Recent changes

Product, API, pricing and model updates.

123Changes shipped
LatestModels

Gemini 4 Hydrogen is here.

Google's fast Gemini 4 model replaces Gemini 3.8 Flash for everyday agent, coding and multimodal work. At its default high effort it scores 51 on the Artificial Analysis Intelligence Index, one point above Argon, and lists 25% below it per token.

November 2026

8 changes

  1. Models

    Gemini 4 Carbon is here.

    Google's strongest Gemini 4 model is built on Argon for long coding, agent and research work, at the same per-token price. At medium effort it scores 58.3 on the Artificial Analysis Intelligence Index and starts answering in about five seconds.

    Explore Gemini 4 Carbon
  2. Models

    Claude Fable 5.2 and Muse Ember are here.

    Anthropic's flagship lists 20% below Fable 5.1 and generates up to 30% faster, and Meta's Muse Ember runs high-volume agent steps at about 6% of Muse Spark 1.3's cost per task. DeepSeek V4.1 Flash 1017 Ultrafast joins them for the fastest replies.

    Explore the new models
  3. Audio

    Turn a script into speech, or a recording into text, in Audio Studio.

    Hear 10 voices across TTS-1, TTS-1 HD and MAI-Voice before you pick one, then download MP3, WAV, FLAC, Opus or AAC. Upload or record audio and transcribe it with Whisper or MAI-Transcribe, with timestamps and subtitles.

    Explore Audio Studio
  4. Video

    Gemini Omni 1.2 Flash makes videos in Video Studio.

    Generate 3 to 15 second clips with sound at 360p, 720p or 1080p, or up to 8 seconds in 4K, in landscape or portrait. Prices start at $0.024 per second, and failed videos are free.

    Try Gemini Omni
  5. Image

    GPT Image 2.6 Flare and Sunburst are here.

    Generate 1024×1024 images with OpenAI's newest image models in Image Studio or through POST /v1/images/generations. Each image costs about $0.026.

    Read about GPT Image 2.6
  6. Image

    Create and edit images in Image Studio.

    Write a prompt on /image to get up to four images from GPT Image, Seedream, FLUX, Recraft, Grok or Muse, or edit a photo with FLUX Kontext Pro or GPT Image 2.5. Each image shows what it cost, and your history stays in your browser, so download the ones you want to keep.

    Explore Image Studio
  7. Video

    You can now share your videos as templates.

    Submit a text-to-video clip from History and mark the words people can replace. Approved templates join the library with your credit, ready for anyone to remix.

    Browse templates
  8. API

    Balance price and speed.

    Set sort to balanced and MiniRouter picks the route with the best price for its speed. Use it per request, in a preset or across your account.

    Explore balanced sort

October 2026

29 changes

  1. API

    Turn an image into a video with 15 models.

    Paste an image link in Video Studio, or send image_url to POST /v1/videos, to start from your own picture on Veo, Kling, Wan, MiniMax, Seedance and Grok. The price per second stays the same, Grok adds a small flat fee, and failed jobs cost nothing.

    Open Video Studio
  2. Video

    Start a video from a template and make it your own.

    Pick a clip on /video, replace the highlighted words and generate. Each template shows the prompt behind the clip and what a run costs.

    Browse templates
  3. API

    Introducing videos and our Video Studio.

    Try 17 text-to-video models from Veo, Kling, Wan, MiniMax, Seedance and Grok in the Video Studio, or send them through POST /v1/videos and the OpenAI SDK. Priced per second of output, and failed videos are free.

    Explore videos and the Studio
  4. Policy

    Deepfakes and intimate images of real people are now prohibited.

    The acceptable use policy now bans realistic media that shows a real person saying or doing something they did not, and sexual images of anyone made without their consent. Keys in breach are suspended.

    Read the policy
  5. Models

    Particle-1, Mistral Large 4 and Joey-1.

    Explore Particle-1 for coding and Mistral Large 4 for code and documents. Learn how Joey-1 handles typed decisions; it is not yet available on MiniRouter.

    Explore the releases
  6. Models

    Meet GPT-6.1 Terra.

