flexinference 1.6.3

Official Python SDK for FlexInference - a deadline-aware, OpenAI-compatible inference router.

FlexInference (Python)

The official Python SDK for FlexInference. FlexInference is an inference router that works with OpenAI, Google Gemini, and Anthropic. You send the OpenAI-shaped requests you already send, you bring your own provider key, and you set one required field called start_within. That field is how long you will wait for the request to start. We try a cheaper tier first to lower your bill, and if it cannot start in time we escalate to your standard tier so the request still runs. The first 100,000 requests each month are free; a $10/month subscription raises the allowance to 10,000,000. The SDK speaks four caller formats: responses, chat.completions, interactions (Gemini shape), and messages (Anthropic shape). Any of them reaches any provider.

pip install flexinference

Quickstart

from flexinference import FlexInference, output_text

client = FlexInference(api_key="flex_live_...")

res = client.responses.create({
    "model": "gpt-5.5",
    "input": "Write a haiku about cheap GPUs.",
    "start_within": "00h-00m-30s",
})

print(output_text(res))

Responses come back as the raw OpenAI JSON and we never reshape the body. That means there is no output_text field on the wire, because OpenAI's own SDKs compute that field rather than the provider. output_text(res) pulls the assistant's text out of either a response or a chat completion for you.

start_within is required on every request. Set it to "default", "priority", "auto", or a duration written as "HHh-MMm-SSs" from 5s to 10m. A duration is how long you will wait for the request to start running. We try OpenAI's cheaper flex tier first on a flex-capable model, and that is where your savings come from. If flex cannot start inside your window, we switch to your normal standard tier so the request still completes. The words "default", "priority", and "auto" map straight to those OpenAI service tiers and work with any model. See the docs.

This escalation is your safety net. You never lose a request just because the cheap tier was busy. Your standard tier always finishes the job. It runs the same model either way, so you trade a little waiting for a lower bill and keep the result you would have gotten anyway.

Providers (OpenAI, Gemini, and Anthropic)

FlexInference routes to OpenAI, Google Gemini, and Anthropic. Send the same OpenAI-shaped request and pass whichever model id you want, such as gpt-5.5, o4-mini, gemini-3.5-flash, or claude-opus-4-8. We translate Gemini and Anthropic to and from the OpenAI shape, so your code is identical for all three.

• OpenAI: default (standard tier), priority, auto, and the flex race (a duration) on flex-capable models.
• Gemini: default maps to Gemini's standard tier, plus priority and the flex race on the Gemini flex models (gemini-3.5-flash, gemini-3.1-flash-lite, gemini-3.1-pro-preview, gemini-3-flash-preview, gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite). Gemini has no auto tier, so start_within="auto" on a Gemini model returns 400.
• Anthropic (Claude): proxy-only. default, priority, and auto work; there is no flex race, so a duration start_within on a claude-* model returns 400 flex_unsupported_for_anthropic. Anthropic requires an output-token field upstream. FlexInference forwards max_output_tokens (max_completion_tokens on Chat, max_tokens on Messages) when set and does not synthesize a default, so omitting it returns Anthropic's own request error. You keep the unified API and tier control, and draw down your own Anthropic credits.

Add the provider key you'll use (OpenAI, Gemini, and/or Anthropic) in the dashboard. Text, streaming, structured outputs, function calling, image input, and web search work across the openai, google, and anthropic direct routes (send a Responses web_search tool; we map it to Gemini's google_search).

Pin the route or set a same-model fallback chain with the optional provider array (for example ["openai"] or ["google", "vertex"]): element 0 is the primary route and the sole flex gate, and later entries are tried in order. A malformed provider array raises a ValueError locally before any network call. All five routes serve: google, openai, and anthropic are direct routes, and the cloud routes bedrock (claude-*, via Amazon Bedrock) and vertex (gemini-*, via Google Vertex AI) serve the same model using your own cloud key. Cloud routes run at the standard tier only, with no flex. The vertex route serves text; keep cloud requests text-focused.

Don't send service_tier. The router picks the tier from start_within, so a request that sets its own service_tier fails fast with 400 service_tier_not_allowed.

