The Migration Bribe

On May 13, 2026, OpenAI launched a 30-day promotion. Two months of free enterprise usage for any organization willing to migrate to Codex from a rival platform. Anthropic answered within hours. Claude Code weekly limits jumped 50 percent through July 13. Two of the biggest names in agentic AI looked each other in the eye and started paying customers to switch.

This is the part most CIOs read wrong. They see a price war. A buying opportunity. A chance to renegotiate. They are missing the more interesting signal. When a vendor offers you two months of free usage to leave a competitor, the vendor is telling you exactly what they think it costs to stay. They have measured the wall. They are quoting the toll.

The vendors are bidding to acquire you. You are the asset on the block.

The promo is the receipt of the lock-in

A migration bribe is a strange artifact. It only exists when switching is hard. If customers could swap between models the way they swap between AWS regions, there would be no two-month coupon. There would be a single line item in a billing dashboard. Both Codex and Claude Code are betting that a coupon is enough to overcome a real, costed, structural friction. They are right.

The Register ran a piece on April 28, 2026 with a title most enterprise architects already feel in their kneecaps. Locked, stocked, and losing budget: AI vendor lock-in bites. The article tracks what is happening across enterprise IT right now. Without deliberate architectural planning, a buyer can reach a point where switching providers requires rewriting applications, retraining staff, migrating prompt libraries, rebuilding integrations, and renegotiating data agreements at the same moment. That is the wall the promo is measuring.

Vendor lock-in in the cloud era took eight years to develop. With AI it is happening inside eighteen months. Prompt libraries grow on the order of thousands per quarter at any company that takes AI seriously. Agent memory accumulates daily. Tool integrations harden as engineers settle in. Each of these creates a small switching cost. Stacked together they form a wall a coupon cannot scale.

The fact that the vendors are now bidding for migrations suggests they have priced that wall and decided the customer-acquisition math still pencils. They are betting that once you arrive on their stack, the next coupon you receive will not move you. They are usually right.

Anthropic's confession in plain sight

If the promo war is the loud part, Anthropic's May launches are the quiet part. On May 8, 2026, VentureBeat covered the new Claude Managed Agents offering and used a word the rest of the industry has been tiptoeing around. Lock-in. The piece pointed out that session data, agent memory, evals, connectors, permissions, telemetry, and orchestration now all live inside Anthropic's own infrastructure. The customer gets faster deployment. Anthropic gets a deeper hold.

A separate analysis from Forrester surfaced Anthropic's services arm. Implementation, training, and process redesign sold by the model vendor itself. OpenAI is doing the same. PwC, Accenture, and Deloitte are all training armies of consultants who tie their billable hours to specific vendor stacks. This is the second layer of the wall. Even if you wanted to port the model, the humans who know how your agents work were trained to think in one syntax.

Across the same month Anthropic shipped Claude Opus 4.7, a SpaceX Colossus capacity deal, Claude for Small Business, three new Managed Agents features, a 200 million dollar partnership with the Gates Foundation, and an expanded PwC alliance. Read that list again. Almost every line item is a tighter knot in the same rope. More capacity to depend on. More managed services to depend on. More named partnerships that make the alliance feel inevitable.

None of this is dishonest. Anthropic is doing what good companies do. They are building a product that wins by being easier to use than the alternatives. The point is that ease is the wall. A platform that does more of your work for you takes a deeper hold of how your work actually gets done.

OpenAI is running the same play with a different mood. The May 2026 Codex push, the merge of ChatGPT and Codex teams, the GPT-5.5 Instant and GPT-5.6 releases. Each one is a step toward an integrated stack you cannot disassemble without losing capability. The promo is the welcome mat. The architecture is the trap.

What the lock-in actually costs

Pricing is the visible bill. Pricing is also the part of vendor risk a procurement team can read. When CIOs talk about lock-in they usually mean future negotiating leverage, and that is real. But it is the smallest part of the cost.

The bigger costs are structural and invisible until you try to move.

Prompt libraries are the first surprise. A company that built a serious AI practice over the last two years has somewhere between two and ten thousand active prompts. They were written for a specific model's behavior. Claude Sonnet's preference for structured reasoning. GPT-4o's tolerance for instruction stacking. Gemini's particular failure modes around long context. When you port a prompt library across vendors, output quality drops, sometimes by a lot. The fix is not a translation. It is a rewrite, prompt by prompt, with a re-evaluation harness. That work takes months of senior engineering time. It is rarely budgeted.

Agent memory is sticky in a way contracts are not

Modern agents accumulate context across sessions. User preferences. Past interactions. Domain knowledge. Tool usage patterns. If that memory lives in Anthropic's managed store, moving it requires an export pipeline that may or may not exist. If it lives in OpenAI's Assistants threads, the same problem. Memory state outlasts every clause in your master agreement.

Tool integrations are the part most teams discover too late

Every tool your agent uses is a piece of glue code written against a specific function-calling protocol. When that protocol changes, your tools break. When you switch model vendors, every integration touches the floor. The work of rebuilding tool graphs across a hundred internal services is a quarter of engineering time minimum.

Fine-tunes and retrieval indexes do not port

A vendor-specific fine-tune does not move across providers. An embedding store built with OpenAI's text-embedding-3-large cannot be queried with Anthropic's embedding API. A retrieval pipeline that was tuned for one model's grounding behavior breaks against another model's instruction following.

Add these up and the cost of switching is rarely less than 18 to 24 months of senior engineering load. The promo gives you 60 days of free tokens. The math does not balance. The vendors know it does not balance. That is why they are willing to write the check.

