Subscribe
Share
Share
Embed
Twenty-four months of operational record across the five major foundation-model providers. AM-136 documents Anthropic, OpenAI, Google, AWS Bedrock, and Azure OpenAI. Every provider had at least one multi-hour outage that exceeded its own SLA-credit thresh
The audio companion to agentmodeai.com. Two analysts pick one claim from the Holding-up ledger per episode, walk the evidence, and give the current verdict: Holding, Partial, or Not holding. For CIOs, IT directors, and senior implementers. 15-20 min, every Sunday.
[00:30] Why status pages do not tell procurement what they need to know
ABBY: A vendor status page reports the operational state of the API gateway and the regional infrastructure surrounding the model. It does not report the operational state of the specific model the customer integrated against. A customer running production traffic through Claude 3.7 Sonnet during a Sonnet-specific degradation can be told by the status page that the API is operational while their production agent is returning errors or degraded responses. This is not a vendor failure. The status page reports what it reports. The procurement gap is that the customer's operational metric is not the gateway's uptime; it is the specific model's uptime against the specific prompts the customer's workload sends.
AVERY: So the customer's own observability is the leading indicator.
[01:45] The customer's own observability is the leading indicator
ABBY: The customer's own observability is the leading indicator. Per-model latency, per-model error rate, per-model degradation patterns, all separately from the gateway's published metrics. The agent observability comparison at AM-123, Langfuse, Arize, Helicone, LangSmith, walks the four-platform decision. The relevant procurement signal there is that the observability platform must capture per-model signals, not just gateway signals. A customer that relies on the vendor's status page as its operational signal has no leading indicator of model-specific issues until production traffic itself reveals them.
AVERY: The SLA-credit gap.
[02:45] The SLA-credit gap and the ninety-nine-point-nine arithmetic
ABBY: The standard published SLA across major foundation-model providers commits to ninety-nine-point-nine percent monthly availability with credits capped at twenty-five to fifty percent of monthly fees on the affected service. The arithmetic on ninety-nine-point-nine is roughly forty-three minutes of allowed monthly downtime. The arithmetic on the credit cap, for an enterprise paying an indicative fifty thousand dollars per month against a single provider, is a maximum of twenty-five thousand returned in the worst month. A fraction of the operational impact a multi-hour outage produces on a customer-facing workload. Those figures are illustrative, not surveyed.
AVERY: What do enterprise workloads actually need.
[04:00] What enterprise workloads actually need: ninety-nine-point-nine-five plus hard-dollar liability
ABBY: Customer-facing agentic deployments typically need ninety-nine-point-nine-five availability with hard-dollar incident liability above the SLA-credit cap, not service credits as the only remedy. The point-zero-five gap is roughly twenty-two minutes per month of additional downtime tolerance. Material on a workflow where every minute of downtime translates to a measurable customer impact. Negotiating up from ninety-nine-point-nine to ninety-nine-point-nine-five is contractually possible at the Enterprise tier of every major provider. The negotiation requires three procurement instruments: a documented business impact analysis showing per-minute cost of downtime, a named incident-severity-tier framework with response-time commitments, and an explicit hard-dollar liability ceiling that exceeds the SLA-credit cap.
AVERY: Walk the five providers briefly.
[05:00] Walking the five providers
[05:15] Anthropic and OpenAI — incident pattern and credential-class failures
ABBY: Anthropic publishes incidents at status.anthropic.com. The operational pattern across 2024-2026 is a small number of multi-hour incidents per year, typically affecting a specific model version or region, with detailed postmortems published within seven to fourteen days. OpenAI publishes at status.openai.com. The operational pattern includes the broader incident class of credential-system or quota-system failures that affect the entire API surface, alongside model-specific incidents. Credential-system incidents are particularly procurement-relevant because they cannot be mitigated by switching models within the same provider.
AVERY: Google.
[06:15] Google Gemini — GCP regional fault-domain inheritance
ABBY: Google Gemini API publishes at status.cloud.google.com within the broader Google Cloud Platform status surface. The operational pattern inherits the GCP regional fault-domain structure. Gemini-specific incidents are reported but require the customer to filter the broader status feed.
