Bedrock Managed Agents vs Calling Model APIs Directly: The Price Check (2026)
AWS runs OpenAI's agent loop in Bedrock. The token bill is +10%, the infra bill is yours. When that trade pays off, with worked math.
In September 2026, AWS and OpenAI shipped something that changes the "which cloud runs my agent" question: Amazon Bedrock Managed Agents (BMA), powered by OpenAI — OpenAI's agent harness running inside AWS, with every model call executing on Amazon Bedrock and the customer's data never leaving AWS. The headline sounds like a new billable service. It is not, yet. The real cost comparison is subtler: Bedrock's per-token rates versus OpenAI's own, plus whatever runtime you feed the agent, plus a GA fee AWS has not published.
What was actually announced (verified)
First, the fact-check, because "Bedrock Managed Agents" is easy to misread:
- Announced as a limited preview on April 28, 2026, as part of the expanded AWS–OpenAI partnership, alongside OpenAI models and Codex on Bedrock (AWS What's New, OpenAI).
- Opened to public preview on September 29, 2026, in three US Regions (us-east-1, us-east-2, us-west-2), API-only on the
bedrock-mantleendpoint (AWS What's New). - It runs OpenAI models — not Anthropic or Amazon models. It is built on a customized version of OpenAI's Agents API, hosts the agent loop in Bedrock, and executes your tools on either self-hosted compute or Amazon Bedrock AgentCore Runtime (Bedrock docs). Non-OpenAI models stay on plain AgentCore.
- Preview pricing, verbatim from AWS: "During preview, there is no additional charge for BMA beyond the underlying AWS resources your agents consume. Pricing is subject to change at general availability." No GA date or GA fee has been published as of November 7, 2026.
So the correct comparison is exactly what the title says: BMA (Bedrock-hosted OpenAI agent loop) versus calling OpenAI's model APIs and Agents API directly. The same +10% in-region pattern also holds for Anthropic models on Bedrock, so the shape of the math below applies to a Claude-based variant too.
TL;DR
- Tokens: Bedrock = OpenAI + 10%. GPT-5.4 is $2.50/$15.00 per 1M in/out direct, $2.75/$16.50 in-region on Bedrock; GPT-5.5 is $5.00/$30.00 direct, $5.50/$33.00 on Bedrock. The model cards state the in-region price already includes the 10% premium.
- The 10% delta mostly evaporates if you were going to buy data residency anyway. Direct OpenAI regional-processing endpoints carry a 10% uplift for models released after March 5, 2026 — the same fee BMA charges by default.
- BMA's service fee is $0 in preview, undefined at GA. Anthropic's competing harness charges $0.08/running session-hour — a plausible analog, but unverified for BMA.
- The runtime is your bill, not AWS's. AgentCore Runtime microVMs bill per-second CPU and memory (v2: $0.1276/vCPU-hr, $0.0169/GB-hr); AWS's own example stack adds a NAT gateway that costs roughly $33/month while idle.
- Batch is the real price lever, and it's currently direct-only for these models. OpenAI batch pricing halves GPT-5.4/5.5 token costs; BMA's preview supports no Batch/Flex tiers and no cross-Region inference profiles.
- Rough scenario: 10 agents × 100 calls/day of GPT-5.4-class tokens lands around $2,025/month direct vs ~$2,462/month on Bedrock + AgentCore — about +21% all-in, of which only 10 points are the token premium.
Per-million-token price: one premium, stated on the card
Cited from the GPT-5.5 and GPT-5.4 Bedrock model cards (which note: "Commercial In-Region prices include a 10% fee over OpenAI rates. You do not need to add this fee") and OpenAI's pricing page. Short context = ≤272K input tokens; long-context rates apply to all tokens, not just the overflow.
| Model (short ctx) | Direct in | Direct cached in | Direct out | Bedrock in | Bedrock cached in | Bedrock out | Delta |
|---|---|---|---|---|---|---|---|
| GPT-5.4 | $2.50 | $0.25 | $15.00 | $2.75 | $0.275 | $16.50 | +10% |
| GPT-5.5 | $5.00 | $0.50 | $30.00 | $5.50 | $0.55 | $33.00 | +10% |
| GPT-6.1 Terra | $1.00 | $0.05 | $5.00 | $1.10 | $0.055 | $5.50 | +10% |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 | $2.20 | $0.11 | $11.00 | +10% |
Notes that matter for the bill:
- Residency changes the comparison. OpenAI charges a 10% uplift on regional (data-residency) processing endpoints for models released on or after March 5, 2026. BMA's inference is in-region by construction (three US Regions only in preview). If data staying in a specific geography is a requirement, the direct API's apples-to-apples price is roughly the Bedrock price. Losing that requirement makes direct API the cheaper of the two for pure tokens.
