Vertex AI Gemini tuning by Google Cloud

# Fix list: Vertex AI Gemini tuning

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/vertex-ai-tuning, the October 2026 research run, assessed 1 October 2026. Grade B, 64.2 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 Vertex AI Gemini tuning: 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, 20 out of 100, up to 10 more on the total

Why it scored 20: No machine payment protocol (0). Per-1M-token tuning prices published without a login, from $1.50 for Gemini 2.5 Flash Lite to $25 for Gemini 2.5 Pro and $10 for Gemini 3.5 Flash (20). No free tier for tuning found (0). A Cloud project with billing is a browser and console flow (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. Agent ergonomics, 48 out of 100, up to 8.5 more on the total

Why it scored 48: ListTuningJobs returns full job objects with no read mask found (10). `filter`, `page_size` and `page_token` on ListTuningJobs (20). Errors use Google's standard status codes; the error docs weren't readable (10). CreateTuningJob has no `request_id` or other idempotency field (0). `client.tunings.tune()` needs only a base model and a dataset URI, but the dataset must sit in Cloud Storage under a project with billing and IAM set up, and the only SDK we checked is Python (8).

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.

## 3. Reliability, 67 out of 100, up to 6.6 more on the total

Why it scored 67: Google Cloud status dashboard with per-product history, including Gemini on Agent Platform (20). No incident for Vertex or Gemini products between 1 July and 1 October 2026. The most recent was a 1 hour 58 minute degradation of the Gemini API global endpoint on 27 February (30). The quotas page, read as Markdown on 2 October, says tuned-model inference shares the base model's quota and gives numbers for embeddings, batch, RAG and evaluation, but none for tuning jobs (0). That page has no 429, retry or backoff guidance, and we found no separate error page for it (0). The Vertex AI SLA, last modified 2026-02-12, promises 99.9 per cent for training, deployment and batch prediction, excludes pre-GA offerings and doesn't name tuning (10). Supervised tuning sits in the v1 API with no pre-GA label on its docs page. Reinforcement learning fine-tuning is marked Pre-GA, runs on v1beta1 only and comes with a warning not to send confidential data, and the google-genai `tunings.tune()` method warns that its tuning implementation is experimental (7).

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.

## 4. Security & auth, 71 out of 100, up to 5.1 more on the total

Why it scored 71: OAuth 2.0 bearer tokens from service accounts or gcloud. We found no API-key route for tuning (30). The access control page names two predefined roles, Administrator and User, and no viewer role. Tuning needs individual `aiplatform.pipelineJobs.*` and `aiplatform.customJobs.*` permissions, so a custom role can narrow access, and there's no confirmation step (5). Returns job state and your own model's output, no third-party content (10). Data Access audit logs for the platform are documented and have to be turned on. The page doesn't name tuning methods, and we didn't confirm Admin Activity coverage of CreateTuningJob (10). Security programme (16). google.com's security.txt is valid, a public Cloud security bulletins page exists, and Google's services-in-scope list (updated 16 July 2026) marks Gemini Enterprise Agent Platform and Generative AI on Vertex AI for ISO 27001, 27017 and 27018 and SOC 1, 2 and 3. We didn't check the bug bounty scope.

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, 82 out of 100, up to 2.9 more on the total

Why it scored 82: Public contract in two forms, the GenAiTuningService proto in googleapis (CreateTuningJob, GetTuningJob, ListTuningJobs, CancelTuningJob, RebaseTunedModel) and the Agent Platform API discovery document (25). No llms.txt, but every docs.cloud.google.com page answers as Markdown when `.md.txt` is added to its URL, which we used on 2 October (10). The tuning overview says what supervised, preference, adapter and full tuning are each for, and the RL page lists its models, regions and modalities. Proto comments stay terse ('Optional. The standard list filter.') (14). Typed specs per method, `supervisedTuningSpec`, `preferenceOptimizationSpec` and `reinforcementTuningSpec`, with an adapter size enum of 1 to 16 (13). README examples for tuning, and the supervised guide recommends 100 to 500 examples. No error documentation for tuning found (5). Versioned /v1 API, platform release notes and a semver SDK (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. Maintenance & community, 80 out of 100, up to 1.8 more on the total

