← Adoption Challenges 🕐 19 min read
Adoption Challenges

The 18-Month Cliff: What Happens to Your AI ROI After the First Deployment Works

Brandon Sneider · April 2026

The business case that got your AI deployment approved assumed a one-time investment with compounding returns. Month 3 looked like the case study. Month 7 started feeling different.

See also (wiki): wiki/ai-maturity-models.md, wiki/ai-vendor-contracts.md, wiki/ai-budget-cfo-decisions.md

Executive Summary

  • The research corpus on enterprise AI is overwhelmingly about getting to first production deployment. Almost nothing covers what happens next. That gap is where the ROI lives or dies.
  • 91% of machine learning models experience degradation over time. Models left unchanged for six months see error rates jump 35% on new data. Most organizations do not detect this until it has already cost them revenue. (MIT research, 32 datasets across 4 industries; 2025 LLMOps report)
  • License fees account for 10-17% of total AI spend. The remaining 83-90% accumulates in integration maintenance, prompt engineering updates, compliance re-audits, team capability upkeep, and consumption-based overages — none of which appear in the original business case. (CloudZero, 2025)
  • 73.8% of organizations are considering switching AI vendors between 2025 and 2028, but the average AI platform migration costs $315,000 per project, and the true switching cost compounds when agentic workflows are built on proprietary orchestration layers. (Futurum Group 1H 2026 survey; Swfte AI enterprise data)
  • The Forrester TEI 207% figure and similar ROI projections assume the deployment stays optimized. Staying optimized requires active monthly maintenance nobody budgets for — and governance re-audits that EU AI Act enforcement (August 2026) now mandates formally.
  • The organizations that sustain AI ROI through Years 2 and 3 budget for four things that never appear in vendor quotes: drift monitoring, compliance re-audit cycles, vendor switching optionality, and prompt engineering maintenance as a recurring line item.

The Problem Nobody Talks About

The business case that got your AI deployment approved assumed a one-time investment with compounding returns. Month 3 looked like the case study. Month 7 started feeling different. By month 12, the ROI numbers do not match the projection, and nobody in the room can explain why.

This is not a failure. It is the normal lifecycle of an AI deployment that was built for launch, not for operations.

The research literature has the same gap. A 2026 corpus analysis of enterprise AI studies finds the overwhelming majority focused on the pilot-to-production transition. The post-deployment maintenance literature is sparse, scattered across MLOps documentation and vendor monitoring guides rather than peer-reviewed evidence. The gap matters because that is precisely where the compounding returns either materialize or erode.

Four mechanisms drive the post-deployment ROI gap: model drift, hidden lifecycle costs, vendor lock-in dynamics, and governance re-audit requirements. Each operates on a different timeline. Together, they explain why the 56% of CEOs reporting no financial benefit from AI (PwC, n=4,454, September-November 2025) are not all running bad pilots — many ran good pilots whose gains simply evaporated after month 9.

Part 1: Model Drift — The Silent Performance Decay

How Fast Outputs Actually Degrade

Model drift is the gap that opens between the world your model was trained on and the world it is operating in now. It is silent. It does not announce itself with error messages. It shows up as gradually declining output quality, slightly higher customer escalations, marginally worse decisions — changes that are easy to rationalize as seasonal variation or team performance until the gap is large enough to be undeniable.

The data on how fast this happens is sharper than most practitioners expect.

MIT research examining 32 datasets across four industries finds 91% of machine learning models experience degradation over time. A 2025 LLMOps report finds that models left unchanged for six months see error rates jump 35% on new data. Performance degradation can occur at rates of 0.5% per week in dynamic environments — which compounds to roughly 26% annual degradation without any retraining. High-change environments (fraud detection, customer behavior prediction, dynamic pricing) see this faster; more stable environments see it over 12-18 months instead of 6.

The detection problem is compounding the damage. 75% of businesses observed AI performance declines without proper monitoring, and over half report measurable revenue losses from AI errors before they identify the source. In 2024, 67% of enterprises reported critical issues going unnoticed for more than a month without proper monitoring infrastructure.

Two types of drift have different time signatures:

Data drift — the statistical distribution of inputs shifts from the training data. Customer demographics change. Seasonal patterns affect input features. A new product launch changes the content of support tickets. This kind of drift can be detected early, before output quality visibly degrades, using input distribution monitoring.