AVERY: The hyperscalers.
[06:45] AWS Bedrock and Azure OpenAI — regional failover plus the OpenAI-side propagation
ABBY: AWS Bedrock publishes at the AWS Service Health Dashboard. Operational pattern inherits AWS regional infrastructure plus Bedrock-specific model availability across regions. Bedrock customers benefit from cross-region inference patterns that mitigate single-region failures within the provider but do not mitigate Bedrock-wide incidents. Azure OpenAI publishes at the Azure Status Dashboard. Operational pattern inherits Azure regional infrastructure plus the Azure-OpenAI integration layer. Azure OpenAI customers can deploy across multiple Azure regions for failover; that pattern mitigates regional incidents but does not mitigate the OpenAI-side issues that propagate through the Azure integration.
AVERY: The procurement-relevant read across all five.
[07:45] No clean choice for single-provider deployment
ABBY: No provider is a clean choice for single-provider production deployment in 2026. The architecture has to assume failure across the provider graph regardless of which provider is chosen. The relative ranking depends on which workload class the customer cares about. Hyperscaler-backed offerings inherit the underlying cloud's regional fault domains plus model-specific failure modes. Frontier-lab APIs have a shorter operational history but more direct upgrade-path optionality on model versions. Both classes have material incidents in the twenty-four-month window.
AVERY: Three architectural patterns for multi-provider routing.
[08:30] Three multi-provider routing patterns
[08:45] Pattern 1 — gateway abstraction (LiteLLM, OpenRouter, Portkey)
ABBY: Pattern one. Gateway abstraction. A gateway sits between the application and the foundation-model providers and routes traffic based on configurable rules. LiteLLM is the open-source reference implementation. OpenRouter and Portkey are commercial alternatives. The gateway provides a unified API surface that the application integrates against, with the underlying provider selection happening at routing time. Failover is observable, configurable per route, and testable in production. The procurement-defensible benefit is that the customer owns the routing logic and is not dependent on the vendor for failover capability.
AVERY: The cost of pattern one.
ABBY: The gateway becomes part of the customer's operational substrate. Its own observability requirements, its own latency overhead, typically fifty to one hundred and fifty milliseconds added per request, and its own change-management discipline. Enterprises running this pattern typically deploy the gateway in their own VPC with redundancy across availability zones, treating it as production-critical infrastructure equivalent to an API gateway or service mesh.
AVERY: Pattern two.
[10:00] Pattern 2 — provider-side regional failover
ABBY: Provider-side regional failover. Within a single vendor, the customer deploys across multiple regions and handles failover at the regional level. AWS Bedrock cross-region inference and Azure OpenAI multi-region deployment are the two production-grade implementations. Partial mitigation. It addresses regional failures within the vendor's fault domain but does not address vendor-wide incidents. A Bedrock-wide outage affects all regional deployments simultaneously. Appropriate for workloads where vendor-wide incidents are tolerable but regional incidents are not.
AVERY: Pattern three.
[10:45] Pattern 3 — explicit multi-provider provisioning at the application layer
ABBY: Explicit multi-provider provisioning at the application layer. The application is built to support two or more model families with prompt-tested compatibility maintained as a deployment requirement. The customer maintains active provisioning with a primary and secondary provider; failover is application-driven. Most expensive to maintain because every prompt change requires testing across the supported providers. Most resilient because no shared failure mode with the gateway or any single vendor's regional architecture. Appropriate for high-availability customer-facing workloads where the per-incident cost exceeds the maintenance overhead of multi-provider support.
AVERY: The default for 2026 enterprise procurement.
[11:30] The procurement default for 2026
ABBY: Pattern one, gateway abstraction, is the procurement default for most enterprise workloads in 2026. Pattern three is the higher-rigour option for the workloads that justify it. Pattern two is a partial mitigation appropriate for narrower workload classes than most enterprises initially assume.
AVERY: Three contract language additions for the 2026 AI MSA red-team checklist.
