Bedrock sells GPT-5.6 Sol at exactly OpenAI's list price and also at 1.10x it. Which one you pay comes down to a single table row, and getting approved for OpenAI's cyber program deletes that row from your card.
AWS publishes five Bedrock model cards for GPT-5.6, and only three of them have a cheap row. The ordinary Sol card carries three routing options, and the cheapest, Global cross-region inference, bills $5.00 and $30.00 per million tokens, which is OpenAI's own list price to the cent on all four lines. The other two bill $5.50 and $33.00. The two gated cards, Daybreak Blue and Daybreak Red, have no Global row at all, so the same Sol weights cost 10% more once you have been approved to use them for security work. None of this appears on the AWS pricing page, which names seven OpenAI frontier models and publishes numbers for none of them. And the largest gap here is one nobody is charging for: OpenAI sells Sol through Batch and Flex at half price, Bedrock supports neither on any GPT-5.6 model, so the floor on one side of the fence is exactly twice the floor on the other.
Photo by Robin Pierre on Unsplash
Five cards, and only three of them have a cheap row
Start with the thing that makes this measurable at all. Of the Bedrock model cards we opened, the only ones printing per-token prices inline were OpenAI's. If you want to count them yourself, they are all reachable from AWS's model card index, which is also where the five GPT-5.6 entries are listed together. That is why this comparison exists and why it does not exist for anything else on the platform, which is a point I will come back to at the end because it bounds how far any of this generalises.
Regular GPT-5.6 Sol landed on Bedrock on July 13, a month before the security models. Its card carries three inference options, and the prices split two against one.
| GPT-5.6 Sol, per 1M | Input | Cache write | Cache read | Output |
|---|---|---|---|---|
| OpenAI direct, Batch or Flex | $2.50 | $3.125 | $0.25 | $15.00 |
| OpenAI direct, Standard | $5.00 | $6.25 | $0.50 | $30.00 |
| Bedrock, Global CRIS | $5.00 | $6.25 | $0.50 | $30.00 |
| Bedrock, Geo CRIS | $5.50 | $6.875 | $0.55 | $33.00 |
| Bedrock, In-Region | $5.50 | $6.875 | $0.55 | $33.00 |
Geo and In-Region are the same number, which tells you the premium is not really about how far your tokens travel. Global is the outlier. That matters for how you describe it: it is more accurate to say Global is 9.09% cheaper than the other two than to say the other two carry a 10% surcharge, even though the arithmetic is identical either way. AWS built one discounted row and two ordinary ones, not two penalised rows and one ordinary one.
Two things about that table are less symmetrical than they look, and both cut against the cheaper rows. Sol's In-Region row is not reachable on the ordinary bedrock-runtime endpoint at all: the card's availability table marks In-Region unsupported in all 32 regions and says so in prose, so the only way to buy it is bedrock-mantle in us-east-1 or us-east-2. And the only Geo profile AWS publishes is us.openai.gpt-5.6-sol, which means a team in the EU or APAC with a residency requirement cannot buy the Geo row either. For most customers outside the United States the three-row table is really a one-row table, and the row is Global.
The same shape repeats on the two cheaper tiers of the family. Terra bills $2.00 direct and $2.20 In-Region. Luna bills $0.20 and $0.22. Sol's long-context tier does it too: $10.00 and $45.00 on Global, $11.00 and $49.50 on the other two. Worth noting that the 2x-input and 1.5x-output long-context step those figures encode appears only inside AWS's tables. OpenAI's pricing page publishes Standard, Batch, Flex and Fast for Sol and no long-context tier at all, so the parity claim in this post is a claim about the 272K tier. Above it there is no OpenAI list price to be at parity with.
1.1000, with no rounding left over
It is worth being precise about how exact this is, because a multiplier that holds to four decimal places on every line is a deliberate pricing device rather than a coincidence of two independently set cards. GPT-5.6 Cyber shows it most cleanly, because Cyber has exactly one Bedrock row and nothing to average against.
| GPT-5.6 Cyber, per 1M | OpenAI direct | Bedrock In-Region | Ratio |
|---|---|---|---|
| Input | $12.50 | $13.75 | 1.1000 |
| Cache write, 30m | $15.625 | $17.1875 | 1.1000 |
| Cache read | $1.25 | $1.375 | 1.1000 |
| Output | $75.00 | $82.50 | 1.1000 |
The $17.1875 is the tell. AWS could have rounded that to $17.19 and nobody would have complained about four hundredths of a cent per million tokens. Publishing the unrounded figure means the card was generated by multiplying OpenAI's number rather than negotiated line by line.
