What hit GA on June 1
On June 1, 2026, AWS shipped general availability for GPT-5.5, GPT-5.4, and Codex on Amazon Bedrock. The April limited-preview moved to GA in eight weeks — a faster cadence than the analyst community had projected, and a deliberate signal that both companies wanted the multi-cloud question settled before AWS re:Invent gets close enough to dominate the news cycle.
The operationally important details:
- Pricing is at parity with direct OpenAI API rates. AWS is not charging a Bedrock-routing markup. Per-token costs match what you'd pay calling OpenAI directly, which removes the but our finance team won't approve a cloud markup on top of vendor cost objection from the procurement conversation.
- The call path is the Bedrock Responses API. Existing OpenAI SDK code requires the standard Bedrock adaptation — endpoint, auth, request envelope — but the model behavior and the output structure carry over. The integration tax is a few days of glue code per service, not a re-architecture.
- Region availability is initial-tier. GPT-5.5 launched in US East (Ohio); GPT-5.4 in US East (Ohio) and US West (Oregon). Additional regions are scheduled to follow, but production teams routing through Bedrock should plan around those two regions for the next quarter at minimum.
- Codex moves to pay-per-token billing on Bedrock. The per-seat licensing model that has historically governed Codex deployment is gone for the Bedrock path; consumption is metered like every other Bedrock model, on tokens.
- The AWS governance layer applies natively. IAM policies, KMS encryption at rest, CloudTrail audit logging, VPC endpoint connectivity — the security and compliance primitives that the rest of your AWS data infrastructure runs on now apply to OpenAI model calls without a separate vendor onboarding.
That last bullet is the under-told part of the story. The compliance burden of routing AI traffic to a non-AWS vendor — we need a separate DPA, separate audit attestations, separate IR coordination, separate SOC 2 mapping, separate procurement cycle — was the structural friction that kept many regulated-industry buyers on Azure or running models direct-to-Anthropic. Bedrock collapses that friction for the OpenAI line, because the same IAM role, the same KMS key, the same CloudTrail audit pipeline that already audits your S3, your Redshift, your DynamoDB, now audits your GPT-5.5 calls.
The 30-month architecture assumption that just collapsed
For essentially the entire post-ChatGPT-launch era — from late 2022 through May 2026 — the architectural assumption baked into every enterprise AI strategy was a vendor-to-cloud mapping: GPT-class capability lived on Azure; Claude-class capability on AWS (or direct API, or — for the last twelve months — on Anthropic's own platform); Gemini-class capability on Google Cloud. The mapping was rarely written down explicitly, because it didn't need to be. The procurement team's existing cloud-vendor relationship effectively settled the AI-vendor decision, because the friction of routing AI traffic to a different cloud than where the data lived was prohibitive on compliance, on networking, and on operational tooling.
This was not great for buyers. It meant the cloud-vendor decision — made years before frontier AI capability mattered, in a strategic context dominated by storage cost and compute reservations — was also the AI-vendor decision, and re-opening one meant re-opening the other. The honest read is that the friction was a feature for the cloud vendors: it converted cloud lock-in into AI lock-in at zero marginal cost.
June 1 collapses the mapping for the largest of the four major frontier vendors on the second-largest of the three hyperscalers. The implication is not AWS is now an OpenAI shop; the implication is the cloud-vendor decision and the AI-vendor decision are now independent for the first time. A Fortune 500 enterprise sitting on AWS, with three years of data-platform investment in S3 + Redshift + Glue + Lake Formation, can now run OpenAI as the primary AI vendor without rebuilding their data plane on Azure. The same is true in the other direction for an Azure-anchored enterprise that wants Claude — Anthropic shipped Claude on AWS GA in early May and has since broadened its multi-cloud surface — and for the smaller GCP-anchored cohort.
The procurement decision that was effectively settled by the cloud-vendor relationship is now genuinely open. Most enterprise AI strategy documents drafted in the last twelve months need a refresh, and the conversation that's been deferred — which frontier vendor actually fits our workload best, independent of where our data already lives? — just became the conversation worth having.
What multi-cloud frontier vendor availability actually changes
Four decisions that change shape when the cloud-vendor and AI-vendor decisions decouple.
The vendor-evaluation conversation gets honest. When the AI vendor was effectively dictated by the cloud vendor, the evaluation was performative — a checkbox exercise to justify a decision that was already made. With both vendors genuinely available on the same data plane, the evaluation has to actually pick, and the procurement team that doesn't run a serious eval against the actual workload will pick on vendor-relationship affinity rather than on capability fit. That's worse than the constrained-by-cloud era, because the affinity decision is harder to revisit than the cloud-imposed constraint was.