    Explore OpenAI's latest Terra for everyday work, with Standard and Fast options. Compare it with GPT-5.6 Terra and review current pricing and API setup.

    Explore GPT-6.1 Terra
  7. Models

    Meet Abliterated Large Flash.

    A million-token context with input rates two-thirds lower and output rates 60% lower than Large v2. Choose low, high or max reasoning, or turn it off for direct answers.

    Explore Flash
  8. Models

    Meet Kimi K3.1 Pro.

    Explore Moonshot's latest model for coding and agent work, with a million-token context window. See what changed from Kimi K3, current prices and API setup.

    Explore Kimi K3.1 Pro
  9. Agents

    Find a better model for your budget.

    Give your agent four skills to choose models, audit spending and test a switch within an approved budget. Start with public advice or a local usage export, without an account.

    Explore the model skills
  10. Models

    Meet DeepSeek V4.1 Flash 1017.

    Explore DeepSeek's October update, with higher reported agent and coding scores at the same price. Compare it with September's release and pin the new ID.

    Explore V4.1 Flash 1017
  11. API

    Introducing the Decisions API.

    Turn shared context into probabilities, choices and scores with Jev. Explore an interactive example and copy a request.

    Explore Decisions API
  12. Models

    Meet Muse Spark Open.

    Explore Meta's open-weight model for coding and agents. Compare current prices, reasoning options and text API setup.

    Explore Muse Spark Open
  13. Models

    Meet Gemini 4 Argon.

    Explore Google's model for sustained coding and complex workflows. Compare reported results, introductory pricing and current API availability.

    Explore Gemini 4 Argon
  14. API

    Voice is here: speech, transcription and audio input.

    Turn text into speech, transcribe recordings, and send audio to chat models that accept it. One API key covers all three, and every hold settles to the real cost.

    Read the audio docs
  15. Models

    Meet Muse Wonder 1.0.

    Explore Meta's model for coding and long-running agents. Compare current prices, reasoning options and API setup.

    Explore Muse Wonder
  16. Models

    Claude Haiku 5.5 is now available.

    Use Haiku 5.5 for text requests through Chat Completions, Responses and Messages. Compare its tiered pricing and get the API setup.

    Explore Haiku 5.5
  17. Models

    Find the right model effort for your budget.

    Compare up to five models across effort levels using Artificial Analysis intelligence scores and benchmark task costs. Share your comparison with a link.

    Compare models
  18. Billing

    Stake MINI, earn credits.

    Stake 1M MINI or more at stake.minirouter.sh and credits land in your MiniRouter balance every hour, from $10,000 this month. Unstake any time; the reserve wallet backing each program is public.

    Stake MINI
  19. Models

    Meet Nano Banana 2.1.

    Generate images from text through Chat Completions with default image settings. See image prices separately from token costs and copy a working setup example.

    Explore Nano Banana 2.1
  20. Models

    Meet Mistral Large 4.

    Mistral's flagship preview joins the text catalog with 524K context. Compare its reported coding results, current prices and API setup.

    Explore Mistral Large 4
  21. Account

    Your accounts and keys, in order.

    Switch between accounts without signing out, rename API keys and set an expiry when creating a key. Keep each project easy to recognise and temporary access time-limited.

    Explore accounts and keys
  22. Chat

    Find a chat. Take it with you.

    Search conversation titles and messages to pick up where you left off. Export individual chats as Markdown or JSON.

    Explore chat search and export
  23. Site

    Ask the FAQ anything.

    Type your own question under the homepage FAQ and get a short answer drawn from the docs, with a link to the right page. No account needed.

    Meet the FAQ assistant
  24. Billing

    Web fetch costs $0.001 per successful call.

    Each successful fetch adds a $0.001 platform fee to model tokens; failed, blocked and timed-out fetches have no fetch fee, and redirects count once. Datetime stays free and previously settled charges stay unchanged.

    Read the fetch pricing
  25. Models

    More image models, one API.

    Generate with 26 image models and edit with FLUX Kontext Pro, Flare or Sunburst. One API key, pay as you go.

