Together AI Fine-tuning by Together AI

# Fix list: Together AI Fine-tuning

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/together-fine-tuning, the October 2026 research run, assessed 1 October 2026. Grade C, 54.9 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 Together AI Fine-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 prices for every tunable model, with minimum charges, published without a login (20). The billing docs say 'Together AI does not currently offer free trials' and access needs a $5 credit purchase (0). Sign-up and the first purchase are browser steps (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, 42 out of 100, up to 9.4 more on the total

Why it scored 42: GET /v1/fine-tunes returns truncated job objects, with no limit or field selection (10). No pagination or filter parameters on the list call found (0). An error page maps 11 status codes to a cause and a fix, scoped to inference; fine-tuning errors aren't listed (12). No idempotency keys or client-chosen job IDs; the docs advise trying 5xx again after a short wait (5). Only `model` and `training_file` are required, defaults are documented, and official SDKs exist for Python and TypeScript plus the `tg` CLI (15). We didn't score POST /v1/fine-tunes/estimate-price, a dry-run quote the checklist has no line for, though it saves an agent a failed or overpriced job.

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, 55 out of 100, up to 9 more on the total

Why it scored 55: Better Stack status page at status.together.ai with component history, but its components are serverless inference models, the website and the playground; nothing covers fine-tuning jobs or dedicated endpoints, so half credit (10). The incident archive says 'No incidents reported' for July, August and September 2026, while the monitors show 30-day uptime of 98.854 per cent for one model and 98.999 per cent for Kimi K3. We count per-model downtime as minor (20). Serverless limits are dynamic and no numbers are published; no limits for fine-tuning calls found (5). 429s carry `x-ratelimit-reset` in seconds and the docs recommend exponential backoff, and GPU quota rejections on dedicated endpoints return 429 naming the GPU since 27 July; no safe-retry guidance for job creation (10). No SLA found on the pricing page or in the docs (0). Fine-tuning carries no beta label (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.

## 4. Security & auth, 50 out of 100, up to 8.8 more on the total

Why it scored 50: Project-scoped API keys, revocable, with an optional expiry from 1 hour to a custom date since 7 August 2026; within its project a key has full access and can spend the credit balance (25). Projects isolate resources, but the project roles are Admin and Editor and there's no read-only role; no confirmation for deletes (5). Returns job state and your own model's output, no third-party content (10). Per-job event lists over the API; no audit log found, and the docs say per-key spend caps don't exist (5). security.txt sends reports to a private HackerOne programme; no SOC 2 or ISO report, bug bounty terms or public advisories found in the docs (5).

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, 78 out of 100, up to 3.6 more on the total

Why it scored 78: A public OpenAPI file at docs.together.ai/openapi.yaml ('Together APIs' 2.0.0); the part we could read covered endpoints, deployments and rollouts and stopped before the /fine-tunes paths, though the fine-tune reference pages show the same typed schema (20). llms.txt with .md twins (10). Field descriptions say what a field does and sometimes the trade-off ('may improve results but increase cost and risk of overfitting'), rarely when to use it (12). Two required fields, enums for `training_type`, `training_method` and the scheduler, defaults and limits such as `n_checkpoints` up to 10 and `suffix` up to 64 characters (13). Python, TypeScript, JavaScript and curl examples on each call; the error code page covers inference only and the fine-tune reference documents only a 200 response (8). Versioned /v1 paths and a changelog with over 50 dated entries since July (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, 70 out of 100, up to 2.6 more on the total

Made of editorial 55, provenance 85.

Why it scored 70: Closed service under terms that name Together Computer, Inc., a Delaware corporation; SDKs are Apache-2.0 (20). The privacy policy (2025-12-17) says no training on customer data without opt-in and the docs agree, but the docs also say prompts and responses are stored by default unless storage or ZDR is turned off, and no retention period for training files or tuned weights found (18). A deprecations page with a model lifecycle policy and dated notices in the changelog, often about two weeks ahead (12). EU-region and VPC options for enterprises are mentioned; no subprocessor list found (5).

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):

- Domain age: together.ai, no registry record we could read (0 of 15)

## 7. Maintenance & community, 80 out of 100, up to 1.8 more on the total

Why it scored 80: Changelog entries on 29 September (LoRA rank up to 128) and 1 October 2026 (30). Over 50 dated changelog entries between 1 July and 1 October, 18 of them about fine-tuning (20). Public changelog and support channels; we didn't get to read the SDK issue tracker, so no evidence of replies either way (10). Python (together 2.x) and TypeScript SDKs are current (15). Two Python repositories publish under the same package name, with together-python marked deprecated, and we didn't confirm CI on together-py (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.

## 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.

- We couldn't confirm that openapi.yaml includes the /fine-tunes paths; the readable part stopped at the deployment endpoints.
- We couldn't read the deprecations page or the together-py issue tracker within this run's fetch budget, so the deprecation and responsiveness scores lean on the changelog alone.
- No SOC 2, ISO or subprocessor page was found in the docs index; Together may publish these elsewhere.
- The pricing page now lists dedicated endpoint GPUs at $5.49 (H100) and $8.99 (B200); the listing's $3.99, $5.99 and $8.19 figures may be GPU cluster rates rather than endpoint rates, so we replaced them in pricingNotes but left unitPrices for the editor.
- Whether failed or cancelled jobs are billed isn't stated in the pages we read.

## Weaknesses

- Fine-tuned models don't run serverless; dedicated endpoints start at $5.49 an hour
- No free trial, a $5 prepaid purchase before the first call, and job minimums up to $60
- The status page covers serverless models only, and no fine-tuning rate limits are published
- No pagination on the job list and no documented error responses for fine-tuning calls
- No read-only project role and no audit log found

## 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.

- Call POST /v1/fine-tunes/estimate-price with the same body before creating the job, and check the model's minimum charge
- Read `lora_training.max_rank` from the model limits response before setting `lora_r`; most models went to 128 on 2026-09-29
- Don't retry a create call blindly after a timeout; there's no idempotency key, so list jobs and check first
- Download with checkpoint=adapter if you'll merge locally; merged weights for a 70B model are a large stream
- Tear down the dedicated endpoint once evaluation ends, since it bills while idle

## What the review panel asked for

- fine-tuning on the status page
- Add per-key spend caps
- Allow serverless serving

## 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.