AI replies that flow in while they are being written

Fluxio relays Claude, GPT, Gemini and a dozen more model families as live token streams — made for chat apps, assistants and anything real-time.

New accounts start with $0.12 of credit.

Replay of a streamed answer

PromptGive me a one-line tagline for a live chat widget.

GPT-6.1 Sol · gpt-6.1-solAnswers that start arriving before your customer finishes reading the question.

PromptIn two sentences, why do server-sent events suit chat replies?

Gemini 2.5 Flash · gemini-2.5-flashThey keep one HTTP response open and push each new piece of text the moment it exists. Browsers read them natively, so a chat window can paint words as they are generated.

PromptSuggest a friendly loading message for a streaming bot.

Claude Haiku 4.5 · claude-haiku-4-5-20251001Thinking out loud — the first words are on their way.

PromptName three things a real-time assistant should never do.

Grok 4.7 · grok-4.7Freeze without any sign of progress, bury the answer under filler, and keep talking after the user has already moved on.

PromptTranslate "the stream is live" into French and Spanish.

MiniMax M3 · MiniMax-M3French: « le flux est en direct ». Spanish: «la transmisión está en vivo».

— chunks have arrived in this reply

A recorded answer replayed in your browser at streaming pace.

How a stream is billed

A streamed reply is priced like any other call: by the tokens it used. What differs is where the numbers come from and what happens when a stream ends early.

The usage frame
OpenAI-style streams report usage in a final chunk, and Fluxio switches stream_options.include_usage on for every streamed chat so that chunk always arrives. Anthropic streams carry the counts in message_start and message_delta; Gemini puts them in each chunk's usageMetadata.
Hanging up midway
If your client closes the connection, Fluxio stops the upstream at once and bills only what was already generated. When no usage frame had arrived yet, input is estimated from the request size and output from the text that already streamed, and the record is marked as estimated.
Calls that cost nothing
Token-counting endpoints are free, and replies that end in an upstream error are recorded at zero. HTTP 402 means the balance or monthly plan behind the key is empty.

For example, GPT-6.1 Sol costs $0.1458 per million input tokens and $0.7292 per million output tokens here; the official list price is $2.00 and $10.00.

Monthly plans

Prefer a fixed monthly spend? Each plan covers Claude or GPT calls for 30 days, with a weekly ceiling.

Claude · $1100 allowance · 30 days · $350 a week

$62.36 / valid for 30 days

Plan allowance: $1100
Most you can use per week: $350
Plan multiplier: × 2.8

Claude models only; a dedicated API key is issued with the plan.

Claude · $2900 allowance · 30 days · $850 a week

$165.14 / valid for 30 days

Plan allowance: $2900
Most you can use per week: $850
Plan multiplier: × 2.8

Claude models only; a dedicated API key is issued with the plan.

GPT · $2600 allowance · 30 days · $800 a week

$44.31 / valid for 30 days

Plan allowance: $2600
Most you can use per week: $800
Plan multiplier: × 2.8

GPT models only, every Codex-ready GPT id included; a dedicated API key is issued with the plan.

GPT · $8000 allowance · 30 days · $2400 a week

$137.36 / valid for 30 days

Plan allowance: $8000
Most you can use per week: $2400
Plan multiplier: × 2.8

GPT models only, every Codex-ready GPT id included; a dedicated API key is issued with the plan.

Three questions about streaming

Does Fluxio buffer the stream before passing it on?

No. Every complete SSE line is forwarded the moment it reaches us; only a half-received line waits for its line break, so nothing is held back longer than it takes to arrive.

My client sets include_usage to false. Will the charge be a guess?

No. For streamed chat completions Fluxio asks the upstream for the usage frame whatever that setting says, so the charge follows the real token counts.

Can I keep several streams open at once?

Yes: up to eight per account at the same time, and each key may start up to 120 requests a minute. A ninth parallel stream receives HTTP 429 until one of the others finishes.