4 GovCloud GPT-5.6 doc conflicts AWS shipped on 24 August 2026

Summary. On 24 August 2026 AWS announced that OpenAI's GPT-5.6 Terra and Luna are "generally available on Amazon Bedrock in AWS GovCloud (US-West) and AWS GovCloud (US-East)". Three other AWS pages disagree. The Bedrock regional availability page, which that same announcement tells you to consult, carries no GovCloud row for Terra, Luna or Sol; the newest OpenAI model with a GovCloud row is GPT 5.4, in us-gov-west-1 only. The AWS FedRAMP and DoD CC SRG certification-status table for Bedrock models, last updated 28 July 2026, also stops at GPT 5.4. And the AWS GovCloud (US) User Guide still names just 7 authorised models, none of them from OpenAI, two months after AWS announced on 25 June 2026 that OpenAI GPT models had cleared FedRAMP High and IL 4/5. For a federal or defence programme, that gap is the difference between a model you can run and a model you can put in an authorisation boundary. The pricing gap is real too: AWS calls Terra "GPT-5.5-level performance at half the cost", while OpenAI lists Terra at $2.00 per million input tokens against GPT-5.5 at $5.00, and files Terra under the mini tier.

What AWS actually shipped on 24 August

The AWS What's New post is dated 24 August 2026 and is unambiguous about availability. Terra and Luna are GA in both GovCloud regions, served through Bedrock's Mantle inference engine on the bedrock-mantle endpoint via the Responses API. The post claims 1 million token context windows, prompt caching with explicit breakpoints at "a 90% discount", and positions Terra as GPT-5.5-class at half the price.

Every one of those four claims is checkable against another AWS or OpenAI page. Three of them do not survive the check in the form the announcement states them.

Conflict 1: the page the announcement links to has no GovCloud row

The announcement closes with "For regional availability, please see the Amazon Bedrock regional availability page." That page is Regional availability by models in the Bedrock User Guide, and it publishes a per-model table with In-Region, Geo and Global columns for every region Bedrock serves.

Under the OpenAI section, GPT-5.6 Sol, GPT-5.6 Terra and GPT-5.6 Luna each list commercial regions only. Neither us-gov-west-1 nor us-gov-east-1 appears in any of the three tables. The OpenAI models that do carry GovCloud rows are GPT 5.4, gpt-oss-120b and gpt-oss-20b.

That is 6 rows, and the top three are the ones the 24 August post says are generally available.

Conflict 2: the FedRAMP table stops at GPT 5.4

AWS maintains a single authoritative page for this: Amazon Bedrock models - FedRAMP and DoD CSP SRG (IL4/IL5) certification status. It carries a visible "Last updated: July 28, 2026" stamp, and its columns are now labelled "U.S. E/W (FedRAMP Class C, formerly Moderate)" and "U.S. GovCloud (FedRAMP Class D, formerly High)", reflecting the FedRAMP impact-level rename.

Under the OpenAI model family, the table lists one entry: GPT 5.4. Not GPT-5.5, not any GPT-5.6 tier. Because the page predates the 24 August launch by 27 days, it could not include Terra or Luna even if authorisation had been granted the same day.

This matters more than a stale doc usually would. FedRAMP and DoD CC SRG authorisation is asserted per service and, for Bedrock, per model. An assessor does not accept a What's New post as evidence. The certification-status page is the artefact.

Conflict 3: the GovCloud User Guide is two months behind AWS's own announcement

The Amazon Bedrock in AWS GovCloud (US) page has a section headed "How Amazon Bedrock differs" that lists the models holding FedRAMP and IL4/5 authorisation. As fetched on 24 August 2026 it named all Titan models, Claude Sonnet 4.5, Claude 3.7 Sonnet, Claude 3.5 Sonnet v1, Claude 3 Haiku, Llama 3 8B and Llama 3 70B. That is 7 entries, and no OpenAI model appears among them.

AWS itself contradicted that list on 25 June 2026, when it announced that "OpenAI GPT, OpenAI GPT OSS, and NVIDIA Nemotron models are now FedRAMP High and Department of Defense Cloud Computing Security Requirements Guide (DoD CC SRG) Impact Level (IL) 4 and 5 approved within Amazon Bedrock in the AWS GovCloud (US) Regions." The GovCloud User Guide has been wrong for two months, in the direction that understates what is authorised. The certification-status page is wrong in the direction that understates what shipped. They are not wrong in the same direction, which is why reading only one of them produces a confident and incorrect answer. This is the second AWS docs conflict on Bedrock third-party models in as many months; the earlier one hit Cost Anomaly Detection coverage of third-party models.

Conflict 4: the caching claim and the cache-write fee

The announcement says repeated context "is billed at a 90% discount and doesn't compound cost as you scale". Bedrock's own prompt caching documentation is more careful, and states plainly that "tokens written to cache may be charged at a rate that is higher than that of uncached input tokens."

