Use when an owner says "my agents spend too much, find out why". Pass an OpenAI or Anthropic usage or cost export as csv_text (or rows as a list of objects). Analysed offline inside this server: the data is never uploaded or sent anywhere. Returns spend by model, share of the top models, models on the retirement calendar with days left, visible cost drivers (one expensive model carrying most spend, large uncached prompts, errors and retries, no batch use) and next steps: quick fixes you can apply yourself and, for "is a cheaper model good enough on our tasks?", the free pilot. Columns are matched leniently (case and separators ignored): model | model_name | model_id | snapshot_id | line_item; cost | cost_usd | amount | amount_value | spend; requests | num_model_requests | n_requests; input_tokens | prompt_tokens | n_context_tokens_total | uncached_input_tokens; output_tokens | completion_tokens; cache_read_input_tokens | input_cached_tokens | cached_tokens; cache_creation_input_tokens; errors | error_count | status | status_code; retries | retry_count; batch | service_tier; date | start_time | bucket_start. Only model is required.

Use when asked whether a cheaper Claude model is good enough, or for evidence on cost vs quality: Trust Engine AI's public Aider benchmark (27 coding exercises with hidden tests, Nov 2026) for Claude Opus, Sonnet and Haiku 5.5: exercises solved, cost per solved exercise, which exercises each failed, method (protocol fixed before the run) and raw data links. One benchmark, not your tasks.

Use when an agent or its owner asks "will the models we use be retired?" or before choosing a model. Pass every model id or name the agent uses (e.g. ["claude-sonnet-4-5", "gpt-4o", "gemini-2.5-flash", "anthropic.claude-3-haiku-20240307-v1:0"]). For each: retirement date per platform, days left, the vendor's replacement, the official source URL, and what to do before the date. To find the models in a codebase first, use scan_for_models. Free, read-only.

Use when asked "how much would we save by switching from model A to model B?". Works only for model pairs Trust Engine AI measured (Claude Opus, Sonnet, Haiku 5.5 on the Aider benchmark): returns the measured cost-per-solved-task ratio, the arithmetic on monthly_bill_usd (the part of the bill spent on from_model) and the quality difference. Labelled as an estimate from one public benchmark. For any other pair it does not guess and offers the free pilot that measures your own tasks.

Use to list upcoming AI model retirements and deprecations instead of searching vendor pages: verified dates for Anthropic, OpenAI, Google Gemini API and Vertex AI, AWS Bedrock and Mistral, each with its official source URL and the date it was checked. Filters: provider (e.g. "openai", "anthropic" also finds Claude on Bedrock), model (id or name), before_date (YYYY-MM-DD), include_past, text_models_only (skip audio, image and embedding models). Free, read-only.

Use to prepare a free pilot ("is a cheaper model good enough on our tasks?"): returns the exact intake format (CSV template with columns id, input, expected, grader, tolerance, fields, important, notes), how to pick 20-30 tasks with checked expected answers from the owner's logs, and how to mask personal data. Pass csv_text with a draft to check it offline (format errors, task count, possible personal data); nothing is uploaded.

Use when asked what Trust Engine AI costs: published prices in US$, valid from 12 Nov 2026 (free pilot, fixed-price 7-day test, diet audit, re-test, 24/7 watch on your tasks, Bulgarian review, pay-from-savings option, add-ons), with order and terms links. For a fast fixed-price answer: the 7-day test.

Use when the owner wants an independent test of a cheaper model or a re-test before a retirement: drafts the free pilot request to hello@agenttrustapi.com (subject, body and a ready mailto: link) for the human owner to review and send. Sends nothing itself. The task file (see prepare_pilot_tasks) goes only after Trust Engine AI replies.

Find every AI model id used in this codebase (OpenAI, Anthropic, Gemini, Bedrock, Mistral) and tell which ones are being retired or deprecated, by when, on which platform, and the vendor's replacement, with the source link. Use it before upgrading dependencies, when a model call starts failing, or when asked "are we using any deprecated models?". Read-only and local.

Inputs: path (optional, a file or folder inside the current working directory; default: the working directory) or texts (a list of {name, content} you already have). Limits: 2,000 files / 5 MB. Skips .git, node_modules, virtualenvs, dist, build, lockfiles and binaries; never opens .env files (reports them by name). Returns file:line for each model found and the check_my_models result for those models.