    Explore image models
  26. Models

    abliteration.ai is live as a direct provider.

    Two uncensored fine-tunes with up to 1M context work on Chat Completions, Responses and Messages. You pay provider cost plus 5%, and abliteration.ai keeps no data by default.

    Meet abliteration.ai
  27. Models

    Three more text models join the catalog.

    Seed 2.1 Turbo, Ling 3.0 Flash VL and Ling 3.1 Flash Free join the text model catalog. All three currently support text input only.

    Browse models

September 2026

65 changes

  1. Models

    Meet GPT-6.1 Sol and Fast.

    GPT-6.1 Sol and Fast have joined the text model catalog. Compare current rates and review Responses setup and reasoning settings.

    Meet GPT-6.1 Sol
  2. Models

    Meet Claude Sonnet 5.5.

    Sonnet 5.5 joins the catalog with 1M context and stronger coding scores at Sonnet 5's token price. Compare Anthropic's results and get the API setup.

    Explore Sonnet 5.5
  3. API

    Web search takes filters.

    Set max_results, country and search_language_filter on minirouter:web_search, alongside domain and recency filters and a cap on result tokens. Narrow each search to the sources, region and languages you want.

    See search options
  4. Routing

    Choose your balance with Fusion.

    Fusion Cheap, Fusion and Fusion Max give you three price–quality priorities. Each routes by task and complexity using benchmark quality, estimated cost and latency.

    Explore the three profiles
  5. Account

    Invite teammates by email.

    Paste up to 10 addresses in Team and each person gets their own invitation, valid for 7 days. They join once they verify the invited email.

    See email invites
  6. Billing

    Agents can open an account and pay their own way.

    POST /accounts/agent returns a key and a Solana deposit address, no browser needed. Fund it with USDC or any Solana token, which converts to credit on the way in.

    See how agents pay
  7. Access

    Get your key first, pay when you are ready.

    Creating a key now comes before payment, and Pay later lets you fund it afterwards. A five-step setup guide then takes you to your first request.

    Get a key
  8. Models

    Qwen 3.8 Max Prime and Ember 1 are live.

    Alibaba's high-speed Qwen 3.8 Max lane and Fireworks' Ember 1, a Kimi K3 tune that reports 40% fewer tokens at the same quality, are in the catalog. Both take 1M context and image input.

    Meet both models
  9. Account

    Teams.

    Share one prepaid balance with up to 10 people, each with their own keys and logs. The owner funds it, sets routing and invites by link.

    See how it works
  10. Billing

    Pay with MINI at 105%.

    MINI you pay credits at 105% of its value while the monthly allowance lasts, then 100%. The price is the lower of spot and a 30-minute average.

    Pay with MINI
  11. API

    Bring your own provider keys.

    Connect OpenAI, Anthropic or DeepSeek with a 0% MiniRouter service fee for one enabled key per provider, or 3% of published-rate usage with multiple enabled keys for the same provider. Your provider bills inference separately.

    Explore provider keys
  12. Billing

    Export your transactions from one place.

    Billing, Usage and Activity share one Export button: deposits and settled charges for any period, as a CSV a spreadsheet can sum. Refunds, free credits and holds are left out.

    Export transactions
  13. Models

    GPT-6 Sol and Luna are live.

    OpenAI's GPT-6 Sol, Luna and their Fast versions are in the catalog with 1.05M context and vision input. Both cost half their GPT-5.6 predecessors, and Sol beats Claude Opus 5 on AutomationBench.

    Meet GPT-6 Sol and Luna
  14. Models

    Claude Opus 5.5 is live.

    Anthropic's Opus 5.5 and Opus 5.5 Fast are in the catalog with 1M context and vision input. It scores 66.4% on Terminal-Bench 4.0 at 20% less per token than Opus 5.

    Meet Opus 5.5
  15. Models

    MiMo V2.6 is live.

    Xiaomi's V2.6 Pro, Flash and Pro UltraSpeed are in the catalog with 1M context and vision input. Pro scores 71.9% on DeepSWE, near Claude Opus 5.