For GPT-5.6 specifically, the same page gives the number: cache writes are billed at 1.25x the uncached input token rate, cache reads at a 90% discount, with a 1,024-token minimum prefix per breakpoint, a maximum of 4 cache checkpoints per request, and a 30-minute TTL. OpenAI's prompt caching guide matches those figures and adds the behaviour that turns them into a bill: on GPT-5.6 and later the service does not automatically fall back to the longest matching unmarked prefix before a breakpoint, so a small change ahead of your breakpoint produces a miss and a fresh write.

We have run this arithmetic before on the Bedrock versus OpenAI direct cost math, and the shape repeats. Do the arithmetic on Terra at its $2.00 per million input rate. Write once and read the prefix even once and caching already wins. Write and never read, because the prefix churns, and you pay $2.50 per million input tokens, 25% more than not caching at all. OpenAI's own prompt caching dashboard exposes the cache-write line, which is the fastest way to see whether your prefixes are earning reads. Bedrock's page adds one more hazard for cross-region inference: "At times of high demand, these optimizations may lead to increased cache writes."

There is a migration trap buried here. Bedrock's page notes that for OpenAI models prior to GPT-5.6, such as openai.gpt-5.5 and openai.gpt-5.4, caching is automatic and "cache writes have no additional fee on these models." GPT 5.4 is the OpenAI model GovCloud is currently documented to support. Moving from it to Terra introduces a cache-write charge that did not previously exist on that workload.

The pricing and context claims, checked against OpenAI

Two of the announcement's softer claims also move once you open the model cards.

That table has 5 rows. The last one carries the largest number. A single 922,000-token Terra request bills input at $4.00 per million, not $2.00, so $3.69 rather than $1.84 for the same prompt, and output at 1.5x for the whole request. Any team sizing a GovCloud budget from the headline rate and the "1 million token context" line will be out by roughly 2x on long-context calls.

For reference, OpenAI's published list prices per million tokens as of 24 August 2026 are Sol at $4.00 input and $20.00 output, Terra at $2.00 and $12.00, and Luna at $0.20 and $1.20. Sol's figure is promotional: the OpenAI API changelog entry dated 21 August 2026 records the cut from $5.00 and $30.00, a 20% input and 33% output reduction, "available at least through November 21, 2026". A GovCloud budget built on Sol's current rate has a known expiry date.

How to tell if this affects you

You are exposed if any of the following is true. Your programme requires FedRAMP Class D (formerly High) or DoD CC SRG IL 4/5 evidence per model. Your ATO package cites a model by name. Your architecture assumes In-Region inference in GovCloud rather than Geo routing. Or your cost model was built from the announcement's headline token rate.

You are probably fine if you are running GPT 5.4, gpt-oss-120b or gpt-oss-20b in us-gov-west-1, all of which appear on the regional availability page with GovCloud rows today.

One operational detail catches teams either way. Bedrock's regional availability page documents that enabling a foundation model in GovCloud starts outside GovCloud: you must accept the model's EULA in a standard region, us-east-1 or us-west-2, through the linked standard account, then enable the model on the Model Access page in the GovCloud account. The CLI path runs aws bedrock list-foundation-models, then aws bedrock list-foundation-model-agreement-offers --model-id <model-id> for the offer token, then aws bedrock create-foundation-model-agreement. A GovCloud-only team without a linked commercial account cannot complete step one.

What to do this week

Treat the certification-status page as the only compliance artefact, and screenshot it with its "Last updated" stamp on the day you rely on it. Ask your AWS account team for the model-level authorisation status of openai.gpt-5.6-terra and openai.gpt-5.6-luna in writing, and ask specifically whether the 25 June 2026 OpenAI GPT approval covers the 5.6 family or only the models named in the July table. Confirm through the console or aws bedrock list-foundation-models in your own GovCloud account rather than from a docs page, because entitlement propagation and regional availability are separate from authorisation. Re-run your token cost model with the >272K multiplier applied and the 1.25x cache-write rate included, using the same discipline our cloud FinOps guide applies to committed spend. Then check whether your prompt prefixes are stable enough to earn cache reads at all; if they are not, explicit-only breakpoint mode avoids paying to write a prefix nobody will read.

If your workload touches Indian data as well as US federal data, the same discipline applies to residency: our note on data residency and DPDP cloud architecture covers the equivalent gap between where a model runs and where its data is processed.

India-specific considerations

GovCloud is a US construct and Indian teams cannot use it, but two things here travel. First, the pattern: a vendor announcement, a linked availability page and a compliance page that disagree is not an AWS-only failure, and the same three-source check applies to any regulated deployment under the Digital Personal Data Protection Act 2023. Second, the cost mechanics are region-agnostic. Terra's Geo routing covers ap-south-1 (Mumbai) and ap-south-2 (Hyderabad) on the regional availability page, so Indian teams get the same 1.25x cache-write rate and the same >272K long-context multiplier without any of the GovCloud authorisation questions. If you are sizing an agent workload from Mumbai, the cache-write arithmetic above is the part that will move your bill.

