Free estimator · GPT-5.5 leaves ChatGPT, ChatGPT Work and Codex on 2026-10-14 · API unaffected

GPT-5.5 retirement October 14: model migration cost estimator

The GPT-5.5 retirement on October 14, 2026 removes the model from ChatGPT, ChatGPT Work and Codex on all plans. The API is not affected. Every workspace default, custom agent, scheduled task and script that names gpt-5.5 has to move, and the cost is not the code change. It is the regression testing: proving every prompt and tool call still behaves on the replacement. Enter your agent count, prompts per agent, eval coverage and target model. You get engineering hours, the eval gap to close first, the run-cost delta, and whether you make the date.

6 agents
8
30 %
400 M tokens
$130 $/h
2
1
Estimated migration cost
$19.0K
146 engineering hours at $130/h
Hours to build missing evals first
50
4 of 6 agents have no eval suite
Ready by 2026-10-14?
At risk
20 days x 6 h x 1 engineer = 120 h
Migration cost by component (USD)
Re-point + smoke test
1.6K
Prompt / tool regression
9.4K
Config surfaces
1.6K
Missing evals
6.6K
Migration cost by component
Re-point + smoke test1560
Prompt / tool regression9360
Config surfaces1560
Missing evals6551.999999999999
Monthly run cost on GPT-5.6 Sol
$5,000 (+$0 vs GPT-5.5 at $5,000)
Tokenizer adjustment
none
Retirement scope
ChatGPT, ChatGPT Work, Codex on all plans; API unaffected
Eval coverage benchmark
Only 52% of teams run offline evals (LangChain, 2026-03)

146 hours, about $19,032, to move 6 agents to GPT-5.6 Sol; readiness by 2026-10-14: at risk.

Illustrative model, not a quote. Cost bands are published third-party 2026 market ranges, never TheoSym or QGI prices. Per-token prices verified 2026-09-19; OpenAI prices are from a secondary tracker. Not legal, financial or tax advice.
Assumptions and sources (15)
ConstantValueBasis
Retirement date and scopeGPT-5.5 leaves ChatGPT, ChatGPT Work and Codex on 2026-10-14; API unaffectedsourcedOpenAI model release notes
Re-point + smoke test2 h per agentestimateStudio estimate, to be replaced by measured data
Prompt / tool regression1.5 h per promptestimateStudio estimate
Config surfaces1 h per surface per agentestimateStudio estimate
Minimal eval suite12 h per agent for ~30 casesestimateStudio estimate; rationale from the 3-level production testing framework
Capacity6 productive hours per engineer per day until 2026-10-14; at risk within 25%estimateStudio estimate
GPT-5.5 price$5 in / $30 out per MToksourcedSECONDARY: CloudZero tracker 2026-09-10 (openai.com returned 403)
GPT-5.6 Sol price per MTok$5 in / $30 outsourcedSECONDARY: CloudZero GPT-5.6 pricing tracker (updated 2026-09-10). openai.com/api/pricing returned HTTP 403 on 2026-09-19
GPT-5.6 Terra price per MTok$2 in / $12 outsourcedSECONDARY: CloudZero GPT-5.6 pricing tracker (updated 2026-09-10). openai.com/api/pricing returned HTTP 403 on 2026-09-19
Claude Sonnet 5 price per MTok$2 in / $10 out (introductory price made permanent)sourcedAnthropic pricing page, fetched 2026-09-19
Claude Opus 5 price per MTok$5 in / $25 outsourcedAnthropic pricing page, fetched 2026-09-19
Gemini 3.6 Flash price per MTok$0.75 in / $3.75 out ($1.50 / $7.50 / $0.15 after 2026-12-31)sourcedGoogle Gemini API pricing page, fetched 2026-09-19
Token split70% input / 30% outputestimateStudio estimate
Claude tokenizer~30% more tokens for the same text on Claude 4.7+sourcedAnthropic pricing page, 2026-09-19
Eval coverage benchmark52% run offline evalssourcedLangChain State of Agent Engineering, 2026-03-03

Key takeaways

  • GPT-5.5 leaves ChatGPT and Codex on 2026-10-14. The API keeps it. Anything that names the model in a UI or config has to move.
  • Re-pointing takes hours. Proving the agent still works takes days, unless the eval suite already exists.
  • GPT-5.6 Sol is priced the same as GPT-5.5 ($5 / $30). Terra and Claude Sonnet 5 are cheaper if evals hold.
  • Only 52% of teams run offline evals. That is why retirements become incidents.