- Batch is a 50% discount on OpenAI direct (GPT-5.4 batch: $1.25/$7.50; GPT-5.5 batch: $2.50/$15.00). Bedrock lists Batch/Flex tiers at 50% off Standard in its pricing UI, but both GPT-5.5 and GPT-5.4 model cards state "Priority and Flex tiers are not supported for this model" — batch support for the OpenAI frontier models on Bedrock is [unverified].
- Direct API also has a cheaper tail BMA can't match today:
gpt-5.4-mini($0.75/$4.50) andgpt-5.4-nano($0.20/$1.25) are not in the Bedrock OpenAI catalog; if you want to route traffic down-tier, you'd usegpt-oss-20b/120bon Bedrock ($0.0721/$0.3090 and $0.1545/$0.6180 in us-east-1-style in-region pricing) instead. - Capacity: OpenAI's developer platform is effectively reservationless; Bedrock still offers its classic on-demand vs provisioned (capacity-reservation) split for other model families, and usage of OpenAI models and Codex on Bedrock counts toward existing AWS commitments (RIP/EDP). For enterprises with negotiated AWS spend, that discount can be worth more than the 10% list premium — the impact is deal-specific and [unverified].
The bill beyond tokens: who runs the agent loop, and what it costs
This is where the hosted-harness comparison gets concrete.
| Cost line | BMA | OpenAI Agents API (direct) | Anthropic Managed Agents (direct) |
|---|---|---|---|
| Agent loop service fee | $0 during preview; GA pricing unannounced — treat any post-GA fee as [unverified] | $0 | $0.08/running session-hour |
| Model tokens | Bedrock in-region rates (+10%) | OpenAI list rates | Claude list rates (US-pinned inference = 1.1x) |
| Execution environment (tools) | You own it: self-hosted host, or AgentCore Runtime | OpenAI-hosted sandbox, self-hosted sandbox, or none | Included in session runtime |
| Idle infrastructure risk | Yes — the AgentCore sample stack provisions a NAT gateway that bills continuously | None (unless you self-host) | None |
Sources: AWS docs "Charges" section, Amazon Bedrock AgentCore pricing, Anthropic Managed Agents pricing.
The AgentCore line items, verbatim from AWS's pricing table:
- Runtime microVM v2 consumption: $0.1276/vCPU-hour and $0.0169/GB-hour (v1: $0.0895 / $0.00945), billed per second with a 1-second minimum. CPU billing pauses during I/O wait; v2 reclaims idle memory after 120 seconds. Committed baselines were listed at $0.0997/$0.0132 (v2).
- The sample-stack booby trap: AWS's own BMA documentation warns that "a deployed stack can continue to incur charges when no BMA turn is running" because the AgentCore example provisions a NAT gateway — about $0.045/hour ≈ $32.85/month at us-east-1 rates, before a single task runs.
- AWS's published worked example (1M sessions/month, 10 minutes each, 90% I/O wait) prices out at $0.0067 per session, or $6,703/month for a million — runtime only, before tokens.
If you pick the self-hosted execution environment instead, the "infra" line is whatever your existing build server costs — BMA just needs its codex exec-server process, and it makes an outbound connection to the Bedrock endpoint. That is the cheapest BMA configuration and often the right one for internal engineering agents.
Worked scenario: 10 agents × 100 calls/day
This is a model scenario built from cited unit prices — not benchmarks and not a real measured workload. Assume: 10 agents, 100 agent-turns each per day = 1,000 turns/day ≈ 30,000 turns/month. Each turn consumes 15,000 input tokens and 2,000 output tokens (a document/tool-heavy turn shaped like the code review and workflow tasks BMA is aimed at). Month = 30 days.
