sign-on for the API organisation that holds cyber access and enforce multi-factor authentication through the identity provider, and Daybreak can only be enabled in a named project, because default projects cannot enable it. The GPT-6 Cyber steps then follow in order.
- Confirm approval. Your account team confirms Daybreak Red access and GPT-6 Cyber approval for the organisation.
- Enable the project. An organisation owner enables Daybreak Red for the project that will send requests, and you use a key from that project.
- Set an IP allowlist. GPT-6 Cyber through the API requires an allowlist of the public IP addresses your application sends from.
- Configure retention if you use ZDR. Zero Data Retention projects must set up Private Safety Processing, show storage as validated and use the retention policy for ZDR with Private Safety Processing.
The allowlist deserves care. It holds up to 50 addresses or CIDR ranges, a project-level list replaces the organisation list rather than adding to it, changes take up to 15 minutes, and a blocked request returns HTTP 401 with the error code ip_not_authorized. List every egress path first, including CI runners and disaster recovery, or a scanner that worked from a laptop will fail from the pipeline. Private Safety Processing writes encrypted records of selected interactions to customer-controlled cloud storage with a 30-day expiry, for automated safety review only.
A minimal request looks like this. The access_programs field selects the tier, and safety_identifier should be a stable, pseudonymous ID for the person making the request.
curl https://api.openai.com/v1/responses \
-H "Authorization: Bearer $OPENAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-6-cyber",
"access_programs": {"cyber": "daybreak_red"},
"safety_identifier": "analyst_7f3a",
"input": "Propose a regression test for the path traversal fix in this diff."
}'
For gpt-5.6-cyber, the API selects daybreak_red when the parameter is omitted and fails the request if the project lacks access. Setting it explicitly makes the intent visible in code review. Each model accepts only its listed value: gpt-6-sol, for example, accepts daybreak_blue and rejects daybreak_red with invalid_access_program.
Errors to handle before go-live
The first three codes below are from the Daybreak guide, the allowlist error from the IP allowlist guide and the policy error from the cybersecurity checks guide.
| HTTP status | Error code | What it means |
|---|---|---|
| 400 | invalid_access_program | The model needs a different access_programs value |
| 400 | unsupported_access_program | The model does not support Daybreak |
| 403 | access_program_not_enabled | The key's project or organisation lacks the program |
| 401 | ip_not_authorized | The request came from outside the allowlist |
| Not stated | cyber_policy | Safeguards limited access after suspicious activity |
Safeguards that can stop a pipeline
OpenAI monitors traffic to its High and Critical cyber models for signs of suspicious activity. If thresholds are crossed, access can be temporarily limited while it is reviewed, and requests return the cyber_policy error. OpenAI acknowledges that legitimate security research may sometimes be flagged while these systems are calibrated.
The scope of a block depends on you. Without a per-user safety_identifier, access may be revoked for the whole organisation; with one, it may be limited to the affected user after human review and warnings. For ZDR organisations, request-level checks also apply, and on streaming requests the error can arrive midway through the stream. OpenAI points to support for appeals if access needs restoring before a 7-day period ends.
Three engineering consequences follow:
- Send a safety_identifier on every request. A shared service account without one puts every analyst's access at risk from one flagged job.
- Treat cyber_policy as a handled state. Queue the job, alert a person and stop retrying, rather than looping on a blocked key.
- Parse streams defensively. A triage UI that streams findings must cope with an error event after partial output.
Trusted Access also does not configure your environment. OpenAI's recommended configuration, aimed mainly at Red and other higher-risk work, asks teams to run in isolated labs without unrestricted internet or production access, to document approved systems, actions and time limits, and to enforce them with independent filesystem, network and identity controls rather than documentation alone. Our guide to AI agent security guardrails covers the same review-before-execute pattern for tool calls.