There is a practical consequence that runs opposite to most pricing changes we cover. When xAI raised only the cached-input line on Grok 4.6, your exposure depended on your cache hit rate, and the teams who had optimised hardest were hit hardest. Here the multiplier is flat across input, output, cache reads and cache writes, so caching buys you nothing in relief. A Sol workload with 90% of its input served from cache costs $278.00 direct and $305.80 In-Region on the numbers below, which is the same 1.10 as an uncached one. Whatever you do to the shape of your traffic, the premium follows it.
The approval that costs 10%
OpenAI announced Daybreak Red and Daybreak Blue on Monday, August 10. AWS blogged about it on the 11th, and both gated Bedrock model cards carry a launch date of the 12th. Blue and Red are access tiers rather than distinct models. AWS describes Blue as GPT-5.6 Sol with safeguards calibrated for defensive cybersecurity work, and reserves the lower-refusal-threshold language for Red, which is the separate GPT-5.6 Cyber model. Both cards carry word-for-word identical Trusted Access wording, so nothing in AWS's documentation says Red is a higher rung of approval than Blue.
Which means AWS now publishes two model cards for what is, on the weights, GPT-5.6 Sol. Here is what each card contains, with GPT-5.5 added at the bottom for a reason that will be obvious in a moment.
| Bedrock card | Endpoint | Global row | Cheapest input | Gated |
|---|---|---|---|---|
| GPT-5.6 Sol | bedrock-runtime | Yes | $5.00 | No |
| GPT-5.6 Terra | bedrock-runtime | Yes | $2.00 | No |
| Daybreak Blue: GPT-5.6 Sol | bedrock-mantle | No | $5.50 | Yes |
| Daybreak Red: GPT-5.6 Cyber | bedrock-mantle | No | $13.75 | Yes |
| GPT-5.5 | bedrock-mantle | No | $5.50 | No |
Read the Daybreak Blue row again. It is Sol. It bills $5.50 and $33.00 short-context and $11.00 and $49.50 long-context, which is line for line the In-Region column of the ordinary Sol card, and it has no Geo or Global row to fall back to. A team that has never spoken to OpenAI about cyber eligibility can call Sol through Global CRIS for $5.00 and $30.00. A team that went through identity verification and legal attestations calls the same weights for 10% more.
My first instinct was that this is a data-residency side effect: gated models are pinned to Ohio, Global routes across regions by definition, so the cheap row cannot coexist with the promise that makes the gating coherent. That explanation is wrong, and the bottom row of the table is what kills it. GPT-5.5 is not gated by anyone. It has no Geo row and no Global row either, and its In-Region price is $5.50, the identical figure. GPT-5.4 behaves the same way.
What actually predicts the cheap row is the endpoint. Across all seven priced OpenAI cards, the three served on bedrock-runtime, Sol, Terra and Luna, carry Geo and Global profiles. The four reachable only through bedrock-mantle, Daybreak Blue, Daybreak Red, GPT-5.5 and GPT-5.4, carry neither, gated or not. Cross-region inference is a bedrock-runtime feature, and the discount rides on it. So the honest version of this finding is narrower than the headline and more useful: approval does cost you 10%, but not because AWS is charging for approval. It costs 10% because approval moves you onto an endpoint where the discounted routing profile has never existed, and that endpoint also costs you Converse, Guardrails, invocation logging and response streaming. The invoice line is the cheapest thing you give up.
The bigger gap is the one nobody is charging for
Ten percent is the number that shows up when you put two cards side by side, so it is the number everyone will quote. It is not the largest gap on this page. AWS's batch inference support table lists four OpenAI entries and all four are open weight: gpt-oss-20b, gpt-oss-120b, and the two GPT OSS Safeguard variants. No GPT-5.6 model appears on it. Provisioned throughput lists no OpenAI models at all, and every GPT-5.6 card footnotes that its prices are Standard tier and that Priority and Flex are not supported either.
