OpenAI On Bedrock

OpenAI’s proprietary models are available on Bedrock, not just the open-weight gpt-oss family. This changes the default shape of every OpenAI → AWS migration: the source model itself is frequently a Bedrock target, so a cross-family swap to Claude/Nova is no longer the only option — and is no longer the default.

This file is the single source of truth for OpenAI-on-Bedrock facts in this plugin. ai-openai-to-bedrock.md (mapping policy), design-ai.md (selection), estimate-ai.md (costing), and ai-migration-guardrails.md (quota risk) all defer to it. Do not restate model IDs, regions, or endpoint paths elsewhere — link here.

Model Catalog

Naming: GPT-5.6 uses generation number + capability tier. Sol = flagship reasoning, Terra = balanced production, Luna = high-volume / low-latency. Tiers advance on independent cadences, so a future Terra may not share a generation with a future Sol.

Context-window conflict (resolved): the GPT-5.6 launch blog states 272K for all three variants; all three model cards state 1M. The model cards are authoritative — use 1M for GPT-5.6. GPT-5.5 and GPT-5.4 are 272K on both sources. Re-check on refresh; if AWS corrects the blog, the cards still win.

Not on Bedrock (as of this refresh): GPT-4o, GPT-4.1, GPT-4 / GPT-4 Turbo, GPT-3.5 Turbo, the o-series (o1/o3/o4-mini), GPT-5 / GPT-5.1 / GPT-5.2, and the *-Pro variants (GPT-5.5 Pro, GPT-5.4 Pro). Sources whose model is on this list have no same-model landing target — see ai-openai-to-bedrock.md for the two-option path.

Access Paths — Split by Family

GPT-5.5 and GPT-5.4 are bedrock-mantle-only and in-region only. Their model cards list a single Programmatic Access row (bedrock-mantle, Geo/Global “Not supported”) and In-Region pricing only.

GPT-5.6 Sol / Terra / Luna have TWO endpoints (verified 2026-08-21 — this changed after this file’s original 2026-08-10 verification; the Refresh Checklist predicted it):

The GPT-5.6 model cards now carry an explicit tip: “Whenever possible, we recommend using the bedrock-runtime endpoint for new applications.”

Constraints that still break naive assumptions:

1. The path segment is /openai/v1 on BOTH endpoints — a bare /v1 404s. Every GPT model card states the mantle path explicitly, and the runtime example gives bedrock-runtime.{region}.amazonaws.com/openai/v1.
2. The endpoints take different model-id forms. Mantle takes the bare openai.gpt-5.6-* id and has no CRIS; bedrock-runtime takes ONLY a CRIS inference-profile id (us. / in. / global. prefixed) and has no in-region form. bedrock:ListInferenceProfiles returns the 5.6 CRIS profiles; it never returns the bare mantle ids, and it returns nothing for GPT-5.5 / GPT-5.4.
3. API surfaces differ by endpoint. The 5.6 cards list Responses, Chat Completions, Invoke, and Converse as supported APIs, with runtime-side feature splits: Guardrails are Converse-only; prompt caching is Responses-only on runtime; server-side tool use, structured outputs, and application inference profiles are NOT supported on runtime (server-side tool calling IS supported on mantle). For GPT-5.5 / GPT-5.4 treat Responses on mantle as the only verified surface.

Client setup

Requires the OpenAI SDK at >= 2.45.0. Preferred client auto-refreshes a short-term Bedrock token:

The alternative — OpenAI(base_url=".../openai/v1", api_key=os.environ["AWS_BEARER_TOKEN_BEDROCK"]) — uses a key that expires within 12 hours and is not refreshed. Do not recommend it for production.

IAM: on the mantle path, the managed policy AmazonBedrockMantleInferenceAccess grants what inference needs, including bedrock-mantle:CreateInference and bedrock-mantle:CallWithBearerToken — bedrock:InvokeModel does not authorize mantle calls. On the GPT-5.6 bedrock-runtime path the usual bedrock:InvokeModel* against the CRIS inference-profile ARN applies, as for any other runtime model.

Reasoning effort: all five accept none, low, medium, high, xhigh, max. Because these models reason before responding, the model’s output items (which may include reasoning items) must be passed back in the next request for multi-turn and tool-calling flows.

Regional Availability — Endpoint-Aware

The mantle in-region matrix (the only reach for GPT-5.5 / GPT-5.4, and the in-region option for GPT-5.6):

Terra and Luna reached AWS GovCloud (US-West, US-East) in August 2026 — newer than the rest of this matrix.

GPT-5.6 additionally reaches most commercial regions via bedrock-runtime CRIS (Geo us. / in., Global global. inference profiles; the Sol card’s runtime footprint spans 30+ regions). So a region outside the mantle matrix does NOT make the same-model path unavailable for GPT-5.6 — it means using the runtime endpoint with a CRIS id, with the data-residency implications of cross-region routing. For GPT-5.5 / GPT-5.4 the mantle matrix is a hard gate: no CRIS, no fallback.

Verify current footprints per model card / aws___get_regional_availability (AWS MCP Server) — the CRIS lists move faster than this file.