What is still unknown

AWS has not published a model-level statement on whether GPT-5.6 Terra and Luna are inside the GovCloud FedRAMP Class D and IL 4/5 boundary as of 24 August 2026. The 25 June 2026 post says "OpenAI GPT" without naming versions, which can be read either way. Nor has AWS published GovCloud token pricing for these two models in the announcement; the figures above are OpenAI's own list prices, and Bedrock rates in GovCloud have historically differed from commercial regions. Until the certification-status page moves past its 28 July 2026 stamp, the honest position is that availability is documented in one place and authorisation in another, and the two do not currently agree.

FAQ

Are GPT-5.6 Terra and Luna FedRAMP authorised in AWS GovCloud?

AWS has not confirmed it at model level. The certification-status page for Bedrock models, last updated 28 July 2026, lists only GPT 5.4 under OpenAI. The 25 June 2026 announcement approved "OpenAI GPT" models generically without naming versions, so treat authorisation as unconfirmed until that table is refreshed.

Which OpenAI models does the Bedrock regional availability page show in GovCloud?

Three: GPT 5.4 in us-gov-west-1 with In-Region inference only, plus gpt-oss-120b and gpt-oss-20b, which each show us-gov-west-1 with In-Region and Geo routing and us-gov-east-1 with Geo only. No GPT-5.6 tier, including Sol, Terra or Luna, has a GovCloud row on that page.

What does prompt caching actually cost on GPT-5.6 in Bedrock?

Cache reads are billed at a 90% discount to the uncached input rate, but cache writes cost 1.25x that rate. On Terra at $2.00 per million input tokens, a written prefix costs $2.50 per million. Minimum prefix is 1,024 tokens, maximum 4 checkpoints per request, TTL 30 minutes.

Does the 1 million token context window really hold?

The context window is 1,050,000 tokens on Terra and Luna, but maximum input is 922,000 tokens and maximum output is 128,000. Prompts above 272,000 input tokens are billed at 2x input and 1.5x output for the entire request, which the 24 August announcement does not mention anywhere.

How much cheaper is Terra than GPT-5.5?

AWS says half the cost. OpenAI's published list prices put Terra at $2.00 input and $12.00 output per million tokens against GPT-5.5 at $5.00 and $30.00, which is a 60% reduction rather than 50%. OpenAI's own model card places Terra in the mini tier, not the GPT-5.5 tier.

Why can't I enable the model from inside my GovCloud account?

Bedrock requires the model's End User License Agreement to be accepted in a standard commercial region, either us-east-1 or us-west-2, through the linked standard AWS account. Only then can you enable it on the Model Access page in GovCloud. Entitlements take a few minutes to propagate after that step.

Is GPT-5.6 Sol's current price permanent?

No. The OpenAI API changelog entry dated 21 August 2026 records Sol moving to $4.00 input and $20.00 output per million tokens, down 20% and 33% respectively, and states the promotional pricing is available at least through 21 November 2026. Budgets past that date should assume the rate can revert.

Does any of this affect commercial AWS regions?

The authorisation conflict does not, because FedRAMP Class D applies to GovCloud only. The pricing mechanics do. Cache-write charges at 1.25x, the 1,024-token minimum prefix and the 2x multiplier above 272,000 input tokens apply in every commercial region where Terra and Luna are offered, including Mumbai and Hyderabad via Geo routing.

How eCorpIT can help

eCorpIT builds and reviews cloud AI architectures where the compliance boundary, not the benchmark, decides the model. Our senior engineering teams read the certification artefacts alongside the release notes, rebuild the token cost model with the cache-write and long-context multipliers applied, and document what an assessor will actually accept. eCorpIT is CMMI Level 5, MSME Certified and ISO 27001:2022 certified, and we design applications aligned with FedRAMP and DPDP requirements. Book a Bedrock cost and compliance review if you are sizing a regulated model deployment.

References

1. OpenAI GPT-5.6 Terra and Luna now available on Amazon Bedrock in AWS GovCloud (US), AWS What's New, 24 August 2026
2. Amazon Bedrock models - FedRAMP and DoD CSP SRG (IL4/IL5) certification status, AWS, last updated 28 July 2026
3. Amazon Bedrock in AWS GovCloud (US), AWS GovCloud (US) User Guide
4. OpenAI GPT, OpenAI GPT OSS, and NVIDIA Nemotron models on Amazon Bedrock receive FedRAMP High and DoD IL-4/5 approval in AWS GovCloud (US), AWS What's New, 25 June 2026
5. Regional availability by models, Amazon Bedrock User Guide
6. Prompt caching for faster model inference, Amazon Bedrock User Guide
7. Model support by AWS Region in Amazon Bedrock, Amazon Bedrock User Guide
8. Pricing, OpenAI API documentation
9. GPT-5.6 Terra model card, OpenAI API documentation
10. GPT-5.6 Luna model card, OpenAI API documentation
11. GPT-5.6 Sol model card, OpenAI API documentation
12. Prompt caching guide, OpenAI API documentation
13. Changelog, OpenAI API documentation, entry dated 21 August 2026

Last updated: 24 August 2026.