OpenAI, meanwhile, sells Sol through both Batch and Flex at $2.50 and $15.00, half the standard card on every line. So the cheapest Sol token you can buy on Bedrock costs twice the cheapest Sol token you can buy from OpenAI. Here is a month of moderate agent traffic, 40 million input tokens and 8 million output tokens, priced every way it can be bought.
| Route | 40M input | 8M output | Month |
|---|---|---|---|
| OpenAI, Batch or Flex | $100.00 | $120.00 | $220.00 |
| OpenAI, Standard | $200.00 | $240.00 | $440.00 |
| Bedrock, Global CRIS | $200.00 | $240.00 | $440.00 |
| Bedrock, In-Region or Geo | $220.00 | $264.00 | $484.00 |
| Bedrock, Daybreak Blue | $220.00 | $264.00 | $484.00 |
The spread between the top row and the bottom two is $264.00 a month, which is to say the dearest way to buy this workload costs more than twice the cheapest, on traffic small enough that most teams would not bother modelling it. The routing decision everyone is arguing about is worth $44.00 of that. The tier decision nobody mentions is worth the rest. If your traffic is genuinely interactive then Batch was never available to you and the comparison collapses back to $440.00 against $484.00. If it is not, and a great deal of agent traffic is not, then the question of whether to buy through AWS is a much more expensive question than the 10% suggests.
One genuine offset I cannot price for you: Bedrock spend counts toward an AWS enterprise discount commitment, and a team with a negotiated EDP is buying at some percentage off its whole bill. That can close the 10% and occasionally more. It cannot close a 2x, and it is a discount on your AWS total rather than a cut to the token rate, so it belongs in a different column of the spreadsheet from everything above.
Parity arrived, but bolted to a routing profile
When OpenAI models first reached Bedrock in April, AWS had published no frontier prices and we wrote that the reasonable expectation was price parity, on the grounds that AWS took its margin through compute rather than a per-token markup. Parity did arrive, and it is exact rather than approximate, which is more than that post claimed. It is also conditional in a way nobody could have guessed from the April evidence: it lives on one routing profile, on one endpoint, and it was never offered on the gated cards or on GPT-5.5 and GPT-5.4 at all. The April post also assumed the interesting question was the per-token rate. It turned out to be which service tiers and routing profiles exist on each side, and on that it was silent.
AWS and OpenAI do not agree on how big Cyber is
OpenAI's model page for GPT-5.6 Cyber states a 400,000 token context window with 128,000 max output tokens. AWS's Bedrock card states 272K, and puts its one pricing table under a heading reading Short Context Window (272K). Both are primary sources from the companies selling the thing, and they disagree by 128,000 tokens.
My read is that AWS printed a billing boundary in a capability field. 272,000 is precisely where the rest of the GPT-5.6 family switches to long-context pricing, at 2x input and 1.5x output for the whole request. On Sol's Bedrock card that same heading sits above a short-context table with a matching Long Context Window (1M) table underneath it. On Cyber's card there is no second table. So either Cyber really does stop at 272K, or it runs to 400K and AWS has published no rate for the top 128,000 tokens of it. I would not build against the upper number until one of the two pages moves.
For what it is worth, Cyber is a clean 2.5x Sol on every line OpenAI publishes: $12.50 against $5.00 on input, $75.00 against $30.00 on output, $1.25 against $0.50 on cache reads. It sells on the Standard tier only, with no Fast mode, no Flex and no Batch on either side of the fence, which is the one thing both vendors agree about.
Why you cannot run this check on Claude
The obvious next question is whether every model on Bedrock carries a 10% In-Region premium. The honest answer is that there is no way to find out from AWS's documentation, and I want to be explicit about that rather than let a tidy number travel further than the evidence.
We opened 38 of the roughly 123 Bedrock model cards. Seven contain a pricing table, and all seven are OpenAI. Every Anthropic, Amazon Nova, Meta Llama, DeepSeek, Mistral, Qwen, Google, xAI, Z.AI, MiniMax, Moonshot and NVIDIA card we opened has a Pricing section consisting of one sentence pointing at the general Bedrock pricing page. That page, in turn, names seven OpenAI frontier models and publishes per-token figures for none of them, only for the OSS models. So the model cards are the sole AWS-owned source for every number in this post. There is no second AWS page confirming them.
What that supports is a narrow claim: 1.1000 is exact and uniform across all five GPT-5.6 cards, and on both context tiers wherever a second tier exists. What it does not support is calling that a Bedrock convention, or calling it unique to OpenAI. AWS clearly does use clean decimal multipliers as a pricing device, because GPT-5.4's card carries a GovCloud table at $3.30 against $2.75 In-Region, an exact 1.2000 on each of the three lines it prints. Whether Claude or Nova carry anything similar is unknowable from the outside, and any article that tells you otherwise is extrapolating.
