Skip to main content
BenchLM
Data

The system of record for AI vendor change

Every change to the models you run. With the receipt, where you work.

Releases, price changes, retirements, API changes, and incidents, read from 20 providers’ own pages and matched to the models, routes, and parameters you declare. Each record carries its source and its date.

Published facts only. Unknown stays unknown.

Five declared models with the retirement date for your route and alerts at 90, 30, and 7 days on Free. No card, no code upload, no provider keys.

Built for the person responsible for a production AI feature: founders, engineers, and platform owners, and the agencies that maintain several of them.

See how a change reaches you

The published record · read from the source

The published record for one exact model and route.

31days of runway from 2026-10-29

claude-sonnet-4-5-20250929

Retires
2026-11-30(exact)
Announced
2026-09-30
Route
Anthropic API
Replacement
claude-sonnet-5-5 · named by Anthropic; equivalence is not asserted

Same model, another route. Azure AI Foundry lists claude-sonnet-4-5 for 2026-10-19 (exact). Use the date for the route you call. Radar does not reconcile them.

Snapshot 2026-10-23. A published record establishes the provider’s own date and scope, not account access or production suitability.

The record, computed at build time

confirmed model releases in the last 30 days
50confirmed model releases in the last 30 days
published retirement dates in the next 90 days
55published retirement dates in the next 90 days
priced model rows with a source and an effective date
883priced model rows with a source and an effective date
providers with public status monitoring, of 20 read at the source
10/20providers with public status monitoring, of 20 read at the source

Counts come from the lifecycle feed, the release registry, and the pricing dataset at build time, never from copy. View source catalog (JSON)

Five kinds of change. Each one has an invoice, a pager, or a 410 behind it.

A provider changes something on a page nobody bookmarked. Radar reads the page, records the change with its source and date, and matches it to what you declared. This is what each kind of change does to a production feature, and what the record holds for it.

Price change

The invoice moves before your estimate does. A retired slug can be billed at the replacement’s rate.

Mistral Large 4 input price raised from $0.68 to $0.7 per million tokens on 2026-10-29, output from $2.09 to $2.1.

Each change with the provider’s page and date, source-backed price history, and a cost scenario from your own usage assumptions.

Retirement

The endpoint returns 410, sometimes on one route and not another. Notice ranges from six months to none, by provider.

claude-3-haiku-20240307: 2026-04-20 on Anthropic API, 2026-08-23 on Vertex AI, 2026-09-10 on Amazon Bedrock. 3 dates, 3 source pages. 55 published dates land in the next 90 days.

The date for the route you call, its precision, the provider-named replacement, and alerts at 90, 30, and 7 days of runway.

API and SDK change

A parameter you rely on changes meaning or disappears, and your tests may not cover it.

Documented changes are checked against the endpoint, SDK version, and parameter names you declared. First scope: OpenAI Responses. A documented change is something to check, not proof of a break.

Compatibility findings against the endpoint, SDK version, and parameter names you declared, with the outcome you record kept beside the evidence.

Incident

Start, update, and resolution arrive as separate notices from separate status systems, and someone has to reconstruct the story.

Public status monitoring for 10 of the 20 providers. Exact notices link into one incident timeline where the identity is qualified; ambiguous links stay separate.

One incident story with the original receipts, matched to the provider and models you declared, delivered where you choose.

Release

A better or cheaper model ships and you hear about it weeks later, from a post that does not say whether it fits your requirements.

50 confirmed releases in the last 30 days, each with its source page and the publisher’s own date.

Candidate checks against the requirements you saved, and a notice when a source-backed candidate newly matches them, with the requirements it met, failed or left unknown.

The line between Free and Pro. Free records retirements for five declared models and sends a morning read. Pro records all five kinds of change across your whole stack, delivers them where you work, and keeps the review behind each decision.

How much notice each provider promises

As published by each provider. A stated minimum is a commitment on their page, not a guarantee Radar makes.

55 published retirement dates land in the next 90 days. Which ones are yours?

From the 2026-10-23 snapshot of provider notices: 26 on first-party APIs, the rest on Azure, Vertex, and Bedrock. 47 carry an exact date; the others are approximate or earliest-possible, and the record says which.

