OpenAI's models on Amazon Bedrock are no longer a preview experiment. Between April 28 and October 26, 2026, AWS turned them into a documented product line: the GPT-5.x and GPT-5.6 families, the GPT-6 flagship, a GPT-6.1 successor, scoped cybersecurity variants, and a server-side web-search tool.[1][2] The enterprise question has moved with it. It is no longer whether OpenAI models run inside Bedrock, but which endpoint, which API path and which inference option a given workload should standardize on. AWS's model cards record the inputs for that decision: launch dates, supported endpoints and APIs and, for the flagship, token prices.

On April 28, AWS announced a Limited Preview that bundled three products at once: OpenAI's frontier models, the Codex coding agent, and Managed Agents, all reached through Bedrock's unified API.[1][3] AWS's developer-community write-up of that preview names GPT-5.4 and GPT-5.5 as the callable models, lists the controls that carry over (IAM permissions, PrivateLink network isolation, CloudTrail logging) and records that usage counts toward existing AWS spend commitments, the Enterprise Discount Program and Private Pricing Agreement (EDP/PPA).[3] GPT-5.5 reached general availability on Bedrock on June 1, 2026.[4] In July, the GPT-5.6 family (Sol, Terra and Luna) became generally available through the Responses API,[5] and in August the same three models reached the bedrock-runtime endpoint through Responses, Converse and Chat Completions, with Global and Geo cross-Region inference profiles (CRIS), including a new US Geo profile, added for higher throughput.[5]

August also carried the scoped additions. Daybreak Red (GPT-5.6 Cyber), purpose-trained for vulnerability research, exploit reproduction and mitigation development, and Daybreak Blue (GPT-5.6 Sol), with safeguards calibrated for defensive cybersecurity work, are limited-availability models for eligible customers.[5] The same month added a Web Search tool that grounds OpenAI GPT models through a single Responses-API parameter against an AWS-hosted index, in US East (N. Virginia), US East (Ohio) and US West (Oregon).[5] GPT-6 Astra followed on September 8,[6] and GPT-6.1 Terra on October 26.[2]

The catalog as of November 1, 2026

Model On Bedrock since Notes
GPT-5.4 April 28, 2026 (Limited Preview) Named in AWS's preview write-up[3]
GPT-5.5 June 1, 2026 Per its card: 1,050,000-token context shared between input and output; 128,000 max output tokens[4]
GPT-5.6 Sol July 13, 2026 Per its card: same 1,050,000 / 128,000 envelope; Responses, Chat Completions and Converse supported on bedrock-runtime[7]
GPT-5.6 Terra, GPT-5.6 Luna July 2026 (family general availability) Announced alongside Sol through the Responses API[5]
Daybreak Red (GPT-5.6 Cyber), Daybreak Blue (GPT-5.6 Sol) August 2026 Limited availability, eligible customers; cybersecurity-specialized[5]
GPT-6 Astra September 8, 2026 OpenAI's most capable model, per its card; 1,050,000 context / 128,000 output; knowledge cutoff April 30, 2026[6]
GPT-6.1 Terra October 26, 2026 Per its card: 1,050,000 / 128,000; positioned for detailed visual reasoning and accurate computer use[2]

AWS's GPT-6.1 Terra card also names GPT-6.1 Sol and GPT-6 Astra as its comparison points; only Terra's Bedrock launch date is card-documented.[2] The benchmark claims on that card — 15 percentage points over GPT-5.6 Terra on Agents' Last Exam, on par with Astra, and more than three times faster than Astra and GPT-6.1 Sol — are OpenAI's claims relayed by AWS, not independently verified here.[2]

Context matters for the routing question: OpenAI's models sit in a Bedrock catalog that also carries Anthropic's Claude 5.x family (Opus 5.5, Sonnet 5.5 and Haiku 5.5 among them), xAI's Grok 4.7 and Z.AI's GLM 5.3.[8] Within a given API surface, model choice is a model-ID change, which is what "unified API" means in practice.

Two endpoints, one set of model IDs

Bedrock exposes two inference endpoints: bedrock-runtime.{region}.amazonaws.com, the recommended surface, and bedrock-mantle.{region}.api.aws. Both run on the same underlying Mantle engine; "mantle" names an endpoint surface, not a different engine. Both accept an existing OpenAI SDK codebase by changing only the base URL and API key, and on runtime the OpenAI-compatible APIs live on /openai/v1 paths.[9]

The API split follows the endpoint. InvokeModel and Converse are runtime-only; Chat Completions and Responses run on both. The Messages API is an endpoint capability for Anthropic-native callers, but not for the OpenAI models whose cards spell out API support: Astra's card marks Messages unsupported on both endpoints, and the GPT-5.6 Sol card marks it unsupported on runtime.[6][7][9]