Cost: what the move does to the bill
There is no published token price for gpt-5.5-cyber-preview, so the comparison uses the three models you can move to. The workload is a monthly triage pipeline sending 50 million input tokens and receiving 5 million output tokens, all under 272,000 input tokens per request, with no caching. Rupees use ECB reference rates of 19 November 2026 (US$1.0992 and ₹102.55 per euro), crossed to ₹93.30 per US dollar and rounded to two decimals.
| Model | Input cost | Output cost | Monthly total, USD | Monthly total, INR |
|---|---|---|---|---|
| gpt-6-cyber | $1,000 | $500 | $1,500 | ₹1,39,950 |
| gpt-5.6-cyber | $625 | $375 | $1,000 | ₹93,300 |
| gpt-6-sol (Blue) | $100 | $50 | $150 | ₹13,995 |
The workings: 50 × $20 = $1,000 and 5 × $100 = $500 for GPT-6 Cyber; 50 × $12.50 = $625 and 5 × $75 = $375 for gpt-5.6-cyber; 50 × $2 = $100 and 5 × $10 = $50 for gpt-6-sol. Each dollar total is multiplied by 93.30.
Two pricing details change real bills.
- Long prompts cost more on GPT-6 Cyber. Above 272,000 input tokens per request, rates rise to $40 input, $4 cached input and $150 output per million. A single request with 300,000 input and 20,000 output tokens costs 0.3 × $40 + 0.02 × $150 = $15, or ₹1,399.50. Chunk or cache whole-repository prompts where quality allows.
- Caching cuts input cost by 90%. Cached input on GPT-6 Cyber is $2 per million against $20, with cache writes at $25 per million for 30 minutes or $50 for 12 hours. Scanners that resend the same system prompt and repository context on every finding benefit most.
In short: gpt-5.6-cyber is the cheaper Red option, GPT-6 Cyber costs 1.6 times as much on input and 1.33 times on output, and a Blue model is an order of magnitude cheaper for work that does not need Red.
A migration plan for the next 73 days
- This week. Search code, configuration, gateway routes and secrets stores for gpt-5.5-cyber-preview and the two gpt-daybreak aliases. Check the usage dashboard for the real call list.
- By 4 December. Classify each workload as Blue-eligible (scanning, triage, review, detection) or Red-only (reproduction, exploit validation, penetration testing). Submit the GPT-6 Cyber request through your OpenAI representative if Red work needs it.
- By 18 December. Build a test set of 50 to 100 past findings with known outcomes. Run it on gpt-5.6-cyber and, for Blue-eligible jobs, on gpt-6-sol. Record accuracy, tokens and any cyber_policy events.
- By 8 January 2027. Add safety_identifier to every request, set the IP allowlist on a named project, and ship the gpt-5.6-cyber change behind a flag.
- By 18 January 2027. Complete the cut-over, two weeks before shutdown, and remove the preview string so nothing falls back to it.
- After approval. Repeat the test set on GPT-6 Cyber in its own project and move only the jobs where it measurably improves results.
If your scanning runs through OpenAI's Codex Security CLI rather than direct API calls, our Codex Security CLI setup guide explains how the client selects its model. The same inventory, test and flag pattern for OpenAI's text models is in our GPT-5 and o3 shutdown migration guide.
India-specific considerations
OpenAI prices the API in US dollars, so the ₹1,39,950 monthly figure above moves with the exchange rate as well as with usage, before whatever tax treatment your finance team applies to imported services. Budget the 2027 line on measured tokens from your test set, not on the illustrative workload here.
Vulnerability work often touches personal data: application logs with customer identifiers, KYC flows in fintech code, or incident evidence. Under the Digital Personal Data Protection Act 2023, check that what you send for triage is covered by your notice and purpose, and strip or pseudonymise identifiers the finding does not need. The safety_identifier OpenAI asks for should itself be pseudonymous, not an employee email. Teams that need data kept out of OpenAI's retention should confirm Zero Data Retention separately, since Trusted Access does not grant it; our note on OpenAI Zero Data Retention limits covers which endpoints qualify.
How eCorpIT can help
eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our cybersecurity services team can inventory your OpenAI cyber model calls, test gpt-5.6-cyber, GPT-6 Cyber and Blue models against your past findings, and set up projects, allowlists and safety identifiers before the February shutdown. Talk to us at /contact-us/.
Last updated: 20 November 2026.