The nearest published retirement dates in the current snapshot
DatePrecisionModelRouteProvider-named replacementSource
2026-10-31exactzai-glm-5-2Mistral APIZ.ai GLM 5.3docs.mistral.ai
2026-11-15exactclaude-haiku-4-5@1Azure AI FoundryNot publishedraw.githubusercontent.com
2026-11-15exactclaude-haiku-4-5@2Azure AI FoundryNot publishedraw.githubusercontent.com
2026-11-15exactcodex-mini@2025-05-16Azure AI FoundryNot publishedraw.githubusercontent.com
2026-11-17earliest possiblegemini-2.5-flash-native-audio-preview-12-2025Google APIgemini-3.8-liveai.google.dev
2026-11-17earliest possiblegemini-2.5-flash-preview-ttsGoogle APIgemini-3.8-flash-tts or gemini-3.8-flash-lite-ttsai.google.dev
2026-11-17earliest possiblegemini-2.5-pro-preview-ttsGoogle APIgemini-3.8-flash-tts or gemini-3.8-flash-lite-ttsai.google.dev
2026-11-17earliest possiblegemini-3.1-flash-live-previewGoogle APIgemini-3.8-liveai.google.dev
Start free

All 55, without an account: the deprecations record and the public feeds below.

Record. Match. Decide. Keep checking.

A change becomes one record with its source, the publisher’s date, and the moment Radar first observed it. Then it is matched to what you run, the review stays beside it, and the choice is checked again after you make it.

  1. 01 · Record

    Every change, with its receipt.

    Radar reads the provider’s own page and keeps one deduplicated record with its source, the publisher’s date, and the moment Radar first observed it. Corrections and retractions stay on the record instead of overwriting it.

  2. 02 · Match

    Matched to the stack you actually run.

    Declare the models, endpoints, SDK versions, and parameter names that matter to you: five on Free, 100 on Pro. The exporter finds literal model calls on your machine; you import only the metadata you select.

  3. 03 · Decide

    Decide with the evidence still attached.

    Reviews are ranked by their supported consequence, so the queue holds reasons to act rather than every provider update. Estimate monthly cost from usage assumptions you can inspect, import a comparison you ran yourself, and save a migration packet with a validation and rollback plan. Radar never edits code, moves traffic, or runs your evaluation.

A retirement notice can help you plan an update. A price change can prompt a cost review. A release can give you something new to evaluate.

Radar gives you the source for that decision. You decide what to do next. A documented change does not prove your app is broken; it tells you what to check.

Declare from your code, locally.

The Stack exporter finds literal model calls in JavaScript, TypeScript, and Python on your machine. Source files, prompts, and credentials never leave it.

node radar-stack-export.mjs --root . --output radar-stack.json

Keep checking after you choose.

A model can change behind an unchanged name. Schedule a check in your own runner at the cadence and budget you choose, import the bounded result, and compare it with the baseline you pinned. No hosted execution, no provider keys, no prompt retention.

Shipping with the roadmap
What if a date or serving route is missing?

Example · coverage gap: a published date for a provider’s own API may not apply to the same model on Bedrock or in another region. A missing date means Radar has no supported published date for that scope. It does not mean the model is safe indefinitely.

One record per application. One owner per decision.

A project is an application or a client, not a model name. Its dependencies, decisions, and reviews live together, each decision has an owner and a reason, and colleagues join with owner, editor, or viewer access that can be revoked. Across projects, one portfolio view shows what changed and which decisions are open.

Shipping with the roadmap

Project spaces

Dependencies, watches, reviews, and packets grouped by the application or client they serve. Personal workflows stay as they are.

Owners and roles

Assign a review owner, record the rationale, invite collaborators as owner, editor, or viewer, and revoke at any time. Every edit is attributable.

Client-ready reports

A portable review packet per project: changed evidence, open decisions, owners, assumptions, next actions. Sharing is explicit; nothing crosses client boundaries.

These arrive inside Pro first, so a colleague can own a review and a client can receive a report. Existing Pro rights are preserved.

Team pricing

There is no separate team price yet.

If you are buying for a team or an agency, tell us how many people need access, what a review has to show a client, and what your approval process expects. What you say shapes what gets built and how it is priced.

Tell us what your team needs

What a bookmark folder, a cloud console, or a gateway will not tell you.

Most teams already have a way to hear about model changes. It is usually a person. This is what changes when the record is a system.

