New models, features and improvements on Flux Router. Because you target stable aliases, most of these reach you without a code change.
Pricing
flux-auto now bills the Fast rate when a Fast model answers
When flux-auto sends a request to the Fast lane and a Fast model answers it, you now pay the Fast rate instead of Standard.
Everything else bills at the Standard rate, as before. flux-auto never costs more than Standard.
Want the Fast rate on every simple request? Send flux-fast yourself. Current rates are on the pricing page.
Feature
A new fluxrouter.ai
A new homepage, with an animated board that shows how one request travels through Flux.
Abilities: one page for each thing your key can do, from models to the live web, voices, images and apps, with what is live and what is still shipping.
Use cases: nine worked examples of real jobs, step by step, with the agent you already use.
Agents: setup pages for Murage, Wayland, Claude Code, Codex, Cursor, Hermes, OpenClaw and Cline, plus a page for any OpenAI-compatible tool.
Compare: honest side-by-side pages for OpenRouter, Vercel AI Gateway and going direct to each provider.
Models by lab, and a directory of the apps your agent can reach.
Feature
A new dashboard
Your first login now walks you through three steps: your key, then credit or a plan, then connecting your tool.
Overview and Billing show your balance and spend at a glance, and a top-up is now a true one-off payment.
Start pay-as-you-go in one step, top up in one click, and switch on automatic top-ups in one click.
A new Usage section: pick a date range, break spend down by model, key and day, and export it as CSV.
Keys and Connect are rebuilt, and the key dialog now states its limits plainly.
Activity and Routing show each call and which model answered it.
The link in your confirmation email now takes you straight to your dashboard.
Feature
Docs rebuilt around your agent
The docs now start from the agent you use, with a guide to the Flux tools server.
New in-depth guides for speech to text, text to speech, web search, reading a page, images, second opinion and apps.
Cleaner docs addresses. Every old docs link still works and takes you to the new page.
Models
Two GLM pins retired
flux-pinned-glm-5 and flux-pinned-glm-4-7 no longer answered, so new keys no longer receive them.
flux-pinned-glm-5-1 is unaffected, along with every newer GLM pin.
Fix
Clearer, more accurate pricing and trust copy
The site now says exactly how flux-auto, pinned models and each plan bill.
The cost header is described accurately: it comes on non-streaming responses, and streamed responses report cost in usage.
Our privacy wording now matches what we do: we never store your raw prompt, our logs keep a scrubbed copy, and Flux never trains on your data.
The FAQ lists every endpoint: images, speech, transcription, search and fetch.
Models
DeepSeek V4.1 Flash: one canonical pin, native vision
flux-pinned-deepseek-flash is the pin for DeepSeek V4.1 Flash. Its vision is native and generally available: no separate experimental pin, and image input costs the same as text.
DeepSeek has retired the models behind the v4-flash names. flux-pinned-deepseek-v4, flux-pinned-deepseek-v4-flash and flux-pinned-deepseek-v4-flash-vision-exp keep working on every key that holds them and now serve V4.1 Flash; nothing you have configured needs to change.
flux-pinned-deepseek-flash-turbo (thinking off) and flux-pinned-deepseek-flash-max (thinking at max effort) are unchanged, as is flux-pinned-deepseek-v4-pro.
Select flux-pinned-deepseek-v4-flash-vision-exp for experimental image understanding. Automatic routing and the existing text-only Flash pin are unchanged.
The experimental vision pin uses the same pricing policy as the existing DeepSeek V4 Flash pin.
Available to paid keys with the standard model catalogue. Deliberately restricted keys keep their selected model list.
Models
GPT-6 Astra: an optional paid pin for text generation
Select flux-pinned-gpt-6-astra explicitly. Automatic routing pools do not select Astra.
Astra supports text generation through FluxRouter. Tool-bearing requests are rejected on both Chat Completions and Responses; changing endpoints does not enable agent tool use.
Standard service only, billed at provider cost plus 12%. Long-context and cache rates apply to the usage reported by the provider; this is not a single flat token rate.
The model supports 1,050,000 total context tokens, with at most 922,000 input tokens and 128,000 output tokens.
Paid keys with the standard model catalogue can use the pin. Deliberately restricted keys retain their selected model list.
Models
Claude Fable 5.1, Qwen 3.8 Max, and Gemini 3.8 Flash are live
Choose Claude Fable 5.1 with flux-pinned-claude-fable-5-1.
Choose Qwen 3.8 Max with flux-pinned-qwen3-8-max.
Choose Gemini 3.8 Flash with flux-pinned-gemini-3-8-flash.
Use these pinned aliases in your request’s model field when you want a specific model.
Models
GLM-5.3 is live, and it’s pin-only
GLM-5.3, Z.ai’s newest GLM, is live on the gateway, on the same pinned pricing GLM-5.2 carries today. If GLM is already the model you reach for on tool-heavy and long-running coding work, this is the newer one.
Read this bit twice, because it’s different from every other model we’ve added: 5.3 is reached by NAME ONLY. Call it as flux-pinned-glm-5-3 and you get it every time. flux-auto, flux-standard and flux-fast will not route to it, so leaving your code untouched gets you nothing here. If you want 5.3, ask for 5.3.
GLM-5.2 and every older GLM pin are untouched and stay pinnable exactly as before. An addition, not a swap: nothing was retired and nothing changed lanes.
It bills like any other pin: cost-plus over the model’s own cost, through the same metered gateway with the same transparency headers on every response.
Models
Gemini 3.7 Flash joins the Fast tier
Gemini 3.7 Flash, Google’s newest flash model, is live in the Fast lane, with the full 1M-token context window.
It runs at roughly a third of what a frontier mid-tier model costs, and still takes a million tokens of context, so the long-context work you used to send to an expensive model now has a cheap home.
Nothing to change in your code: keep calling flux-auto or flux-fast and you pick up 3.7 automatically. Pin it directly with flux-pinned-gemini-3-7-flash.
Gemini 3.6 and 3.5 Flash stay pinnable exactly as before. 3.7 takes the Fast lane’s rotation slot, but nothing was retired.
Models
Grok 4.6 joins the Reasoning tier
Grok 4.6, xAI’s newest flagship, is live in Reasoning, with the full 500K context window.
Grok 4.5 stays exactly where it was. This is an addition, not a replacement: 4.5 has the cheaper cached-input rate, so it remains the better pick for prompt-heavy work you send repeatedly.
Nothing to change in your code: keep calling flux-auto or flux-reasoning and you pick up 4.6 automatically. Pin it directly with flux-pinned-grok-4-6.
Models
New frontier models across every tier
GPT-5.6, OpenAI’s newest family (Sol, Terra, Luna), is live: Sol in Reasoning, Terra in Standard, Luna in Fast.
Grok 4.5, xAI’s latest flagship, joins the Reasoning tier.
Kimi K3, Moonshot’s trillion-scale reasoning model, joins the Reasoning tier.
MiniMax M3: frontier-class reasoning at a fraction of the price, added to Reasoning and Standard.
Nothing to change in your code: keep calling flux-auto or a flux-* lane and you pick up the new models automatically. Pin any of them directly with its flux-pinned-* alias.