    Meet MiMo V2.6
  16. Models

    Explore Grok 4.7.

    Review SpaceXAI's new coding model, sourced benchmarks and temporary provider pricing. The release guide tracks MiniRouter catalog availability.

    Meet Grok 4.7
  17. Models

    Meet Step 5 Preview.

    StepFun's flagship preview gets a release guide covering its 1M-token context, vision and published evaluations. Check the guide for MiniRouter catalog availability.

    Explore Step 5 Preview
  18. Routing

    Build your own router visually.

    Choose task type, complexity, both, or a custom question, then drag models into routes while comparing intelligence, speed and price. Jev's highest score wins, your answer order breaks ties, and you can test and preview the configuration before saving.

    Build a router
  19. API

    Get structured evaluations with Jev.

    Use /v1/evaluate or the native TypeSafe /typesafe/v1/systemone endpoint for typed decisions. Availability follows the qualified catalog, with finalized provider inference cost plus 10% pricing.

    Meet Jev
  20. Billing

    Auto and Fusion have their own routing rates.

    Auto now charges provider cost plus 8%, and Fusion plus 10%, including classification. Published router rates include these total markups. If Fusion cannot classify a request, it uses Auto at 8%; Auto Free and chat trials remain free.

    See request pricing
  21. Models

    Explore Bonsai 2 27B.

    A new release guide explains PrismML's ternary model, Darkbloom pricing, and text API support. MiniRouter API access is not live yet.

    Read about Bonsai 2
  22. Chat

    Chat is free to try, no key needed.

    Twenty messages per account on Qwen 3.7 Flash, with Ling 3.0 Flash standing in when it is busy. Past conversations sit in a sidebar.

    Open chat
  23. Billing

    Holds shrink and clear the moment a stream ends.

    A streamed answer settles from the provider's reported cost as it finishes, instead of waiting up to forty seconds for a lookup. Continuing a thread holds its cached prefix at the cache-read rate.

    Read about holds
  24. Codex

    Codex compaction turns go through.

    When Codex summarises a long session it sends an empty tool list, which was refused as invalid. It is accepted now, and a refused request names the field.

    Codex setup
  25. Site

    Analytics shows the tokens and requests MiniRouter serves.

    See tokens and successful requests over 30, 90 or 180 days, with daily or weekly totals and the most-used published models. Totals update once a day and exclude internal and test traffic.

    Open analytics
  26. Models

    Compare models side by side.

    Click up to six rows on the models page to see input price, output price, context window and intelligence side by side. The link carries the selection.

    Browse models
  27. Chat

    Chat shows a router choosing your model.

    Pick Fusion or Auto in Chat and watch it choose: the candidates spin, land on the model that answers, and Fusion adds the task, tier and effort it read from your message.

    Open Chat
  28. API

    Fusion reads your request and picks the model.

    Send minirouter/fusion and a classifier labels the request's task and complexity, then routes it to the cheapest model and reasoning effort that fits: a quick question goes to a fast, cheap model, a hard problem to a frontier one. The x-minirouter-route-reason header shows the label, and your own reasoning setting always wins.

    How Fusion picks
  29. API

    Claude models cache long prompts automatically.

    Fable 5.1, Opus 5, Sonnet 5 and Haiku 4.5 now cache repeated tools and history for five minutes on Responses and Chat Completions, so each agent turn pays the cached rate for what it already sent. On Responses, cache_ttl: "1h" keeps it longer; caching: "disabled" opts out on either.

  30. Integrations

    Codex lists your presets and the newest models.

    Codex now loads MiniRouter's model list, presets first and then the newest tool-capable models, and switching model mid-session keeps working. Long sessions on OpenAI models hold counted tokens instead of request bytes, so a funded account no longer meets a 402.

    Set up Codex
  31. Integrations

    The current Claude Code release works with Messages again.

    Messages accepts the context_management field and the system-role turns recent Claude Code versions send, which it used to refuse. Nothing changes in your settings.

    Set up Claude Code
  32. Dashboard

    A guardrail on a key caps its spend and redacts sensitive text before any model sees it.