Concept drift — the relationship between inputs and correct outputs changes. A fraud model trained before a new attack vector cannot detect the new pattern. A customer sentiment model trained in 2023 interprets post-ChatGPT customer communication differently than the current baseline. This is harder to detect because the output format looks correct — the model is confidently producing plausible answers to the wrong question.

Leading Indicators Before Performance Visibly Degrades

The critical insight from monitoring research is that waiting for business metrics to visibly degrade means the damage is already done. There are early warning signals in the data that appear before output quality drops:

Population Stability Index (PSI): measures how much the distribution of input features has shifted from the training distribution. PSI values above 0.10 warrant attention; above 0.25 require investigation and likely retraining. This is measurable continuously, in real time, without labeled outcome data.

Prediction distribution changes: when a model’s confidence scores shift — more outputs clustering near threshold values, unusual spikes in high-confidence wrong-direction predictions, or changes in the distribution of output labels — the model is likely encountering data that resembles but differs from its training distribution.

Feature importance drift: when the relative importance of input features shifts, the model is responding to different signals than it was trained on. This shows up before output accuracy visibly declines.

Calibration degradation: well-calibrated models produce confidence scores that accurately reflect actual accuracy. When calibration degrades — the model claims 90% confidence on predictions it gets right only 70% of the time — it is a leading signal of distributional shift.

What Triggers a Retraining Cycle

Retraining frequency depends on the deployment environment, but the evidence points to a hybrid approach rather than calendar-based schedules alone:

Environment Baseline retraining frequency Accelerated trigger
High-change (fraud, dynamic pricing, customer behavior) Weekly to monthly PSI > 0.25 on any core feature
Medium-change (credit risk, operations, HR scoring) Quarterly Accuracy drop > 3 points or PSI > 0.10 sustained 2 weeks
Stable (classification, document processing, content categorization) Every 3-6 months Measurable output quality decline > 5%
LLM-based workflows (summarization, drafting, Q&A) No model retraining (vendor-managed); prompt engineering review quarterly Output quality benchmarks drop 10%+

For LLM-based deployments — which describes most enterprise AI deployed in 2024-2025 — the model itself is not retrained by the customer; the vendor releases model updates on their own schedule. The customer’s equivalent of “retraining” is prompt engineering maintenance: updating system prompts, few-shot examples, and guardrail configurations to match current model behavior and current business context. This work is invisible in most budgets, costs $20,000-$60,000 annually for a mid-market deployment across 3-5 workflows, and is the primary mechanism by which LLM-based deployment quality stays calibrated over time.

Arize AI benchmark data finds that proactive retraining policies outperform reactive updates by 4.2x in maintaining prediction stability. The organizations that sustain ROI are the ones that build monitoring into their deployment architecture before launch, not after the first quality complaint.

Part 2: The Year 2-3 Cost Reality

What the Vendor Quote Does Not Include

The 3-year total cost analysis (see research/07-adoption-challenges/cfo-3-year-ai-tco-model.md) documents the full cost arc from Year Zero through Year Two. The lifecycle management layer adds five recurring cost categories that most organizations discover rather than budget for:

Prompt engineering maintenance ($20,000-$60,000/year): System prompts and few-shot examples that worked at launch degrade as model vendors update their underlying models. GPT-4o behaved differently from GPT-4. GPT-5 handles system messages differently from GPT-4o. OpenAI deprecated GPT-4 in April 2025 and GPT-4o’s Assistants API in August 2026. Each deprecation event requires prompt re-engineering, output quality re-testing, and often integration code changes. For a 500-person company running AI across 4-6 workflows, this is a $20,000-$60,000 annual line item that compounds as the number of workflows grows.

Monitoring and drift management ($15,000-$40,000/year): Running the monitoring infrastructure, reviewing alerts, and managing retraining cycles for ML-based deployments (vs. LLM-based). For organizations with custom-built or fine-tuned models, this includes the engineering time to execute retraining cycles, re-test against validation datasets, and roll out updated models with appropriate change management.

Compliance re-audit costs ($25,000-$75,000 biannually): The risk assessment and governance documentation that supported the original deployment does not stay current on its own. Regulatory requirements change — the EU AI Act reached full enforcement in August 2026, requiring operational evidence rather than policy documents. Model behavior changes with vendor updates. Business context changes. A deployment classified as medium-risk at launch may be high-risk 18 months later if the company’s regulatory exposure has changed. Quarterly reviews for medium-risk systems and monthly control audits for high-risk systems are the 2026 standard — neither was a budget line in the 2024 business case.