Why it scored 80: google-genai 2.27.0 tagged on 2026-10-01, after 2.26.0 on 2026-09-30 (30). 19 SDK releases since 4 July, eight of them between 25 August and 1 October (20). python-genai has 190 open issues and 103 open pull requests; platform release notes are public (10). The SDK is current (15). The SDK repository's workflows run mypy and import checks; test status not checked (5).

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.

## 7. Transparency & trust, 89 out of 100, up to 1 more on the total

Made of editorial 78, provenance 100.

Why it scored 89: Closed service under the Google Cloud terms, and the SDK is Apache-2.0 (20). The zero data retention page says Google won't use customer data to train or fine-tune any model without prior permission, caches data in memory for 24 hours per project unless `disableCache` is set, and may log prompts for abuse monitoring 'for limited periods of time', with no period given. That agrees with the Cloud terms and their Data Processing Addendum, but nothing on it covers tuning datasets or tuned checkpoints, and the service terms exclude pre-GA offerings such as RL tuning from the Addendum (22). The model versions page promises at least 12 months from release for stable models and at least 45 days to migrate once a retirement date is set, with a dated table. Gemini 2.5 Pro, Flash and Flash-Lite retire on 20 October 2026. It doesn't say what happens to tunes of a retired base model (16). The Cloud subprocessor list, last modified 20 August 2026, names third-party and Google subprocessors with their locations and lists Gemini Enterprise Agent Platform among the services they touch, and the API is served from regional endpoints such as us-central1 (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.

## 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.

- Whether the SLA's 'Training' service covers managed Gemini tuning jobs. The SLA doesn't name tuning
- No quota for tuning jobs on the quotas page. Its pipelines link may hold one and wasn't read
- unchecked: a 429 or retry guidance page for the platform. The old error-code URL is a 404 on the new docs site
- What happens to tunes of Gemini 2.5 Pro, Flash and Flash-Lite after those models retire on 2026-10-20. The model versions page doesn't say
- unchecked: the bug bounty scope, and whether Admin Activity audit logs name CreateTuningJob
- The service terms page was too long for our reader to find a training-restriction clause. The zero data retention page states one, and that's what we count
- The listing's lastRelease of 2026-09-30 was google-genai 2.26.0. 2.27.0 was tagged on 2026-10-01, corrected in patch

## Weaknesses

- No weight export. The tuned model exists only as a Google Cloud endpoint
- Tuned Gemini 3 inference costs 1.5x the base model for as long as you serve it
- Setup needs a project, billing, IAM and a Cloud Storage bucket before the first job
- RL tuning is Pre-GA on v1beta1, and the SDK's `tunings.tune()` is marked experimental
- Gemini 2.5 Pro, Flash and Flash-Lite retire on 20 October 2026, and the docs don't say what happens to their tunes

## 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.

- Use `client.tunings.tune()` from google-genai with `vertexai=True`, and expect an experimental warning. Tuning isn't available on the Gemini Developer API
- Add `.md.txt` to any docs.cloud.google.com URL to read the page as Markdown
- Tune Gemini 3.5 Flash or 3.1 Flash-Lite. The 2.5 models retire on 2026-10-20
- List jobs with a filter before re-sending a create after a timeout. There's no request ID to deduplicate it
- Count dataset tokens times epochs before submitting, since that product is the bill, and price serving at 1.5x base for Gemini 3 tunes

## What the review panel asked for

- say what happens to tunes of retired bases
- a GA tuning method in the SDK
- Publish tuning-job quotas
- State idle endpoint billing

## 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.