    Create one on Guardrails with a daily, weekly or monthly budget, model and provider lists, built-in detectors for emails, phone numbers, card numbers, secrets and prompt injection, and your own patterns, then assign it to a key. A blocked request answers 403 guardrail_blocked; a redacted one carries x-minirouter-guardrail: redacted.

    Read the guardrails docs
  33. API

    Spend caps and disallowed models now answer with a 403.

    A reached daily or monthly spend cap returns 403 spend_limit_exceeded naming when it resets, and a model this key's policy excludes returns 403 model_not_allowed. Rate limits stay 429 with Retry-After, so retry logic keyed on 429 no longer loops on a cap.

    Read the error reference
  34. API

    MiniRouter can fetch pages and tell the time for your model.

    Add minirouter:web_fetch or minirouter:datetime to tools on Chat Completions. MiniRouter runs each call and returns the final answer, with usage and cost totalled. Two plugins join it: one fits a long prompt to the model, one repairs malformed JSON output.

    Read the request parameters docs
  35. API

    Structured output works on 153 models through Chat Completions.

    Send response_format with a JSON Schema and MiniRouter routes only to providers verified to enforce it. Elsewhere the format is dropped and flagged by x-minirouter-structured-output: ignored, or refused with provider.require_parameters: true.

    Read the structured output guide
  36. API

    Save a reusable request setup as a preset.

    Save models, instructions, tools and routing settings under a versioned slug in Dashboard → Presets, then send @preset/your-slug as the model to Chat Completions, Responses or Messages. Fields you send override saved ones, and account and key limits still apply.

    Read the presets docs
  37. Startups

    Startups and developers can apply for API credit.

    Up to $10,000 across three tiers, one form, and a reply within two business days. Approved teams redeem a single code in Billing.

    Startup program
  38. Models

    Auto can be held under a price ceiling per account.

    Set a Low, Medium or High cost tier on Routing and minirouter/auto only picks candidates under that output price. Low also starts bare auto from the cheap list.

    See how the tier applies
  39. Dashboard

    Routing now holds a default model, provider sort and override lock.

    Set the model a request gets when it names none, the provider sort it inherits when it sends none, and whether requests may override provider routing at all. Every key on the account follows them.

    Set routing defaults
  40. Dashboard

    Routing sets the models every key on your account may use.

    A new Routing page holds the account-wide allowed-model list, and a family such as openai/* covers models that author publishes later. Keys keep their own lists and can only narrow the account's.

    Open Routing
  41. Models

    Explore the DeepSeek V4.1 Flash release.

    Compare DeepSeek's latest Flash generation using published coding benchmarks, pricing, and migration notes. The release guide shows its current MiniRouter availability.

    Explore DeepSeek V4.1 Flash
  42. Dashboard

    Balance and model alerts have a new home.

    Set your own low-balance threshold and choose model retirement, price or weekly new-model updates in Alerts. Email delivery is not active yet.

    Configure alerts
  43. Models

    More Qwen, Claude, GPT and Grok models to choose from.

    The catalog adds Qwen 3.8 models, Claude Opus 5 Fast, Mercury 2.5 and more GPT and Grok variants. These additions support text requests.

    Browse models
  44. Chat

    Choose how much room to give a chat response.

    Paid chat defaults to 4,096 output tokens, with a selectable limit up to 16,384, subject to the model's limit. Replies now indicate when they reach the cap.

    Open chat
  45. Integrations

    OpenCode can use more complex tools.

    Nested tool schemas, including OpenCode's question and option fields, no longer fail an overly restrictive depth check. Unsupported streaming-tool combinations also return clearer errors before a request is sent.

    Set up OpenCode
  46. API

    Messages accepts both structured-output formats.

    Use either Anthropic's current or legacy JSON-schema request format with Claude Opus 5 and GPT-4.1 mini on supported routes. Both streaming and non-streaming responses are supported.

    Read the API docs
  47. Rankings

    See which models people use on MiniRouter.

    Daily rankings show eligible token usage, request counts and model-author shares. Compare trends across models using aggregated usage that meets the publication thresholds.