Team capability maintenance ($30,000-$80,000/year): The person who built the deployment leaves. The prompt engineer who knew how the system worked moves to a competitor. The data analyst who ran the governance review is on a different project. Mid-market organizations run this risk harder than enterprises because their AI teams are thinner — often two or three people who are simultaneously the architects, operators, and trainers. Rebuilding this institutional knowledge after turnover costs more than maintaining it. The remediation literature (Pertama Partners, 2026) finds abandoned AI projects average $4.2M in sunk costs, with the median time to abandonment at 11 months — the exact window where maintenance knowledge gaps become deployment quality gaps.

Integration maintenance ($20,000-$50,000/year): The ERP gets upgraded. The CRM changes its API. The HR system moves to a new version. Every integration point between the AI deployment and the rest of the technology stack requires ongoing maintenance as the surrounding systems evolve. Integration maintenance costs run 15-25% of the original integration investment annually — often the largest hidden cost in Year 2.

The Realistic Year 2-3 Maintenance Budget

For a 500-person company with a $400,000 Year One deployment across 4-6 workflows, Year 2-3 operational maintenance runs:

Category Annual Range Notes
Prompt engineering maintenance $20,000-$60,000 Per vendor model update cycle
Monitoring and drift management $15,000-$40,000 Scales with number of workflows and ML vs. LLM mix
Compliance re-audits $25,000-$75,000 Biannual; higher if EU-facing or regulated industry
Team capability maintenance $30,000-$80,000 Training, documentation, knowledge transfer
Integration maintenance $20,000-$50,000 15-25% of original integration investment
Annual maintenance total $110,000-$305,000 On top of licensing and usage costs

That $110,000-$305,000 annual maintenance figure represents the gap between what the business case projected and what the deployment actually requires. Most Year One ROI analyses assume the deployment runs at launch-day performance indefinitely, with only licensing costs increasing. The maintenance reality means Year 2-3 economics look structurally different from the pilot projection.

Part 3: Vendor Lock-In and Switching Costs

The Multi-Layer Lock-In Problem

The vendor relationship that looked like a straightforward API subscription at deployment often turns into a significantly more durable dependency by month 12. This is not deliberate entrapment — it is the natural accumulation of organizational investment in a specific platform.

Lock-in now accumulates at four simultaneous layers, not one:

Model layer: Fine-tuning investments, custom training data, and workflow-specific prompt libraries represent organizational knowledge embedded in a specific model architecture. These artifacts often require significant re-engineering to port to a different model — not because the intellectual content cannot transfer, but because the operational mechanics differ.

Orchestration layer: Organizations deploying agents or multi-step AI workflows via a vendor’s proprietary orchestration framework (OpenAI’s Assistants API, Microsoft Copilot extensions, Salesforce Agentforce) face the highest switching costs. When OpenAI deprecated the Assistants API with an August 2026 shutdown deadline, enterprises that had built production workflows on it faced complete migration projects — not because they wanted to switch vendors, but because the vendor discontinued the product.

Integration layer: The longer a deployment runs, the more deeply the AI system is wired into surrounding infrastructure. Customer data flows through it. Human workflows depend on it. Approval chains reference its outputs. Each of those connections is a switching cost.

Data layer: Prompt engineering history, evaluation datasets, monitoring baselines, and organizational calibration data represent months of investment that live inside vendor platforms. Without contractual provisions for artifact export, this intellectual capital stays with the vendor at termination.

The Realistic Migration Cost

The existing corpus documents the migration cost anatomy directly (see research/07-adoption-challenges/ai-contract-portability-and-exit-terms.md). Swfte AI enterprise data shows the average AI platform migration at $315,000 per project, ranging from $200,000 to $500,000 for a 300-person company migrating a single AI platform. When the migration does not go smoothly, costs can reach twice the initial implementation investment.