Streaming

stream = client.responses.create(
    {"model": "gpt-5-nano", "input": "Count to ten.", "start_within": "00h-00m-20s"},
    stream=True,
)
for event in stream:
    if event.get("type") == "response.output_text.delta":
        print(event["delta"], end="")

Chat Completions

res = client.chat.completions.create({
    "model": "gpt-5.5",
    "messages": [{"role": "user", "content": "Hello!"}],
    "start_within": "default",
})
print(res["choices"][0]["message"]["content"])

Interactions (Gemini shape)

Speak Google's Interactions shape and reach any model. interaction_output_text(res) pulls the assistant text out of the interaction's steps.

from flexinference import interaction_output_text

res = client.interactions.create({
    "model": "gemini-3.5-flash",
    "input": "Summarize this contract.",
    "start_within": "00h-01m-00s",
})
print(interaction_output_text(res))

Messages (Anthropic shape)

Speak Anthropic's Messages shape and reach any model. Anthropic requires max_tokens upstream. FlexInference forwards it when set and does not synthesize a default, so omitting it returns Anthropic's own request error. message_output_text(res) pulls the assistant text out of the message content.

from flexinference import message_output_text

res = client.messages.create({
    "model": "claude-opus-4-8",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "Summarize this contract."}],
    "start_within": "default",
})
print(message_output_text(res))

Closing the client

The client holds a pooled httpx.Client, so close it when you're done to release connections. Use it as a context manager:

with FlexInference(api_key="flex_live_...") as client:
    res = client.responses.create({"model": "gpt-5.5", "input": "Hi.", "start_within": "default"})
    print(output_text(res))
# connections are released on exit

Or close it yourself:

client = FlexInference(api_key="flex_live_...")
try:
    ...
finally:
    client.close()

Timeouts

Streaming and non-streaming requests are timed differently, so a long generation is never cut off yet a hung request never hangs forever:

• Non-streaming requests bound only the body read with timeout, default 600s - the read after the first response has arrived, not a total wall-clock budget (the worst case is roughly connect_timeout + first_byte_timeout + timeout).
• Streaming requests have no total cap (a healthy stream runs as long as tokens keep arriving). A stalled stream is caught by an idle watchdog: if no chunk arrives within idle_timeout (default 60s), the request is aborted with a typed FlexInferenceError (code idle_timeout).
• Both wait up to first_byte_timeout (default 60s) for the first response, and connect_timeout (default 10s) to reach the router. The first-byte wait is auto-raised for a flex start_within (the router withholds headers until the race resolves), so a long deadline just works.

client = FlexInference(
    api_key="flex_live_...",
    timeout=120.0,  # non-streaming total budget
    first_byte_timeout=90.0,  # wait for the first response
    idle_timeout=30.0,  # max silence between streamed chunks
)

Any of these can be overridden per request:

res = client.responses.create(
    {"model": "gpt-5.5", "input": "Hi.", "start_within": "priority"},
    timeout=30.0,
)

for event in client.responses.create(
    {"model": "gpt-5.5", "input": "Count to ten.", "start_within": "priority"},
    stream=True,
    idle_timeout=15.0,
):
    ...

A custom httpx.Client is still accepted for transport, auth, or SSL config; the timeouts above are applied per request and override the client's own timeout.

Request validation

Before a request leaves your machine, the SDK validates the parts it owns. start_within is required and must be "default", "priority", "auto", or a duration "HHh-MMm-SSs" between 5s and 10m; model and input/messages must be present. A missing or bad value raises a ValueError locally instead of making a round trip to a provider 400:

client.responses.create({"model": "gpt-5.5", "input": "hi"})
# ValueError: Invalid request body:
#   Missing required parameter: `start_within`. Set it to "default", "priority", "auto", or a duration "HHh-MMm-SSs".

Validation is request-only. Unknown fields pass straight through to the provider (so new OpenAI parameters keep working), and responses are never validated or reshaped.

Retry

Pass an optional retry and the router re-hits the upstream on a transient transport failure, so a request that would otherwise error because the tier was briefly down instead quietly succeeds:

res = client.responses.create({
    "model": "gpt-5.5",
    "input": "Write a haiku about cheap GPUs.",
    "start_within": "00h-00m-30s",
    "retry": {"count": 3},  # up to 3 re-hits; backoff "exponential", jitter True
})

retry is {"count", "backoff"?, "jitter"?}. count is required and sets how many times the router re-hits the settled tier (1 to 5). backoff is "exponential" (the default) or "linear", and jitter (default True) spreads each delay so clients don't retry in lockstep. The object is validated locally before any network call -- a missing or out-of-range count, a bad backoff, a non-boolean jitter, or an unknown sub-key raises a ValueError.