What happens to GPT-5.5 on October 14, 2026

OpenAI retires GPT-5.5 from ChatGPT, ChatGPT Work and Codex on all plans on 14 October 2026 (OpenAI release notes). The OpenAI API is not affected. Codex users signed in with ChatGPT are pointed to GPT-5.6 Sol. Workspace defaults, custom agents, scheduled tasks and scripts that select gpt-5.5 stop working as configured.

GPT-5.5 shipped to the API on 2026-04-24. Six months from launch to retirement is the new normal, and it is a line item. TheoSym ships every production agent with the model id behind a router and an eval suite that gates the swap, so the next retirement is a pull request, not a project.

How model migration cost is calculated

Four lines. Re-pointing and a smoke test per agent. Regression per prompt and tool definition, because a model change alters tone, tool-call formatting and refusal behaviour. One hour per configuration surface per agent. And the eval suites you do not have yet, at about twelve hours each for a minimal 30-case set.

The last line is usually the biggest. Without an eval you are not migrating, you are guessing. LangChain's 2026 survey found only 52% of teams run offline evals (State of Agent Engineering). The per-agent hours here are studio estimates; the AI Factory workproofs show measured runs.

  • Hours = agents x (2 + 1.5 x prompts + surfaces) + missing evals x 12
  • Cost = hours x engineer rate
  • Ready = days to Oct 14 x 6 h x engineers >= hours

GPT-5.5 to GPT-5.6 Sol vs Claude Sonnet 5 vs Gemini 3.6 Flash

Sol is the like-for-like: $5 in and $30 out, identical to GPT-5.5, so the swap is not a price change (CloudZero tracker; secondary source, openai.com returned 403 on 2026-09-19). Terra at $2 and $12 halves the bill if the task tolerates it. Claude Sonnet 5 at $2 and $10 is cheaper still on list, with about 30% more tokens on the Claude 4.7+ tokenizer, which the estimator applies. Gemini 3.6 Flash is the floor at $0.75 and $3.75 through December.

The right target is the cheapest model that passes your evals, not the cheapest on the price list. That comparison is a standard step in TheoSym's AI consulting engagements. Regulated decisioning agents route to QGI, where a model swap has to leave the decision output reproducible.

When do you need a model migration checklist?

Now, if any surface names gpt-5.5, and before the next retirement in any case. A working checklist: inventory every place a model id is set; write or complete the eval per agent; run the candidate model against it; compare cost per task; swap behind a router; monitor the online evals for a week (Developers Digest checklist).

Model churn is a control problem, not a vendor problem. An agent whose behaviour you cannot regression-test is an agent you do not control, whatever model sits under it. Dr. Sam Sammane and the TheoSym team publish shipped work in the press center; bring your inventory to the strategy call.

GPT-5.5 retirement October 14: questions people ask

What happens to GPT-5.5 on October 14, 2026?+

OpenAI retires GPT-5.5 from ChatGPT, ChatGPT Work and Codex on all plans on 14 October 2026. The OpenAI API is not affected. Codex users signed in with ChatGPT should switch to GPT-5.6 Sol, and any workspace defaults, custom agents, scheduled tasks and scripts that select gpt-5.5 must be updated.

How much does it cost to migrate an AI agent to a new model?+

The engineering cost is dominated by regression testing, not the code change. Re-pointing an agent takes hours; verifying that every prompt and tool call still behaves takes days unless an eval suite already exists. The estimator prices both, plus the run-cost difference of the replacement model.

Will GPT-5.6 Sol cost more than GPT-5.5?+

No. Sol is priced identically at $5 per million input tokens and $30 per million output, so the swap is not a price increase. Terra ($2 / $12) and Luna ($0.20 / $1.20) are cheaper tiers if your evals show they hold quality for the task. Prices are from a secondary tracker dated 2026-09-10.

How do I avoid this problem at the next model retirement?+

Keep a repeatable offline eval suite per agent, abstract the model id behind a router, and gate every model swap on eval pass rates. Only about half of teams run offline evals today, which is why retirements become incidents. Eval-gated swaps across providers are what a production harness provides.

Are the per-agent hours in this estimator measured?+

No. Two hours to re-point, 1.5 hours per prompt to regress, one hour per configuration surface and twelve hours per missing eval suite are studio estimates, labelled as such under assumptions. They are meant to size the work before October 14, not to quote it. Replace them with your own timings as soon as you have one migration done.

Next step

Turn the estimate into a scoped plan

Send the calculator result. You get one reply from a human within one business day, with an eval plan for the workflow and a real scope.

Prefer to talk? +1 657-888-0688 or contact@theosym.com

Turn the estimate into a scoped plan

No spam. One reply from a human. Unsubscribe any time.