The 2025-2026 vendor landscape has accelerated the frequency of forced migrations:

  • OpenAI deprecated GPT-4 in April 2025, GPT-4o’s Assistants API in August 2026 (3-month notice)
  • The model replacing GPT-4o handles system messages differently, enforces stricter JSON schema adherence, and uses a different API architecture — requiring re-engineering even for customers staying on the same vendor
  • Open-weight model performance has narrowed the gap with closed models from 8% to 1.7% on some benchmarks in a single year (Stanford AI Index), meaning the switching cost justification (“our vendor has better AI”) becomes weaker every quarter

73.8% of organizations are considering switching vendors between 2025 and 2028 (Futurum Group 1H 2026 survey). The barrier is not technical capability judgment — it is the migration economics. Organizations that negotiated exit provisions at contract time (data export SLAs, transition assistance obligations, model deprecation notice periods) have measurably lower switching costs when the time comes. Organizations that did not are stuck in increasingly expensive incumbent relationships.

What a Multi-Model Strategy Actually Requires

The mature enterprise response to vendor lock-in is a multi-model architecture: different vendors for different use cases, with open standards (Model Context Protocol, standardized prompt formats, vendor-neutral evaluation frameworks) creating interoperability. This is the right architectural answer. It is also 30-40% more expensive to operate than a single-vendor deployment in Years 1-2, because it requires internal capability to manage multiple vendor relationships, different API patterns, and divergent monitoring approaches.

For a 200-2,000 person company running its first deployment, the practical answer is narrower: negotiate the exit provisions now, before they matter. The specific provisions that reduce migration cost — data export SLAs, artifact ownership clarity, model deprecation notice periods, transition assistance obligations — add weeks to procurement but save months in forced migrations.

Part 4: Governance Re-Audit Requirements

What “Compliance” Looks Like at Month 18

The governance documentation that supported the original deployment had a specific scope: the system as it was designed, deployed on a specific model version, for a defined set of use cases. By month 18, several things have changed that affect the original risk assessment:

  • The underlying model has been updated by the vendor — often without formal notification
  • Use cases have expanded beyond the original scope as users find new applications
  • The regulatory environment has evolved (EU AI Act full enforcement August 2026)
  • The organization’s regulatory exposure may have changed (new markets, new products, M&A)
  • Documented human oversight procedures may have drifted from actual practice

The EU AI Act enforcement framework makes the re-audit requirement concrete. Before August 2026, screenshots and policy declarations were acceptable compliance evidence. The 2026 standard requires operational evidence: audit logs connected to production systems, demonstrable human oversight mechanisms in runtime, and conformity assessments complete and inspection-ready.

For medium-risk AI systems: quarterly reviews to confirm usage remains within authorized boundaries, with documentation updated to reflect current model behavior.

For high-risk AI systems: monthly control audits, formal human-in-the-loop validation, and continuous monitoring during operation.

Most enterprises are not positioned for this cadence. Grant Thornton’s 2026 AI Impact Survey (n=950, US, February-March 2026) finds 78% of organizations lack confidence they could pass an independent AI governance audit in 90 days. That is not a compliance posture — it is a liability exposure that grows with each quarter the deployment runs without a re-audit.

The Re-Audit Trigger Framework

Beyond regulatory mandates, four events should trigger a formal governance re-audit regardless of schedule:

Model version change by vendor: When the underlying model is updated, the original risk assessment is invalidated. Output characteristics, edge case behavior, and bias profiles may have changed. A re-audit does not need to be a full reassessment — but it needs to confirm that the deployment’s behavior matches the documented expectations.

Use case scope expansion: When a workflow built for one purpose is extended to adjacent applications, the original risk classification may no longer apply. A document summarization tool extended to generate customer-facing communications shifts from lower-risk to higher-risk with different oversight requirements.

Significant personnel change: When the people who built the system leave, institutional knowledge of why certain safeguards were implemented, what edge cases were documented, and how the monitoring baselines were established leaves with them. A re-audit after significant turnover rebuilds the governance foundation.

Material business context change: New regulatory filings, M&A activity, geographic expansion, or new customer segments can all affect the risk classification of an existing deployment.

The organizations that build re-audit triggers into their governance calendar before problems arise are the ones that pass the audit when it counts. The organizations that treat re-audits as one-time events discover they have compliance gaps at the worst possible moment.

The Maintenance Deficit: Why ROI Projections Miss

The Forrester TEI framework that produces figures like 207% ROI over three years assumes deployments remain at launch-day performance for the entire projection period. The lifecycle management evidence shows that assumption fails unless active maintenance is funded.