Pick the endpoint before you pick the vendor
Check whether your workload can run as a batch job first, because that decision is worth twice what the vendor decision is worth. If it can, buy from OpenAI and stop reading. If it cannot, and you want the tokens on your AWS bill for procurement, compliance or EDP reasons, use ordinary GPT-5.6 Sol through Global CRIS and you will pay exactly what OpenAI charges. Only accept the 1.10 if something concrete forces you onto a single region, and price that constraint honestly rather than assuming Bedrock is at parity because one row of one card is.
And if you are being onboarded to Daybreak Blue, ask whether you need it. It is Sol with safeguards recalibrated for defensive work, on a reduced-feature endpoint, at 10% over the rate you can already get, and for a lot of security work ordinary Sol with your own prompt scaffolding will do the job on the cheap row. Current rates for every model named here sit on the pricing page, and you can run your own token split through the calculator.
Seven cards and one silent pricing page
- AWS: Bedrock model card, GPT-5.6 Sol - The three-row table behind every Sol figure here. Global CRIS $5.00 input, $6.25 cache write, $0.50 cache read, $30.00 output; In-Region and Geo CRIS identical to each other at $5.50, $6.875, $0.55 and $33.00. The Long Context Window (1M) table at $10.00 and $45.00 on Global against $11.00 and $49.50 on the other two. Also the July 13, 2026 launch date, and the availability table behind the region facts above: Global enabled in 32 regions, Geo published only as us.openai.gpt-5.6-sol, and In-Region marked unsupported on bedrock-runtime everywhere, leaving us-east-1 and us-east-2 on bedrock-mantle as the only way to buy that row
- AWS: Bedrock model card, Daybreak Red GPT-5.6 Cyber - One pricing row, In-Region only, at $13.75, $17.1875, $1.375 and $82.50. Geo and Global both explicitly marked not supported in the Programmatic Access table. Model ID openai.gpt-5.6-cyber, served on the bedrock-mantle endpoint rather than bedrock-runtime, us-east-2 only, launch date August 12, 2026, Standard tier only with Priority, Flex and Reserved all unsupported. This is also the page that says 272K, and the page that carries no long-context table
- AWS: Bedrock model card, Daybreak Blue GPT-5.6 Sol - The card that makes this post. Model ID openai.gpt-daybreak-blue-5.6-sol, launched August 12, 2026, us-east-2 only, 1M context, priced at exactly regular Sol's In-Region row and carrying no Geo or Global alternative, in either of its two context tiers
- OpenAI: API pricing - Every direct rate above. Sol at $5.00, $6.25, $0.50 and $30.00 Standard, $10.00 and $60.00 Fast, and Batch and Flex identical to each other at $2.50, $3.125, $0.25 and $15.00. Cyber at $12.50, $15.625, $1.25 and $75.00 with no Batch, Flex or Fast rows at all. Also the note that Priority Processing was renamed Fast mode on July 30, 2026, so anything still describing a Priority tier is out of date
- AWS: Amazon Bedrock pricing - Named here for what it does not contain. All seven OpenAI frontier models are listed by name with no per-token figures, under an instruction to select a model to view detailed pricing. Token prices on this page exist only for gpt-oss-20b, gpt-oss-120b and the GPT OSS Safeguard variants. That is why nothing in this post is corroborated by a second AWS source
- AWS: models supported for batch inference - The four OpenAI entries, all open weight, and the absence of every GPT-5.6 model. Paired with the provisioned throughput support list, which contains no OpenAI models at all, this is the basis for the claim that Bedrock's floor is 2x OpenAI's
- AWS: Daybreak Red and Daybreak Blue on Bedrock - The August 11 announcement, and the Trusted Access wording repeated on both gated cards: eligibility runs through OpenAI first and then a second access request through your AWS account team
- TokenCost: OpenAI on AWS Bedrock, April 30 - Our own parity prediction, made when AWS had published no frontier prices, and the EDP point that still stands
- Three things this post does not know. OpenAI's own Daybreak announcement page and its help-centre overview both returned 403 to us, so the August 10 announcement date and the framing of Blue and Red as access tiers come from AWS's card and from secondary press rather than from OpenAI directly. We have not seen an invoice for any of these rates, only published cards. And the Cyber context window is genuinely contested between two primary sources; we have printed both numbers rather than picking the one that reads better