Retry fires only on a genuine 429 or 5xx, and only before the response commits (its status line, or the first streamed byte); it honors the provider's Retry-After first (clamped to 60s). A client error, an auth failure, an upstream timeout, or a stream that has already started is never retried, and only the tier the router settled on is re-hit (never the flex leg during a flex race). The response carries an x-flexinference-retries header with the count the router spent, so if your own client also retries you can turn it down to avoid multiplying attempts.

Errors

Non-2xx responses raise FlexInferenceError, carrying status, type, code, and param. The router shapes error bodies to match the endpoint you called (OpenAI on responses/chat, Anthropic on messages, Google on interactions) so the SDK you would use for that surface parses them, and FlexInferenceError reads all three. message and status are always set; code, param, and doc_url are populated on the OpenAI surface.

from flexinference import FlexInferenceError

try:
    client.responses.create({"model": "gpt-5.5", "input": "hi", "start_within": "priority"})
except FlexInferenceError as err:
    if err.code == "no_byok_key":
        print("Add your OpenAI key in the dashboard.")
    else:
        raise

Every FlexInference error tells you the same four things. It says what went wrong, why it went wrong, how to fix it, and it shows an example of a request that works. This is built for agents as much as for people. An agent can read the message and correct the call instead of guessing and burning tokens. Provider errors are reshaped into the same surface envelope and normalized into FlexInferenceError, so you get one consistent error type no matter which model ran.

Billing / 402

The first 100,000 requests each month are free - bring your own provider key, no card required. A $10/month subscription raises the monthly allowance to 10,000,000 requests. Past that, buy overage at $1 per extra 1,000,000 requests. Pre-buy a block, or turn on auto-charge to top up automatically up to a cap you set. The allowance counts every request, not just flex.

When you run out of allowance, the router returns 402 Payment Required on every request until the allowance resets on the 1st of the month (UTC) or you buy more. The SDK raises a typed PaymentRequiredError (a subclass of FlexInferenceError) for HTTP 402, and its .code tells you which case you hit and what to do:

• free_quota_exceeded - a free org used its 100,000 requests for the month. Subscribe ($10 for 10,000,000/month) or wait for the 1st-of-month UTC reset.
• quota_exceeded - a subscription used its 10,000,000 requests plus any purchased overage, with auto-charge off. Buy more requests or enable auto-charge, or wait for the reset.
• payment_required - an auto-charge overage payment failed. Update the card on the Billing page to clear it.

from flexinference import PaymentRequiredError

try:
    client.responses.create({"model": "gpt-5.5", "input": "hi", "start_within": "00h-00m-30s"})
except PaymentRequiredError as err:
    if err.code == "free_quota_exceeded":
        print("Free monthly allowance used up - subscribe or wait for the 1st-of-month reset.")
    elif err.code == "quota_exceeded":
        print("Monthly allowance used up - buy more requests or enable auto-charge.")
    else:  # payment_required
        print("Auto-charge payment failed - update your card on the Billing page.")
except FlexInferenceError:
    raise

Because PaymentRequiredError subclasses FlexInferenceError, existing except FlexInferenceError handlers keep catching 402s too.

Configuration

Argument	Default	Description
api_key	(required)	Your flex_live_ key.
base_url	https://api.flexinference.com/v1	Override the router endpoint.
client	new httpx.Client	Provide your own httpx.Client (transport/auth/SSL); its timeout is overridden per request.
timeout	600.0	Non-streaming body-read budget in seconds, after headers (streaming has no total).
first_byte_timeout	60.0	Wait for the first response in seconds; auto-raised for a flex start_within.
idle_timeout	60.0	Max silence between streamed chunks (seconds) before a stream is treated as hung.
connect_timeout	10.0	Wait to reach the router in seconds.

timeout, first_byte_timeout, and idle_timeout can also be passed per create() call.

License

MIT