    Explore usage rankings
  48. Models

    Auto Free adds no-deposit API access.

    Auto Free supports no-deposit playground and API requests with a permanent $0 customer charge and shared account limits. Availability and daily limits apply; Auto Free never switches to a paid model.

    Check free availability
  49. Models

    DeepSeek V4.1 Flash Beta joins the catalog.

    DeepSeek's temporary Flash beta supported text requests through MiniRouter and expired on September 10, 2026 at 00:00 UTC. See the permanent V4.1 Flash release for current availability.

    Read the permanent release
  50. Pricing

    Base prices are shown with clear peak-hour disclosures.

    Models with peak pricing now show off-peak rates by default, with the provider's 2× surcharge and schedule called out on model pages, model lists, and the calculator. Published rates include MiniRouter's markup; final charges reflect any applicable peak-hour surcharge.

    Compare model prices
  51. Models

    Darkbloom-only models now appear in the public catalog.

    Three Qwen models already available through the API now appear on the website with their Darkbloom routes and prices. The public catalog recognises models served only by Darkbloom.

    Browse models
  52. Account

    Your last-used sign-in method is easier to find.

    Google, GitHub, and passkey options now show a visible Last used badge, making it easier to recognise how you signed in before.

    Sign in
  53. Account

    Sign in with Google, GitHub, or a passkey.

    Create an account with Google or GitHub, or link them to your existing account. Manage connected sign-in methods and add, rename, or remove passkeys in Settings.

    Manage sign-in methods
  54. Reliability

    Service updates are less likely to be interrupted.

    Overlapping releases no longer interrupt an update already in progress. Service can still be unavailable for several minutes during an update.

    View live status
  55. Models

    Darkbloom is live on /v1: open weights, one field away.

    One request field routes chat completions or Responses to Darkbloom's quantized open-weight models on Apple Silicon, with token rates up to 60% lower than the default route and a 105 µUSD floor per request. Four more models joined, including three only Darkbloom serves.

    Read the Darkbloom guide
  56. Dashboard

    The chat playground is a full page with a searchable model picker.

    Try any catalog model from the dashboard without leaving it, on mobile too. Loading stages are shown as the request moves, and a low balance prompts a top-up in place.

    Open the playground
  57. Payments

    Solana deposits are matched by signature, on chain.

    A deposit lookup now reads the transaction directly instead of waiting for the address scan, and a failed lookup says why.

    View Solana payments
  58. Models

    minirouter/auto picks the model for you.

    Send model "minirouter/auto" (or :cheap, :smart, :code, :roleplay) and the edge resolves one concrete catalog model before pricing, from a public table you can read. The response names the pick.

    See the resolution table
  59. Billing

    Billing problems no longer pause everyone's requests.

    Two short outages on 5 September came from the upstream reconciliation refusing admission. Over-bound provider costs are now absorbed under a budget, upstream failures before the first byte are written off, and reconciliation trouble alerts instead of refusing.

    Read billing behavior
  60. Dashboard

    Activity and Usage gained a yearly heatmap and paging.

    A spend heatmap shows the year at a glance on both pages, and long lists page at a size you choose.

    Open Usage
  61. Support

    A help menu and direct feedback from the app.

    The desktop header has a help menu, and feedback typed into the app reaches the team's Slack immediately.

  62. Models

    GPT-6 Astra joins the catalog.

    OpenAI's new flagship is listed with a 1.05M-token context window. Its model page brings together current MiniRouter prices, published benchmarks, independent measurements, and Responses API setup.

    Explore GPT-6 Astra
  63. Payments

    Pay with MINI, from the normal top-up page.

    The MINI token is a first-class way to add credits: pick it on top-up and the swap and credit happen in one step, with the same zero deposit fee as other crypto.

    See payment options
  64. Dashboard

    Charts, usage tables and API keys were polished.

    Usage charts read at a glance, tables share one register, and key actions fit on small screens.

    Manage keys

August 2026

17 changes

  1. Pricing

    Pricing became faster to compare.

    A searchable live-catalog calculator now leads the page, with current offers and the 0% crypto-deposit fee kept in view.