The gap between projected and actual ROI in Year 2-3 can be traced to four specific deficits:

  1. No monitoring budget: 75% of businesses that observe performance declines do so without proper monitoring infrastructure. Drift accumulates for 6+ months before it is detected.

  2. No prompt engineering maintenance budget: Vendor model updates, which are not announced on enterprise-friendly timelines, require prompt re-engineering that organizations treat as zero-cost because it is invisible in the license fee.

  3. No governance re-audit budget: Compliance review was a one-time pre-deployment activity. Operating companies treat it as the cost of launch, not the cost of running.

  4. No switching cost reserve: Organizations commit to three-year vendor contracts without budgeting for the migration costs that forced deprecations or better competitors will eventually require.

The correction is not to project lower returns. It is to budget for the maintenance that earns the projected returns.

Key Data Points

Metric Value Source Date Credibility
AI models experiencing degradation over time 91% MIT research, 32 datasets across 4 industries 2025 HIGH — peer-reviewed multi-industry study
Error rate increase for models unchanged 6+ months 35% 2025 LLMOps report 2025 MEDIUM — industry report, n not specified
Businesses observing AI performance declines without monitoring 75% SmartDev aggregation of monitoring vendor data 2025 MEDIUM — vendor-aggregated, directionally consistent
Critical issues undetected for 1+ month (2024) 67% of enterprises SmartDev (2024 survey data) 2024 MEDIUM — monitoring vendor data, published 2025
Proactive retraining advantage over reactive 4.2x stability Arize AI benchmarks 2025 MEDIUM — vendor benchmark, but directionally robust
Performance degradation rate (dynamic environments) 0.5%/week Industry monitoring standards 2025 MEDIUM — widely cited, derived from practitioner data
License fees as % of total AI spend 10-17% CloudZero, n=500 March 2025 HIGH — platform data from 500 US software leaders
Average AI platform migration cost $315,000 Swfte AI enterprise survey 2025 MEDIUM — enterprise survey, n not published
Organizations considering vendor switches 2025-2028 73.8% Futurum Group 1H 2026 Decision Maker Survey 2026 MEDIUM-HIGH — enterprise decision-maker survey
Organizations unable to pass AI governance audit in 90 days 78% Grant Thornton, n=950 Feb-Mar 2026 HIGH — independent survey, US operations-weighted
EU AI Act full enforcement date August 2, 2026 EU Official Journal 2024 HIGH — regulatory source
Medium-risk AI systems: required review frequency Quarterly EU AI Act / Governance Intelligence 2026 HIGH — regulatory requirement
High-risk AI systems: required audit frequency Monthly EU AI Act / Governance Intelligence 2026 HIGH — regulatory requirement
Open-weight vs. closed model performance gap 8% → 1.7% in one year Stanford AI Index 2025 HIGH — independent research institution
CEOs reporting no AI financial benefit 56% PwC, n=4,454 Sep-Nov 2025 HIGH — large independent CEO survey
Annual AI platform maintenance as % of implementation cost 15-30% Industry benchmarks, multiple sources 2025-2026 MEDIUM — range from multiple independent analyses

What This Means for Your Organization

If your AI deployment is between 6 and 18 months old and the ROI is not matching the business case, the first diagnostic question is not “what’s wrong with the technology?” It is “which of the four maintenance deficits do you have?”

Three immediate actions for organizations in this position.

First, run a deployment health audit before the next renewal cycle. Compare current output quality against the baseline quality assessment from the first month of production. For LLM-based deployments, benchmark the current system prompt against a fresh prompt build for the same task — model updates often mean the original prompt is underperforming against what the current model can do. The delta between where the system is and where it should be is the prompt engineering backlog.

Second, build the governance re-audit into the annual calendar now, rather than when an audit is demanded. The 78% of organizations that cannot pass a governance audit in 90 days are not non-compliant because they made bad decisions — they are non-compliant because governance was a launch activity, not an operational cadence. Quarterly documentation reviews, annual full assessments, and event-triggered re-audits for model updates and use case expansions are the three-layer calendar that keeps the governance posture current.

Third, negotiate switching cost reductions at next renewal. If current vendor contracts lack data export SLAs, transition assistance obligations, and model deprecation notice periods, the renewal conversation is the moment to add them. The EU Data Act (effective September 2025) mandates data portability for EU-facing services; use it as the floor position regardless of jurisdiction. The cost of negotiating these provisions now is weeks of procurement time. The cost of not having them when a forced migration occurs is $315,000 plus the organizational disruption of a six-month migration project.

