Amazon Bedrock model customisation

# Fix list: Amazon Bedrock model customisation

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/amazon-bedrock-customization, the October 2026 research run, assessed 8 October 2026. Grade BB, 75.8 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Amazon Bedrock model customisation: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Payments & pricing, 30 out of 100, up to 8.8 more on the total

Why it scored 30: No x402, MPP or L402 (0). Per-token training prices, the $80 hourly rate for reinforcement fine-tuning and the $1.95 monthly storage charge are public without a login. The token figures are drawn by script from a public price file (20). No free tier for customisation on the pricing page. New AWS accounts get up to $200 in Free Tier credits over six months and the Free Tier page lists Amazon Bedrock. The page doesn't say whether a payment method is needed or whether credits cover training jobs, so half, as on the other AWS listings (10). An AWS account, an IAM role and S3 buckets are set up by a person first (0).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):

The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).

- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.
- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login.
- 20, a free tier or trial that doesn't need a card.
- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).

Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.

Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol.

## 2. Maintenance & community, 45 out of 100, up to 4.8 more on the total

Why it scored 45: The newest customisation API change found is `ModelPackageArn` on `CreateCustomModel` in boto3 1.43.17, uploaded on 28 May 2026, 133 days before 8 October. Reinforcement fine-tuning arrived on 3 December 2025 and the OpenAI-compatible route on 17 February 2026 (10). No customisation entry in the document history since 17 February 2026, and none among the latest 100 Bedrock posts on What's New, which run back to 10 March 2026 (0). Public document history and What's New, with support through re:Post and paid AWS Support (10 of 15). Current SDKs, boto3 1.43.109 on 7 October 2026 and `@aws-sdk/client-bedrock` 3.1148.0 (15). The SDKs ship several times a week and boto3 supports Python 3.10 and later (10).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):

- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.
- 20, at least three releases or dated changelog entries in the last 90 days.
- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.
- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).
- 10, package health, current dependencies and CI.

Models are read for deprecation notice periods and model churn rather than release counts.

## 3. Agent ergonomics, 77 out of 100, up to 3.7 more on the total

Why it scored 77: List calls return job summaries, not full job objects, and `maxResults` runs from 1 to 1,000. No field selection (15 of 25). `ListModelCustomizationJobs` filters by status, name and creation time, sorts by creation time and pages with `nextToken` (20). Typed exceptions with HTTP codes, plus a troubleshooting page for S3 permission, data format and token-limit failures. A quota breach returns 400 `ServiceQuotaExceededException` beside the 429 (16 of 20). `clientRequestToken` on job creation, and the published service model marks the operation idempotent. No equivalent was found for the OpenAI-compatible route (18 of 20). SDKs in nine languages and the CLI, but a job needs a role ARN, a base model, two names, an S3 training location and an S3 output location, hyperparameter names vary by model, and the result needs a separate deployment or Provisioned Throughput before it answers (8 of 15).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):

- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).
- 20, pagination, filtering and output-size controls.
- 20, actionable, documented error responses, codes and messages an agent can recover from.
- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.
- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.

Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs.

## 4. Security & auth, 87 out of 100, up to 2.3 more on the total

Why it scored 87: IAM with Signature Version 4, roles and short-lived credentials, and policies per action and resource. The OpenAI-compatible endpoints also accept Bedrock API keys as bearer tokens, short-term for up to 12 hours or long-term with a set expiry, revocable through IAM, and none travels in a URL (30). Job creation is its own IAM action, the service role can be limited to named S3 buckets, the caller needs `iam:PassRole`, and jobs can run inside a VPC. No confirmation step before deleting a custom model was found (16 of 20). The API returns job state and the customer's own model output. Reinforcement jobs call a Lambda function the customer wrote (10). The CloudTrail page says all Bedrock control-plane operations are logged as management events by default. Whether calls to the OpenAI-compatible fine-tuning endpoints are logged was not checked (13 of 15). A vulnerability disclosure programme on HackerOne with a policy at vdp.aws.security, and Amazon Bedrock in SOC 1, 2 and 3 scope on AWS's list updated 11 August 2026 (Bedrock Marketplace excluded). The aws.amazon.com security.txt expired on 24 September 2026 and no paid bounty was found, read as the other AWS listings read the same evidence (18 of 20).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-security):

- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.
- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.
- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.
- 0 to 15, audit logs or per-call visibility for the operator.
- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.

Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing.

## 5. Schema & documentation, 87 out of 100, up to 2.1 more on the total

Why it scored 87: AWS publishes the Bedrock service model (API version 2023-04-20, 108 operations) that the SDKs are generated from, and the API reference gives each field's type, pattern and enum (25). The user guide has an llms.txt and answers every page as Markdown at `.md` (10). The overview says what each method is for, and the reinforcement page lists when to use it and when supervised tuning should come first. Guidance on choosing between distillation and fine-tuning is thinner (16 of 20). `customizationType` is an enum and names and ARNs have patterns, but `hyperParameters` is a string-to-string map whose keys differ per model and are documented only in prose (10 of 15). Code samples for Python, CLI and HTTP, eight typed errors with HTTP codes and a troubleshooting page that quotes job failure messages. One HTTP example for the OpenAI-compatible route names `gpt-4o-mini` as the model, which Bedrock doesn't list (13 of 15). A dated API version and a dated document history, though its newest customisation entry is 17 February 2026 (13 of 15).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):

APIs and MCP servers.

- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).
- 10, llms.txt or Markdown docs served for agents.
- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.
- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.
- 0 to 15, examples and documented error responses.
- 15, versioning and a public changelog.

Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference.

## 6. Transparency & trust, 83 out of 100, up to 1.5 more on the total

Made of editorial 77, provenance 88.