Avoiding AI vendor lock-in starts at the wiring diagram

The good news is that the infrastructure for portability is now real and mostly free. The choice not to use it has become a strategic decision rather than a technical limitation.

Three pieces matter.

The first is an AI gateway. LiteLLM crossed 40,000 stars on GitHub during 2025 and now ships an OpenAI-compatible interface across more than 100 model providers. Portkey open-sourced its full gateway under Apache 2.0 in March 2026, including guardrails, PII redaction, and audit logging. Kong AI Gateway lives in the same neighborhood for teams already on Kong. Any of these layers let your application code talk to a single API while the routing layer decides which model takes the call. The day Anthropic raises prices, you change a config line. The day OpenAI breaks a deployment, traffic flips to a backup provider. Failover, cost optimization, and quality-based routing all stop being a vendor's roadmap and start being your own.

The second piece is the Model Context Protocol. Anthropic introduced MCP in November 2024 as an open standard for how AI agents connect to tools, data sources, and memory. By December 2025 it had crossed 10,000 active public servers and 97 million SDK downloads, with adoption inside ChatGPT, Cursor, Gemini, Microsoft Copilot, and VS Code. CIO Magazine ran a piece in May 2026 with a headline that says the strategic part out loud. Why Model Context Protocol is suddenly on every executive agenda. MCP turns tool integration from an N times M problem into N plus M. Every tool exposes itself as an MCP server once. Every MCP-compatible agent consumes it. The day you change models, your tool graph travels with you.

The third piece is the memory layer. If agent memory lives in a vendor's managed database, you do not own it. If it lives in your own Postgres or vector store, you can hand any model the same context tomorrow. The architecture decision is whether your AI's institutional knowledge is a deposit you make with a single vendor, or a substrate the vendor rents access to.

These three moves, gateway plus MCP plus owned memory, take a working AI stack and make it portable in a way the vendor cannot undo by quietly raising prices on you. The work is not free, but it is bounded. It is the kind of build that pays for itself the first time a vendor calls a meeting about renewal terms.

Why the discount is a trap

There is a subtle case to take the OpenAI Codex deal. Two free months. New capability. A chance to compare. Read the case carefully. The promotion converts your current lock-in into a future lock-in. Two months of free usage is enough time for your engineers to settle in. Long enough to migrate prompt libraries. Long enough to rebuild integrations. Long enough to forget that the next coupon is going to come from someone else.

The vendors are not stupid. They know the migration window is the highest-leverage moment in the entire sales cycle. They are paying you to spend that window inside their walls. Once you build the new dependencies, the next renewal is on their terms.

The strategic answer is to build a stack that makes the discount irrelevant. A gateway-fronted, MCP-wired, memory-owned AI architecture can take the OpenAI promo and the Anthropic promo at the same time. It can route the cheap work to Codex during the free period and route the high-quality work to Claude. It can switch back on July 14 when the promo ends and nothing breaks. The discount becomes free margin instead of a future hostage situation.

This is the asymmetry the vendors are trying to prevent. They are betting you will not build the stack. Their pricing only works if most enterprise buyers are locked in by default. The minority that architect for portability will eat their lunch on cost while the majority pay the next round of price hikes.

Why this is a board problem, not a procurement problem

Most enterprises are treating AI vendor lock-in as a question for the CIO and the procurement team. That is the wrong altitude. Procurement can win marginal concessions on a contract. They cannot architect a portable stack. A portable stack is a six-quarter capital allocation. It touches data architecture, application architecture, agent design, and operations. It requires a CFO and a CTO who agree on what "owned AI capability" actually means.

The InfoWorld piece on the new AI lock-in put the strategic question well. Lock-in shows up in architecture, not in procurement contracts. Each agent that depends on a vendor-specific memory or tool calling format is structural vendor risk. Each agent built on a portable layer is a hedge. The aggregation of those decisions over the next two years will decide who owns the cost curve in AI.

This is the moment the question matters most. The vendors are still spending venture capital and recently issued debt to buy market share. They are bidding against each other to acquire you. The discounts are real today. They will not exist in 2028. By then the prompt libraries, agent memories, and tool integrations of every enterprise that did not architect for portability will be the moat the vendors price against.

If you take only one signal from May 2026, take this one. The companies that built AI to be portable will spend the rest of the decade paying the vendor's compute cost. The companies that did not will spend it paying the vendor's profit margin.

Skip the bribe. Build the architecture.

OpenAI's promo and Anthropic's response are the loudest moment yet of two vendors admitting how much they need you to stay. The honest reading of the moment is not to accept the discount. It is to ask what architecture makes you immune to the next one.

Building that architecture is a serious project. It involves a model gateway. It involves an MCP-based tool layer. It involves owned memory. It involves an evaluation harness that runs across providers so you know which model wins on which task this week, and next month, and the month after. It involves rebuilding parts of your AI practice that already feel like they are working, because the version that works today is also the version that is most exposed.

This is the work most enterprises are still postponing. The promos are designed to keep you postponing it. Every month you spend deeper inside a single vendor's stack is a month the wall grows. By the time the next discount arrives, the wall is taller than the discount.

Avoiding AI vendor lock-in is no longer a procurement skill. It is a structural commitment that has to be designed in, layer by layer, with the same care your data architecture used to require. Agor AI Advisory works with operators who have decided to take that work seriously. We architect portable AI stacks that route across providers, expose tools through open standards, own their own memory, and refuse to sit at the mercy of a vendor's roadmap. The promos will keep coming. The companies we work with treat them as free margin.

The vendors have shown you what they think the wall costs. Build the architecture that makes the wall irrelevant.