Doing it yourself compared with Radar
 On your ownWith Radar
CoverageThe pages you remember to check, on the clouds you have a console for.20 providers’ release, pricing, lifecycle, API, product and status pages, plus the Azure, Vertex and Bedrock routes, read continuously.
Route divergenceEach console shows its own dates. Azure will not tell you about Bedrock.Every route’s published date side by side, with the spread between them stated.
ReceiptsA screenshot, if someone took one.Source page, published date, and first-observed time on every record. Corrections are kept, never overwritten.
RelevanceEverything, or nothing.Matched to the models, routes, and parameter names you declared. Runway alerts at 90, 30, and 7 days.
Silent changesA gateway can re-route to another model without telling you. A retired slug can be billed at another model’s rate.A retirement, a price change, or a redirect is a record you can read before it reaches your traces or your invoice.
The decisionA chat thread.A packet with the cost scenario, your own comparison, the validation plan, and the rollback conditions, kept with its revisions.
MachinesCopy and paste.Public feeds, an Event API with stable IDs, remote MCP for your agent, and a GitHub Action that fails a review inside the runway window.

Radar reads public pages only. It has no private provider access, never estimates a date, and never edits code or moves traffic.

In your CI, your editor, and your agent.

The same records, wherever the decision gets made. Read changes in Radar, send selected alerts to your inbox or chat, or ask your agent to retrieve the evidence. The public feeds and the CI check need no account.

Read and review

Review the Feed, match changes to your Stack, and use Follow-up for supported questions you want to revisit.

Open Radar

Choose your alerts

Email, Slack, Discord, and signed webhooks. Select your providers, alert types, and delivery preferences.

See delivery setup

Connect an agent

Ask for releases, deprecations, and model status with sources attached. The setup guide shows which API and MCP tools your account can use.

Read the MCP setup guide
Agent setup and complete API & MCP tool inventory

Grok Bot

Add Radar’s remote server as a custom MCP connector.

OpenClaw

Install the Radar skill, then connect over Streamable HTTP.

Hermes

Install the Radar skill, then authenticate with your Radar API key.

Four stable source tools are always present for active Pro keys. Twenty-one additional tools, including authorized stack writes, require a current evidence receipt, product gate, and account entitlement. Reading evidence does not authorize writes. Decision-watch writes are not granted to API keys or MCP.

Stable source tools

  • list_changes
  • get_change
  • get_deprecations
  • check_model

Stack writes and runway

  • declare_dependency_v1
  • set_dependency_state_v1
  • set_action_workflow_v1
  • get_model_runway_v1
  • get_lifecycle_calendar_v1

Account and Follow-up

  • list_declared_dependencies_v1
  • get_declared_dependency_lifecycle_v1
  • list_active_actions_v1
  • list_account_events_v1
  • get_account_event_v1

Exact reviews

  • review_dependency
  • review_dependency_v2
  • review_incident_response
  • review_api_contract_changes_v1

Evidence exploration

  • query_model_fit_v1
  • explore_benchmarks_v1
  • get_capability_taxonomy_v1
  • list_capability_candidates_v1
  • assess_capability_v1
  • list_model_value_endpoints_v1
  • explore_openrouter_activity_v1
Event API documentation

Follow published lifecycle notices without an account.

Subscribe to published notices and retirement dates, with original sources. Provider filters are available.

Feed formats and provider filters

Append ?provider=openai to any feed URL to request that provider’s published notices.

  • feed.json · JSON Feed 1.1 with structured provider, model, route, and date fields
  • feed.xml · RSS 2.0 for any reader or Slack RSS app
  • calendar.ics · Published retirement dates on your team calendar

Shipping with the current roadmap

Marked on this page until it is live.

Everything else on this page is live today. Each item here ships only when its evidence and operating limits pass review; no date and no new price is promised, and the marker disappears when it lands.

  • Shortlists, saved trade-offs, and the review trigger behind each decision.In development
  • Project spaces, review owners, scoped roles, decision rationale, portfolio view, and client-ready reports.Next
  • Scheduled checks in your own runner and regression review against a pinned baseline.Next
  • Not offered: Model Fit, Model Value, Dependency Review, and Incident Response stay gated until their results are trustworthy.Gated

The sources behind every record.

Release notes, pricing pages, retirement notices, developer documentation, model catalogs, benchmark tables, and public status pages. Coverage changes as sources pass or fail review, and the live catalog is the receipt.

Check the current catalog for monitored coverage.

View source catalog (JSON)

Every record keeps its source. Coverage varies by provider, model, platform, and region. Published dates come from the source; observed dates tell you when Radar first saw a change. A record for a provider’s direct API does not apply to Bedrock, Vertex, or Azure unless that route has its own published notice.

See all monitored providers and source categories

Provider list for this page. Open the source catalog to inspect configured sources and their scope.