Multi-model routing becomes a first-class procurement requirement, not an engineering preference. When OpenAI was Azure-exclusive and Claude was AWS-strong, the multi-model routing engineering layer was a nice-to-have that some teams built and most teams skipped. With both vendors available on the same data plane, routing across them at workload granularity becomes the operational pattern that justifies the procurement decision — we put Claude on hard reasoning, GPT on long-horizon agent work, Codex on the in-IDE coding load, and the cost model rolls up across all three on the same Bedrock invoice. The teams that already built a portability layer have an easier path. The teams that didn't are about to discover that their single-vendor AI procurement was load-bearing in their architecture in ways they didn't appreciate.
The negotiation leverage shifts toward the buyer. A vendor whose availability was constrained to one cloud could anchor pricing on that cloud's customer base. A vendor available on every cloud has to win every customer against the other vendors available on the same cloud, and the negotiating posture moves accordingly. The enterprise buyer with a credible threat to route 30% of their AI spend to a different vendor on the same infrastructure has structural leverage in the next contract cycle that they did not have before June 1. The buyers who plan their procurement strategy around that leverage will get materially better contracts in Q3 than the buyers who assume nothing changed.
The exit conversation gets cheaper. A future scenario where one vendor's pricing, capability, or contractual posture becomes structurally worse than the alternatives is a scenario where the exit cost dominates the decision-making. When the exit required re-platforming to a different cloud, the exit was prohibitively expensive and the vendor knew it. When the exit is change the routing config and the IAM policy, the exit cost is real but not prohibitive. The vendor that knows the exit is cheap behaves materially differently than the vendor that knows the exit is expensive. That structural change applies to every multi-cloud frontier-vendor relationship from June 1 forward.
What this does not change
Three honest caveats.
It does not eliminate the vendor lock-in surface; it relocates it. A workload built tightly against OpenAI's API specifics, prompt patterns, and tool-calling conventions is still a workload that needs to be ported when you swap to a different model. Multi-cloud availability decouples the cloud-vendor and AI-vendor decisions; it does not eliminate the AI-vendor lock-in surface. The portability layer above the API — the prompt adapter, the tool-call translator, the evaluation harness that grades each vendor on the same gold set — is still the engineering investment that converts theoretical portability into operational portability.
It does not change the model-quality decision. The right model for a workload is still determined by how well it performs on that workload, not which cloud it's hosted on. Multi-cloud availability widens the option set; it does not make the option set easier to evaluate. The eval discipline that grades models honestly on your workload is more important after June 1, not less.
It does not eliminate the operational lock-in that comes with vendor-specific tooling. Bedrock's Responses API has its own conventions; Azure OpenAI's API has its own conventions; OpenAI's direct API has its own conventions. A workload built against Bedrock-specific request envelopes is a workload that is operationally tied to Bedrock even when the underlying model is portable. The teams that want the optionality the multi-cloud availability theoretically grants need to consciously design for portability — adapter layers, request normalization, response normalization — rather than assume it falls out for free.
Where Sonnet Code fits
Multi-cloud frontier vendor availability is the easy half of the story. The hard half is the engineering above the procurement decision — the portability adapter that treats Bedrock-hosted OpenAI as one vendor among several on a common API surface, the eval harness that grades each vendor on your actual workload, the multi-model routing layer that splits the load by workload class with budget guardrails, the IAM and audit posture that captures every frontier-AI call into your existing compliance pipeline — that turns the June 1 GA announcement into a Q3 procurement advantage. AI development at Sonnet Code is that engineering: designing the portability and routing layer that lets your team treat OpenAI-on-Bedrock, Claude-on-AWS, and direct-vendor APIs as a unified procurement surface, extending the IAM and audit configuration so every model call is captured in CloudTrail with the structured metadata your security team needs, and building the cost-per-successful-task dashboard that surfaces which vendor is winning on which workload class in your stack. AI training is the human-judgment half: senior engineers, domain experts, and procurement specialists who run the honest vendor-vs-vendor eval on your workload, calibrate the routing rubrics that distinguish Claude-best from GPT-best from cheaper-tier-acceptable at the granularity production needs, and stand up the contract-renewal data pack that turns the new multi-cloud leverage into a real cost win at the next negotiation cycle.
The Azure-exclusivity era of OpenAI ended at midnight on June 1. The cloud-vendor-decides-the-AI-vendor era of enterprise AI architecture ended with it. The teams that walk into Q3 with the procurement strategy refreshed, the portability layer extended, and the multi-model routing layer in production are the teams that will compound the new vendor leverage into real budget room. The teams that treat the GA announcement as a footnote will spend the same leverage back to the vendors in renewal cycles they could have shaped differently.