[12:00] Three contract language additions for the AI MSA red-team
[12:30] Hard-dollar incident liability above the SLA-credit cap
ABBY: First, hard-dollar incident liability above the SLA-credit cap. Named incident severity tiers, typically Sev-1 customer-impacting, Sev-2 degraded service, Sev-3 minor impact, Sev-4 informational, with per-tier response-time commitments and an explicit hard-dollar liability ceiling that exceeds the SLA-credit cap by a multiplier reflecting the customer's per-minute downtime cost. Achievable at the Enterprise tier of every major provider.
AVERY: Second.
[12:45] Non-degradation clauses on model-deprecation events
ABBY: Non-degradation clauses covering model-deprecation events. The vendor's right to deprecate or update models is the customer's procurement risk. The 2026 standard is a contractually defined transition window, typically ninety days for major version changes, thirty days for minor, during which the customer's prior model version remains available alongside the new version. The clause is required because forced cutover with no transition window produces a regression-test sprint the customer's engineering organisation cannot consistently absorb on the vendor's schedule.
AVERY: Third.
[13:00] Right to multi-provider routing without contract penalty
ABBY: Right to multi-provider routing without contract penalty. Some 2024-vintage MSAs included vendor-exclusivity clauses that prohibited the customer from routing identical traffic through a competitor's model. Those clauses do not survive 2026 procurement diligence because they make pattern one and pattern three contractually infeasible. The procurement-defensible language explicitly preserves the customer's right to deploy multi-provider architectures and prohibits the vendor from using identical-traffic-routing as a basis for contract enforcement action.
AVERY: What does the AM-136 verdict look like.
[13:30] Verdict and cadence on AM-136
ABBY: AM-136 is Holding on a thirty-day cadence. The reliability record is the kind of evidence that ages monthly, not quarterly. The procurement-defensible posture has to refresh against the most recent operational data the providers publish. Three triggers would shift the analysis. A foundation-model provider publishing a sustained ninety-nine-point-nine-nine operational record across twelve consecutive months would close the gap between the standard ninety-nine-point-nine and the enterprise-required ninety-nine-point-nine-five. A regulatory development requiring multi-provider provisioning for high-risk AI deployments under the EU AI Act would shift the procurement question from architectural choice to compliance requirement. A landmark vendor outage producing material customer harm and follow-on litigation would shift the SLA-credit-versus-hard-dollar-liability negotiation.
AVERY: The honest read on what to do Monday.
[13:15] What to do Monday: three concrete actions
ABBY: Three actions. First, audit the current foundation-model architecture against the three patterns. If the deployment is single-provider with no routing layer, prioritise pattern one in the next quarterly engineering plan. Second, audit the current vendor MSA against the three contract additions. The renewal window before 2 August 2026 is the natural moment to negotiate them. Third, instrument the customer's own observability stack against per-model signals, not just gateway signals. The lead time on detecting model-specific degradation is the difference between a bounded incident and a customer-impact incident.
AVERY: Final word.
[13:40] Outtro
ABBY: AM-136 and the five vendor status pages are linked at agentmodeai dot com slash holding. The Sunday brief ships every week with what moved on the ledger.
AVERY: Holding-up. See you next Sunday.
Claims tracked in this episode
AM-136 — Foundation-model uptime SLA track record
https://agentmodeai.com/holding/?claim=AM-136
Related claims
AM-026 — The 60-question agentic AI RFP
https://agentmodeai.com/holding/?claim=AM-026
AM-138 — Vendor MSA renewal post-EU-AI-Act-enforcement
https://agentmodeai.com/holding/?claim=AM-138
AM-123 — Agent observability comparison
https://agentmodeai.com/holding/?claim=AM-123
Primary sources
Anthropic status — https://status.anthropic.com/
OpenAI status — https://status.openai.com/
Google Cloud status (Gemini API) — https://status.cloud.google.com/
AWS Service Health Dashboard — https://health.aws.amazon.com/
Azure status — https://status.azure.com/
Multi-provider routing reference implementations
LiteLLM — https://www.litellm.ai/
OpenRouter — https://openrouter.ai/
Portkey — https://portkey.ai/
Subscribe
Newsletter and the full Holding-up ledger — https://agentmodeai.com
Editorial charter — https://agentmodeai.com/standards/
How it's written — https://agentmodeai.com/how-its-written/
Written by Claude · Curated and signed by Peter