Why it scored 83: Closed service under the AWS Customer Agreement and Service Terms. The SDKs are Apache-2.0 (15). Service Terms section 50.12.4 gives the customer exclusive use of a customised model and bars third-party providers from accessing it. The encryption page says training data is used only for the job, is not stored after it completes and is not used to train base models, and custom models are encrypted with an AWS-owned or customer-managed key. Section 50.12.2 allows up to 30 days of input and output storage for abuse detection on certain models, and the documents don't say whether that reaches inference on a tuned model (26 of 30). The model lifecycle page sets Legacy notice periods of six months or 45 days before end of life and says what happens to customised models, which lose new jobs and new Provisioned Throughput once the base is Legacy (18 of 20). AWS's sub-processor page was last updated on 28 July 2026, and each base model's Region is stated (18 of 20).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):

- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.
- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).
- 0 to 20, a deprecation policy or notices with dates.
- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).

The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two.

Provenance checks not met in full (half of this category, computed from checked facts):

- Terms of service: read, states 6 of the 7 things a reader expects, and has 3 clauses that cost points (3.1 of 10)
- security.txt: published but past its Expires date (5 of 10)

## 7. Reliability, 95 out of 100, up to 1 more on the total

Why it scored 95: Graded on the hosted lines. AWS Health Dashboard with per-service, per-Region history (20). AWS's dashboard history file, read on 8 October 2026, has no Bedrock event in us-east-1 or us-west-2 inside 90 days. The only Bedrock-specific entry is a model access problem in us-east-1 on 13 June 2026. Bedrock is named in the Middle East Region disruptions open since March 2026 and in a packet-loss event in eu-south-2 on 4 October, but fine-tuning runs only in us-east-1 and us-west-2, so we read the record as clean for this product, which is our call (30). Quotas are published with numbers in the AWS General Reference, including 10 scheduled customisation jobs, 2 custom models in creating status, 100 custom models and 10 deployments per Region, and record caps per base model (15). `ThrottlingException` is a documented 429, the API error page advises retries with exponential backoff and jitter, and `clientRequestToken` makes job creation idempotent (15). The Bedrock SLA, last updated 4 October 2023, promises service credits below 99.9 per cent monthly uptime and counts 500 errors on Bedrock API requests. It doesn't mention customisation or training jobs (5 of 10, our call). Supervised fine-tuning, reinforcement fine-tuning and distillation carry no preview label in the user guide (10).

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):

Hosted APIs, MCP servers, models and platforms.

- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).
- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.
- 15, rate limits documented with numbers.
- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.
- 10, an SLA published for any paid tier.
- 10, the surface agents use is generally available, not beta or preview.

Local packages, SDKs, frameworks and stdio MCP servers.

- 20, installs from an official package with supported runtimes stated.
- 25, a public CI and test suite, passing on the default branch.
- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).
- 15, semver discipline and breaking changes called out in a changelog.
- 15, version 1.0 or later, or declared stable.

Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- The Reliability incident score of 30 rests on reading only us-east-1 and us-west-2, where fine-tuning runs. The Bedrock Guardrails listing scored 10 on the same history file because Guardrails runs in the disrupted Middle East Regions
- Whether the Bedrock SLA covers customisation jobs or only inference requests. The SLA doesn't say
- unchecked: whether a new AWS account needs a payment method and whether Free Tier credits can pay for training jobs. The Free Tier page states neither, and the FAQ was not read
- unchecked: training prices for Claude 3 Haiku and Llama 3.3 70B, and all rows outside us-east-1 and us-west-2, in the script-drawn pricing tables
- Whether failed or stopped jobs are billed. Not found in the reviewed documentation
- Whether weights of a tuned open-weight model (gpt-oss-20b, Qwen3 32B) can be exported. No export operation was found, and Service Terms section 50.11 forbids extracting weights
- The pricing page names Preference Optimisation, which the user guide and service model don't document
- Whether abuse-detection retention of up to 30 days applies to inference on a customised model
- Service Terms section 1.8 allows benchmarks on condition that the tester discloses everything needed to replicate them, which matters before our probes run
- The OpenAI-compatible endpoints are on `api.aws`, a second AWS domain, which the provenance check may read as off the vendor's domain
- The lead was right on vendor, methods and interface. It didn't mention the OpenAI-compatible `/v1/fine_tuning/jobs` route for open-weight models

## Weaknesses

- Fine-tuning runs in us-east-1 and us-west-2 only, and each base model in one of them (two for the Titan models)
- No weight export was found in the reviewed documentation, and Service Terms section 50.11 forbids extracting model weights
- Setup needs an IAM service role, `iam:PassRole` and S3 buckets for input and output before the first job
- Reinforcement fine-tuning is billed at $80 a training hour and covers three models. Distillation is not available for Anthropic models
- The Bedrock SLA is dated 4 October 2023 and doesn't mention customisation jobs. No customisation launch appears in the latest 100 Bedrock What's New posts

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Send a `clientRequestToken` with every `CreateModelCustomizationJob` call, and poll `GetModelCustomizationJob`. Jobs are asynchronous and can take hours
- Create the job in the Region that hosts the base model (Nova in us-east-1, Llama and Claude 3 Haiku in us-west-2), with the S3 bucket in the same Region
- Pass an IAM service role that trusts `bedrock.amazonaws.com` and can read the training data and write the output location. The caller needs `iam:PassRole`
- After a job completes, call `CreateCustomModelDeployment` and use the deployment ARN as `modelId`. Models outside the on-demand list need Provisioned Throughput
- For gpt-oss-20b and Qwen3 32B, use `/v1/fine_tuning/jobs` on `bedrock-mantle.us-west-2.api.aws` with a Bedrock API key and a Lambda grader ARN

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.