Model IDs encode the routing decision. On mantle, GPT-6 Astra is openai.gpt-6-astra, served from us-east-1 and us-west-2 (Standard tier only in Oregon). On runtime there is no in-region option for this model: the card states it "is not available for in-Region inference on that endpoint," so callers name us.openai.gpt-6-astra for US geographic cross-Region inference or global.openai.gpt-6-astra for global. The Global profile reaches 18 commercial regions, Frankfurt to Tokyo to São Paulo; Geo profiles serve US and Canadian source regions.[6]

Where the endpoints differ: what does not carry over

  • Server-side tool use and pre-configured tools, including Web Search, exist only on mantle; runtime requests cannot invoke them.[9]
  • Long-running work: on runtime the Responses API is always synchronous, and background=true is rejected with a 400 error. Asynchronous inference is mantle-only.[9]
  • Guardrails is a runtime-only feature in the endpoint matrix, and Astra's card narrows it further: Converse API only.[6][9]
  • Cost attribution: on Astra, application inference profiles work with InvokeModel and Converse only, not with Responses or Chat Completions; on runtime's Responses API, usage attributes by IAM principal, and a request naming an application inference profile is rejected with a 400 error. Cross-Region inference profiles are unaffected.[6][9]
  • Projects run with the default project only on runtime; Workspaces are mantle-only; a stored response on runtime belongs to the region that served it.[9]

Prompt caching and structured outputs are the conveniences that do carry over. Astra supports implicit caching across all four runtime APIs and explicit caching everywhere except Converse; JSON-schema structured output works on runtime through Chat Completions, Responses and Converse (Converse additionally requires additionalModelRequestFields.text.format.strict=true). One content-type catch: Converse accepts text and image blocks in user messages but rejects document blocks — documents go through the Responses API as input_file parts.[6]

The practical reading: agentic workloads that need server-side tools or background execution route to mantle; compliance-filtered workloads that need Guardrails route to runtime. Both endpoints can be called from the same application, per use case.[9]

What AWS's cards disclose on price

Per-token pricing for the same model is identical on the two endpoints, so the price question is about inference options, not endpoint choice.[9] The figures below are as disclosed on GPT-6 Astra's card:

  • Standard, short context (272,000 input tokens or fewer): In-Region and Geo CRIS cost $11.00 input / $55.00 output per 1 million tokens; Global CRIS $10.00 / $50.00. Long context (above 272,000 input tokens): $22.00 / $82.50 In-Region and Geo; $20.00 / $75.00 Global.[6]
  • The premium is explicit: "Commercial In-Region and Geo CRIS prices include a 10% premium over the corresponding OpenAI rates for the same service tier." Global CRIS is the route without the 10% premium where data-residency rules allow.[6]
  • Cache economics: a 30-minute prompt-cache write costs $13.75 per 1 million tokens (short-context In-Region) and a cache read $1.10.[6]
  • Ultrafast, the speed tier, costs six times Standard — $66.00 / $330.00 short-context In-Region — and OpenAI reports API speeds up to six times faster than Standard, a vendor claim. It is requested with service_tier="ultrafast" on mantle in us-east-1 or through the CRIS routes on runtime. Priority and Flex tiers are not supported for this model.[6]

Within the GPT-6.1 generation, Terra's card positions it as the balanced option, with input and output token prices at least 50% below GPT-6.1 Sol's and GPT-5.6 Terra's — again, the vendor's claim on the card.[2] The commitment detail predates the cards: routing OpenAI usage through AWS can land it inside an existing enterprise agreement rather than beside it.[3]

Controls, quotas and the watch list

The controls carry over: IAM governs model access, PrivateLink keeps traffic inside the AWS network and away from NAT-gateway egress charges, CloudTrail records calls, and Guardrails rules apply to OpenAI models.[3][9] Quotas differ by endpoint. Runtime runs fixed RPM/TPM quotas that can be raised; with Astra, one output token consumes ten quota tokens (a 10x burndown rate). Mantle uses scheduling, with higher initial limits and occasional brief queueing.[6][9]

Two things stay on the watch list. Codex (CLI, desktop and VS Code, authenticating with AWS credentials) and Managed Agents (an OpenAI agent harness running on Bedrock AgentCore, with an IAM role per agent and action logging) are documented at their April Limited Preview state in AWS's announcement materials, with preview limits (us-east-1 and us-west-2 only, no custom tools in Managed Agents, TPM below general-availability levels) that are date-stamped to April, not descriptions of today.[3] The catalog table above is bounded by what AWS has documented in cards and announcements: the snapshot stands as of November 1, 2026, and the choices it supports (endpoint, API path, inference option) are the parts AWS documents most completely.