Pro provider coverage20/20 visible
  • OpenAI logo
    OpenAI
  • Anthropic logo
    Anthropic
  • Google logo
    Google
  • xAI logo
    xAI
  • DeepSeek logo
    DeepSeek
  • Cursor logo
    Cursor
  • NVIDIA logo
    NVIDIA
  • Cohere logo
    Cohere
  • Moonshot AI / Kimi logo
    Moonshot AI / Kimi
  • MiniMax logo
    MiniMax
  • Arcee AI logo
    Arcee AI
  • Alibaba / Qwen logo
    Alibaba / Qwen
  • ByteDance Seed logo
    ByteDance Seed
  • Inception Labs logo
    Inception Labs
  • Meta logo
    Meta
  • Mistral logo
    Mistral
  • Sakana AI logo
    Sakana AI
  • Tencent Hunyuan logo
    Tencent Hunyuan
  • Thinking Machines logo
    Thinking Machines
  • Z.ai logo
    Z.ai
  • Model and catalog releases
  • API and documentation changes
  • Published pricing changes
  • Lifecycle and deprecation notices
  • Benchmark and research updates
  • Product and service-status changes

Free proves the first result. Pro keeps the decision maintained.

Free gives you a five-model Stack with runway alerts and an optional morning read. Pro gives you all five kinds of change across your whole stack, the complete included Feed and history, the Reviews, and control over what reaches you.

Free

$0

No card required

For the models you run and a morning read.

  • Add up to five models to your Stack
  • Retirement dates for your route, where published, with alerts at 90, 30, and 7 days
  • Optional morning Brief with up to five confirmed changes
  • Seven Brief issue dates
  • Public lifecycle feeds and the GitHub Action, no account needed

Radar Pro

$19.99/month

Billed monthly

For the models your work depends on.

  • Up to 100 current declared dependencies, matched against the complete included Feed and history
  • All five kinds of change: releases, price changes, retirements, API changes, and incidents
  • The Reviews: compatibility findings, cost scenarios with source-backed price history, workload comparisons, migration packets, candidate checks
  • Choose which providers and change types trigger alerts
  • Deliver to email, Slack, Discord, or signed webhooks
  • Event API and remote MCP; writes where enabled
  • Save Feed views you want to return to
  • Shortlists, project spaces, portfolio view, client reports, and scheduled checks, as each ships

Eligible new customers can try Pro for seven days. A card is required. Radar confirms eligibility, renewal price, and the first charge date before checkout. Cancel from the billing page; access continues to the end of the paid period.

Choose Free if

You run a handful of models, want the retirement date and the runway alert for each, and can live with a morning read. Most people should start here.

Choose Pro if

You are responsible for a production feature or several client applications, need all five kinds of change across your whole stack, alerts routed to Slack or a webhook, and a saved review behind each model decision.

Compare all plan details
Free and Radar Pro plan comparison
FeatureFreeRadar Pro
Declared stackUp to five models, with retirement dates for your route where published and runway alertsUp to 100 current declared dependencies, matched against the complete included Feed and history
CoverageYour declared models, plus the morning editorial selectionSupported providers and benchmarks; see current source coverage
Changes and historyUp to five confirmed stories on qualifying mornings; seven Brief issue datesAll five kinds of change, with the complete included event stream, Feed, and history
ReviewsNot includedCompatibility findings, cost scenarios, workload comparisons, migration packets, candidate checks; history and downloadable packets
NotificationsRunway alerts by email; morning Brief email if you opt inProvider, category and channel rules, selected alerts, and daily recaps
DestinationsEmailEmail, Slack, Discord, and signed webhooks
API and MCPPublic tools, public lifecycle feeds, and the GitHub ActionEvent API and remote MCP; private tools and stack writes depend on feature availability, evidence, and account rights
Agent supportPublic MCP tools for any connectionHermes, OpenClaw, and Grok Bot through remote MCP, with private Stack tools
Saved viewsNot includedIncluded
Model Fit, Capability, and Model ValueNot includedNot currently offered as included benefits
Dependency, incident, and private delivery reviewsNot includedNot currently offered as included benefits
Billing$0; no card$19.99/month or $199.90/year
TrialNot applicableSeven days for eligible new customers; Radar confirms charge timing before Stripe

Legacy Watch subscriptions remain supported inside Radar and are not offered to new customers. Some time-limited pilots carry different watch limits.

Looking for a specific review tool?

The tool guide explains more ways to review models and dependencies, and which ones are currently available. Check it if one specific workflow is the reason you are considering Pro.