Monthly tokens: input 15,000 x 30,000 = 450,000,000 (450M)
output 2,000 x 30,000 = 60,000,000 ( 60M)
GPT-5.4, short context, uncached:
Direct: 450M x $2.50 = $1,125.00 + 60M x $15.00 = $ 900.00 -> $2,025.00/mo
Bedrock: 450M x $2.75 = $1,237.50 + 60M x $16.50 = $ 990.00 -> $2,227.50/mo
Token premium: $202.50/mo (~10%), ~$2,430/yr
With realistic caching — say 80% of input tokens hit a cache read on a stateful, multi-turn agent:
Direct: 90M x $2.50 = $225.00 + 360M x $0.25 = $90.00 + $900.00 -> $1,215.00/mo
Bedrock: 90M x $2.75 = $247.50 + 360M x $0.275 = $99.00 + $990.00 -> $1,336.50/mo
Now the infra lines for the Bedrock-only option. Per-turn runtime modeled on AWS's example session shape (10 minutes, 90% I/O wait, 1 vCPU, 1–2.5 GB, v2 consumption rates): ~$0.0067/turn × 30,000 = ~$201/month. Add one NAT gateway if you use the sample stack: ~$32.85/month. Self-hosted execution takes both lines to zero (you already own the box).
| Scenario (30,000 turns/mo) | Tokens | Runtime | Total/mo |
|---|---|---|---|
| Direct OpenAI API + your own agent loop | $2,025 | $0 | $2,025 |
| Direct + 80% cache hits | $1,215 | $0 | $1,215 |
| BMA (Bedrock tokens, no cache) | $2,227.50 | $233.85 | ~$2,461 |
| BMA + 80% cache hits | $1,336.50 | $233.85 | ~$1,570 |
| [unverified] If GA adopts a $0.08/session-hour fee (Anthropic-style), 30,000 × 0.167 h running | — | +$400 | ~$2,861 no-cache |
Three honest footnotes: the runtime figure scales linearly from AWS's published example, so a workload with more I/O wait than 90% prices lower; a 20-turn task resends the conversation each turn, so caching is not optional — it is the difference between the two middle rows and the bottom two; and batch (direct-only for these models) would cut the direct no-cache total to $1,012.50, a discount BMA cannot currently replicate for GPT-5.4/5.5.
When each side wins
Direct API wins on price when: you don't need residency (skip the 10%), your workload can tolerate batch, you can route a large share of traffic to mini/nano tiers, or you're already below any enterprise discount threshold. Cache-heavy agent loops shrink the absolute premium too — $202/month becomes $121 at 80% cache hit rate in the scenario above.
BMA wins when: inference must stay in-region (that's its headline: "all inference on Amazon Bedrock", "customer data never leaves AWS"), tokens should count toward AWS commitments instead of a separate OpenAI invoice, you want the agent loop inside IAM/CloudTrail with per-agent roles and human-approval gates for consequential actions, or you simply do not want to build and operate an agent loop against raw completion APIs. None of those capabilities is cheaper; you pay roughly 10% of tokens plus your runtime for the privilege.
A useful read of the whole market: OpenAI's Agents API bundles the loop free with tokens; Anthropic charges $0.08/session-hour for the same idea; AWS charges nothing yet and +10% on tokens. The agent-harness layer is racing to zero margin while the tokens stay expensive — which tells you where the pricing war actually is.
Wrap-up
- The Bedrock Managed Agents token premium is a fixed, documented +10% over OpenAI list — and roughly a wash against direct API once you demand in-region residency, where OpenAI charges the same 10% uplift.
- The real cost risk is neither the premium nor the $0 preview fee; it's the unpriced GA fee plus always-on infrastructure (AgentCore compute, the sample stack's NAT gateway) that bills between turns. Budget on lines, not on a single "Bedrock agent price."
- For a 10-agent × 100-calls/day GPT-5.4-shaped workload, model ~$2,025/mo direct vs ~$2,461/mo on BMA+AgentCore (+21% all-in, +10% token-only); if your decisions are driven by residency, IAM/CloudTrail, or AWS commit consumption, that premium buys the things you were about to build anyway.
Sources
- Amazon Bedrock Managed Agents, powered by OpenAI, is now available in preview — AWS What's New, Sep 29, 2026
- Amazon Bedrock now offers OpenAI models, Codex, and Managed Agents — AWS What's New, Apr 28, 2026
- Amazon Bedrock Managed Agents, powered by OpenAI (preview) — Amazon Bedrock User Guide
- GPT-5.5 model card — Amazon Bedrock
- GPT-5.4 model card — Amazon Bedrock
- Amazon Bedrock Pricing
- Amazon Bedrock AgentCore Pricing
- OpenAI models and Codex on Amazon Bedrock are now generally available — AWS ML Blog, Jun 1, 2026
- OpenAI models, Codex, and Managed Agents come to AWS — OpenAI, Apr 28, 2026
- OpenAI API Pricing
- Bedrock Managed Agents — OpenAI developer guide (API comparison table)
- Claude Managed Agents pricing — Anthropic documentation