The organizations that sustain AI ROI through Year 3 are not running fundamentally different deployments. They are running the same deployments with the operational discipline that the original business case assumed but never funded. If building that maintenance architecture for a specific deployment is a decision worth an outside perspective, I’m reachable at brandon@brandonsneider.com.

Sources

  1. MIT Research — Model Degradation Study (2025). 32 datasets across four industries; 91% of models experience degradation over time. Peer-reviewed academic research. Cited via SmartDev aggregation at https://smartdev.com/ai-model-drift-retraining-a-guide-for-ml-system-maintenance/ — HIGH credibility.

  2. LLMOps Community Report — “State of LLMOps 2025.” Models unchanged for 6+ months see 35% error rate increase on new data. Industry practitioners’ report. Cited via SmartDev. — MEDIUM credibility (n not specified, but directionally consistent with monitoring vendor data).

  3. Arize AI — Retraining Policy Benchmarks (2025). Proactive retraining outperforms reactive by 4.2x in prediction stability. Monitoring platform vendor data. https://arize.com — MEDIUM credibility (vendor data, potential commercial interest, but methodology consistent with practitioner consensus).

  4. Futurum Group — 1H 2026 Enterprise Software Decision Maker Survey (2026). 73.8% considering vendor switches 2025-2028; 66% favor platform-first. Enterprise decision-maker survey. https://futurumgroup.com/insights/will-technology-friction-derail-the-roi-promise-of-enterprise-ai-investments/ — MEDIUM-HIGH credibility.

  5. Swfte AI — Enterprise AI Migration Cost Analysis (2025). $315,000 average platform migration cost; 2x implementation cost when migrations fail. Enterprise survey. Cited in research/07-adoption-challenges/ai-contract-portability-and-exit-terms.md. — MEDIUM credibility (n not published, enterprise-weighted).

  6. Grant Thornton — “2026 AI Impact Survey” (February-March 2026, n=950, US). 78% of organizations cannot pass AI governance audit in 90 days; 58% fully integrated report revenue growth vs. 15% still piloting. Independent survey. https://www.grantthornton.com — HIGH credibility.

  7. Kai-Waehner — “Enterprise Agentic AI Landscape 2026: Trust, Flexibility, and Vendor Lock-in” (April 2026). Multi-layer lock-in analysis; MCP as architectural mitigation. https://www.kai-waehner.de/blog/2026/04/06/enterprise-agentic-ai-landscape-2026-trust-flexibility-and-vendor-lock-in/ — MEDIUM credibility (independent practitioner analysis).

  8. EU AI Act (Official Journal of the European Union, 2024). Full enforcement August 2, 2026; medium-risk quarterly reviews; high-risk monthly audits; penalties to €35M or 7% global turnover. https://artificialintelligenceact.eu/ — HIGH credibility (regulatory source).

  9. Stanford AI Index — “AI Index Report 2025” (2025). Open-weight vs. closed model performance gap narrowed from 8% to 1.7% on benchmarks in one year. Stanford HAI. https://aiindex.stanford.edu/report/ — HIGH credibility.

  10. CloudZero — “State of AI Costs 2025” (March 2025, n=500 US software leaders). License fees 10-17% of total AI spend; average monthly AI spend $85,521. Platform data. https://www.cloudzero.com/state-of-ai-costs/ — HIGH credibility.

  11. PwC — 29th Global CEO Survey (January 2026, n=4,454). 56% report no AI financial benefit; 12% report both revenue and cost gains. https://www.pwc.com/gx/en/ceo-survey/ — HIGH credibility.

  12. Fulcrum Digital — “AI Model Drift in Production: What Enterprises Must Monitor” (2025). Enterprise monitoring framework; proactive vs. reactive detection distinction. https://fulcrumdigital.com/blogs/ai-model-drift-in-production-what-enterprises-must-monitor/ — MEDIUM credibility (consulting firm practitioner guidance).

  13. OpenAI — Deprecation notices (2025-2026). GPT-4 deprecated April 2025; Assistants API shutdown August 26, 2026; Sora API shutdown September 24, 2026. https://platform.openai.com/docs/deprecations — HIGH credibility (primary source).

Brandon Sneider | brandon@brandonsneider.com April 2026