Pricing

Read off the model cards, 2026-08-31. All rates per 1M tokens, Standard tier (Priority and Flex are NOT supported for these models). Sol rates reflect the Aug 21, 2026 reduction (−20% input / −33% output vs launch rates), which the AWS What’s New announcement lists as promotional through at least Nov 21, 2026. Pricing now has an inference-option dimension:

• In-Region and Geo CRIS: 1.10x OpenAI’s standard list price (parity with OpenAI’s data residency tier).
• Global CRIS: OpenAI’s standard list price — cost parity, available for GPT-5.6 only, and only when the workload has no data-residency constraint.

So the honest cost statement is conditional, not flat: a same-model GPT-5.6 move on Global CRIS is cost-neutral; the same move in-region or Geo (and any GPT-5.5 / GPT-5.4 move) is ~10% more expensive. Never state either number without stating the inference option it belongs to.

GPT-5.6 — short context (272K)

GPT-5.6 — long context (1M): 2.0x input / 1.5x output of short-context, per option

A workload above 272K context must be priced at the long-context tier.

GPT-5.5 / GPT-5.4 — In-Region only (no CRIS, no long-context tier; usable window 272K)

The Luna “blog discrepancy” resolved differently than first recorded. The AWS News Blog’s 0.20 / 1.20 is not an error — it is the Global CRIS rate, now published on the Luna card. An earlier revision of this file said global pricing was unpublished and treated the blog figure as wrong; both statements are corrected here. Separately, the AWS Price List API still carries no GPT-5.x rows (checked 2026-08-04): a missing or empty price-list result must not be read as “model unavailable.”

Prompt caching — GPT-5.6 only

Listed as a supported feature on the Sol, Terra, and Luna model cards. The GPT-5.5 and GPT-5.4 cards list client-side tool calling in that slot instead and do not list prompt caching. Do not assume caching on 5.5/5.4.

Explicit mode uses prompt_cache_options={"mode": "explicit"} plus a prompt_cache_breakpoint on the content block ending the reusable prefix; a stable prompt_cache_key improves match reliability. Cached input tokens do not count against the input-TPM quota, which compounds the benefit at scale.

Quotas

Inference on bedrock-mantle is governed by two per-model, per-region quotas: input tokens per minute and output tokens per minute. There is no requests-per-minute quota. Exceeding a TPM quota returns HTTP 429.

This corrects two claims that were previously applied to all Mantle traffic in this plugin:

• There is no shared 10,000 RPM account limit governing these models — the quota dimension is TPM, per model, per region.
• “Switch to bedrock-runtime“ is not a throughput remedy for GPT-5.5 / GPT-5.4 — they have no bedrock-runtime path at all. For GPT-5.6 the runtime path DOES exist (CRIS only; see Access Paths above), so moving there is a legitimate option — but treat it as an endpoint/architecture choice with its own quota family and residency implications, not a free throughput escape hatch.

The supported mitigations are: exponential backoff with a bounded retry count (max_retries on the OpenAI SDK), spreading load across minutes rather than bursting, ramping request rate gradually, and prompt caching (cached input is exempt from the input-TPM quota). For sustained volume beyond that, pursue a quota increase.

The model cards now state tier support in text: pricing shown is Standard, and Priority and Flex are not supported for these models. Do not recommend Flex as a cost lever for any GPT model here; Reserved is account-level via the AWS account team.

Features With No Bedrock Equivalent

These are the remaining legitimate reasons to keep a workload on OpenAI’s own API. Cost is no longer one of them.

Data handling: these are third-party models under OpenAI terms. Classifier-flagged traffic is retained up to 30 days for automated abuse detection; retained inputs/outputs are stored and processed by AWS and not shared with OpenAI unless the customer opts in. Prompts and completions are not used to train models. Calls run under the customer’s IAM policies, inside their VPC, logged to CloudTrail, and in-region inference keeps data in-region.

Codex on Bedrock is GA with pay-per-token pricing, inference through Bedrock, and usage counting toward AWS commitments — relevant when the source workload is a coding agent.

Refresh Checklist

This model family is moving fast (two GA waves and a repricing inside 10 weeks). On each refresh:

1. Re-read the OpenAI model card index (opens in a new tab) for models added or removed, and each per-model card for lifecycle state and EOL date.
2. Recheck the region matrix AND the GPT-5.6 CRIS footprints — for 5.5/5.4 a region change is migration-blocking; for 5.6 the runtime/CRIS path usually covers the region instead.
3. Recheck rates on the Bedrock pricing page OpenAI tab, and resolve any row still marked unverified.
4. Recheck whether the Price List API has gained GPT-5.x coverage; if it has, drop the caveat above and let estimate-ai.md price these models from the MCP.
5. Recheck whether Chat Completions and bedrock-runtime support have been added or clarified.
6. Feed any lifecycle change into ai-model-lifecycle.md and any rate change into pricing-cache.md.
7. Re-verify within 14 days of any merge touching this file. Item 5’s prediction fired on 2026-08-21: between 2026-08-10 and 2026-08-21 the GPT-5.6 family gained a bedrock-runtime/CRIS path, published Global CRIS pricing at standard-price parity, and listed Chat Completions/Converse as supported — invalidating three of this file's then-central claims in under two weeks. This family moves faster than a normal refresh cadence.