    Estimate a workload
  2. Reliability

    Provider access is now checked end to end.

    Representative requests verify that each provider can return generated output. Recovery was hardened for stale capacity data, while the paid canary cadence moved to every two hours.

    View live status
  3. Billing

    Billing lag no longer pauses recoverable requests.

    Admission now honours billing grace and bounded reconciliation while keeping the spend controls around every request. Cost convergence, overages and tiny residual balances received the same recovery treatment.

    Read billing behavior
  4. Pricing

    Zero-fee crypto deposits are explicit.

    The public comparison now shows MiniRouter's 0% crypto-deposit fee beside OpenRouter's published 5%, while keeping network costs and token-swap pricing separate.

    See pricing
  5. Platform

    Release bundles now refresh the reverse proxy.

    A release recreates Caddy when its mounted configuration changes, preventing the gateway from serving an older route map. The production image also refreshes its Alpine security packages.

  6. Access

    Keep creating and funding keys during service updates.

    Available payment methods remain usable during service updates, so you can continue creating and funding a key.

    Create a key
  7. Catalog

    Model IDs are tappable on every screen.

    The identifier itself is now the copy control, including on touch devices. Confirmation is written and announced instead of depending on an icon that only appears on hover.

    Browse models
  8. Interface

    Tables now share one visual register.

    Six header treatments and eight container styles became one accessible table system. Redundant rules were removed across the public site and dashboard so data carries the hierarchy.

  9. Payments

    The minimum supported deposit moved to $0.50.

    Deposit capabilities, payment copy and validation now agree on the lower floor. The API host root also points clients to the LLM API instead of leaving discovery at a dead end.

    Review deposit rules
  10. Discovery

    More clients can discover MiniRouter.

    Client metadata and integration guidance expanded for models.dev, Continue and LiteLLM-style consumers, with reasoning options carried through the catalog contract.

    Choose a client
  11. Rankings

    Coding-model rankings became a first-class reference.

    The new ranked boards separate source-led benchmark references from routable catalog data, with the underlying board data also available as JSON and CSV.

    Open rankings
  12. API

    Anthropic Messages gained tools, thinking and images.

    The Messages surface now accepts structured content blocks, tool loops, images, thinking and cache controls. Authenticated token counting and Anthropic SDK parity tests cover the same protocol.

    Read the API docs
  13. Models

    Models gained release dossiers and real brand marks.

    The catalog sorts by publisher release date, uses local brand artwork and links to source-reviewed release dossiers that keep launch claims separate from MiniRouter availability.

    Explore releases
  14. Operations

    Missing catalog updates are detected automatically.

    Hourly checks flag persistent differences between our model list and the provider catalog so we can investigate missing updates.

  15. Payments

    Solana monitoring stopped scaling with every old address.

    Frequent scans now follow recently active deposit addresses while a full six-hour reconciliation preserves dormant coverage. Retry caps keep one bad observation from amplifying RPC load.

    View Solana payments
  16. Experience

    Quick Start shipped, and mobile overflow was removed.

    The dashboard gained a direct first-request path. A viewport audit then fixed page-level horizontal scrolling across 27 routes at 390, 360 and 320 pixels.

    Open dashboard
  17. Guides

    Personal-agent and roleplay guides launched.

    Ten workload and integration pages now cover spend limits, cost-per-message assumptions and setup for OpenHands, RisuAI, Agnai and other clients.

    Browse guides

July 2026

3 changes

  1. Product

    The full public and dashboard surface came online.

    Models, providers, comparison, pricing, payments, legal, documentation, status and integration routes joined the accountless key and nine-page dashboard flows.

    Start with the docs
  2. Foundation

    The complete request and money path passed end to end.

    A cold local stack proved accountless key creation, idempotent credit, streaming inference, reserve and settlement accounting, compatibility endpoints and the zero-balance error path.

  3. Backend

    Gateway, ledger, pricing and payment rails were joined.

    The edge relay, exact-decimal pricing, double-entry credit ledger, provider health, deposit rails and dashboard API were wired into one tested system rather than remaining isolated packages.

See what routes right now.