Review tool descriptions and availability
01

Follow-up

Keep one bounded release, incident, or dependency question current until you mark it reviewed, complete, or stopped.

Available for supported evidence and account access.

02

Model Fit

Query exact runtime and access requirements against source-backed model evidence. Missing support stays unknown.

Not currently offered as an included benefit. See current tool availability.

03

Benchmarks and Capability

Inspect comparable cohorts and the evidence behind a capability assessment without turning uncertainty into a universal rank.

Not currently offered as an included benefit. See current tool availability.

04

Model Value

Compare feasible priced endpoints for one workload. Radar records assumptions and exclusions; it never switches production traffic.

Not currently offered as an included benefit. See current tool availability.

05

Dependency Review

Pin one declared endpoint dependency to lifecycle, reliability, and API evidence with a private review checklist.

Not currently offered as an included benefit. See current tool availability.

06

Incident Response

Keep one exact upstream incident timeline beside a private response checklist. Provider resolution and customer recovery remain separate.

Not currently offered as an included benefit. See current tool availability.

07

API and SDK changes

Review exact structural changes for a declared endpoint or pinned SDK method. Radar does not label them breaking or inspect customer code.

Not currently offered as an included benefit. See current tool availability.

08

Private delivery

Route owner-scoped Fit, Capability, Value, Follow-up, and review activity through selected email or signed-webhook preferences.

Not currently offered as an included benefit. See current tool availability.

Check current tool availability

Questions people ask before the first check

What can I do without an account?

Check the published record for an exact model and route, read the public source coverage catalog, run the GitHub Action, and follow the public lifecycle feeds as JSON Feed, RSS, or a calendar. None of that needs sign-in, a code upload, or a provider key. Declaring a stack needs a free account.

What does the free account include?

A Stack of up to five declared models with the retirement date for your route where a source publishes one, alerts at 90, 30, and 7 days of runway, and the optional morning Brief: up to five confirmed changes on qualifying mornings, with seven issue dates of history. No card is required.

What changes when I choose Pro?

Pro records all five kinds of change against up to 100 declarations, opens the complete included Feed and history, and adds the Reviews: compatibility findings, cost scenarios, workload comparisons, migration packets, and candidate checks. Notifications go to email, Slack, Discord, or signed webhooks, and the Event API and remote MCP expose the same records.

Is there a team plan?

Not as a separate price yet. Project spaces, review owners, and scoped roles ship with the current roadmap and arrive inside Pro first, so a colleague can own a review and a client can receive a report. If you are buying for a team, tell us what you need and it will shape the offer.

Does Radar edit my code, run my tests, or change traffic?

No. The hosted workflows read published evidence and record your review state. The Stack exporter and the Promptfoo exporter run on your machine and upload only the metadata you select. A change suggestion is a download for manual review; the optional GitHub Action opens a pull request only when you enable it.

What if my model or route is not covered?

Missing evidence stays unknown. A notice for a provider’s direct API does not transfer to Bedrock, Vertex, or Azure; each route needs its own published record, and the same model can carry three different dates. A missing retirement date never means a model is safe indefinitely. The coverage catalog shows where sources stop.

How does the seven-day trial work?

Eligible new customers can start a seven-day Pro trial with a card on file. Radar checks eligibility and shows the renewal price and first charge date before you reach Stripe. Returning or previously trialed accounts may be charged immediately; the order review states the timing. Cancel from the billing page at any time.

Does signing in subscribe me to email?

No. Sign-in and email preferences are separate. The Brief and Pro notifications are each an explicit opt-in with their own unsubscribe, and saving your work does not enroll you in anything. Quiet days produce no email at all, because Radar sends only when a qualifying change exists.

Can my agent or CI pipeline use Radar?

Yes. Remote MCP works over OAuth with an API-key fallback: public tools are open to any connection, and private tools appear for active Pro accounts. The Event API pulls a replayable stream with stable IDs and corrections. Downloadable packages include a GitHub Action and client recipes; write actions stay separate, explicit choices.

What happens to my history if I stop paying?

Eligible saved history stays readable after paid access ends: cost scenarios, workload comparisons, migration packets, and review records within each workflow’s cutoff. You can read and download them but cannot create or change paid records. Reading saved history is not continued access to new evidence, and legacy Watch subscriptions keep their existing rights.

Start with one model you run.

Check its published record, declare it, and let the next change find you with the source attached. When the decision gets serious, Pro keeps the review beside the evidence.

Seven-day trial for eligible new customers. Card required. Radar confirms renewal price and first charge date before checkout.

Already have an account? Open Radar