​

Billing address and tax ID each have one place to edit

• Edit the billing address in the Stripe billing portal (Manage billing info).
• Edit the tax ID on the billing account page.
• An EU billing account takes an address in an EU-27 country only.
• The keys list hides revoked keys by default. Filter by status and sort by any column.

​

Admission holds balance for the largest output cap

• We hold balance for the largest output cap a request carries: max_tokens, max_completion_tokens, or either inside extra_body. When that cap is more than the key’s balance covers, the request gets 402 insufficient_balance.

​

The Activity chart follows your time zone

• The dashboard Activity chart buckets by your local time zone and shows quiet periods as zero.
• Invoices and the usage export stay on UTC.

​

Member spend caps count a rolling 90 days

• A member spend cap counts spend over a rolling 90 days, labelled Rolling 90 days.
• A key blocked only by older spend reopens as that spend ages out.

​

Requests and streams run up to 30 minutes

• A request or stream runs for up to 1800 s unless the model sets a shorter limit.

​

Usage presets cover whole UTC days

• The 7d and 30d presets on the Usage page cover whole UTC days.
• Tokens and cost are counted over the same window.

​

Per-token rates changed on one model

​

Two models leave serverless on 2026-10-28

​

Three models added to serverless

​

DeepSeek-V4.1-Flash-1017 is on serverless

​

Per-token rates changed on five models

​

We bill only for the output a request produced

• We bill a generation on its output tokens, counted as the greater of completion and reasoning tokens. Embeddings and rerank are billed on the prompt.
• A failed call is not billed. Nor is a request rejected for exceeding the context window.
• Auto-recharge turns off when the last payment method is removed.

​

reasoning_effort rejects unknown words, and rounds up where a model has no dial

• We reject a reasoning_effort value that is not an effort level with 400. The vocabulary is minimal, low, medium, high, xhigh, max and none.
• On DeepSeek-V4-Flash-0731, minimal is the off switch. low and medium round up to high and are reported in x-flexai-adjusted-params: reasoning_effort=high — reasoning=high for OpenRouter’s reasoning.effort and for /v1/responses.
• We reject an image sent to a model that cannot read one with 400. The response was previously 429, which SDKs retry.

​

Rates changed on sixteen models

​

Context-window rejections carry context_length_exceeded

• An engine context-window rejection is normalised to an invalid_request_error with code context_length_exceeded, and param set to messages, input or prompt to match the endpoint.
• The context window and weight precision in /v1/models are the ones the fleet is serving, which can differ from the catalog’s.

​

Request account deletion from the dashboard

​

Video results are single-download and expire after an hour

• Downloading a clip deletes it and the prompt that made it. A second request is refused with 410 job_already_retrieved.
• An uncollected clip expires after one hour. A request for the job afterwards is refused with 410 job_expired.

​

Usage lookup by request ID works again

​

reasoning_effort on hybrid models, tool calling on the Responses API

• reasoning_effort turns reasoning on for GLM-4.5-Air-FP8, NVIDIA-Nemotron-3.5-Lightning-30B-A3B, Qwen3-8B-FP8, Qwen3.5-9B, Qwen3.6-27B-FP8, Qwen3.6-35B-A3B-FP8, gemma-4-26B-A4B-it and gemma-4-31b-it. low, medium, high, xhigh and max turn it on; none and minimal leave it off. Depth is not adjustable on these models, so x-flexai-adjusted-params: reasoning_effort=on is set on the response — reasoning=on for OpenRouter’s reasoning.effort and for /v1/responses.
• thinking: {"type": "disabled"} and enable_thinking: false in chat_template_kwargs win over reasoning_effort. The dropped effort is reported in x-flexai-dropped-params.
• Tool calls from GLM-4.5-Air-FP8 come back as function_call items on non-streaming /v1/responses with thinking off. A reply that is only a tool call comes back as a message item plus one function_call per call; anything else is passed through as the engine sent it, including a forced tool_choice and a reply cut off at max_output_tokens.
• The call is included for a forced tool_choice on streaming /v1/responses.
• tool_choice: "required" on GLM-5.2, MiniMax-M2.7 and NVIDIA-Nemotron-3.5-Lightning-30B-A3B sets parallel_tool_calls: false, named in x-flexai-dropped-params. Sending false, or omitting it, adds no header.
• An unrecognised /v1/responses input item type is rejected with input[i].type naming the item. reasoning, item_reference and hosted-tool calls pass through.
• DeepSeek-V4-Flash-0731 may fill top_p 0.95 when you send none. Any value you send takes precedence, including temperature: 0. Filled fields are named in x-flexai-sampling-defaulted.

​

Billing emails follow your top-ups, and EU accounts bill in euros

• Low-balance emails key off your last top-up: “running low” below a fifth of the balance after that top-up, “out of credit” at $0. Each top-up re-arms both.
• With auto-recharge on, only the “out of credit” email is sent.
• Postpaid budget alerts fire at 80% and 100% of the cap.
• There is no “card added” confirmation email.
• EU accounts are charged and invoiced in euros. Prices, usage and balance stay in US dollars; a euro top-up converts once, at charge time, at a reference rate fixed for the calendar month. Regions.

​

New model: Qwen3.8 Flash Next

curl https://api.flex.ai/v1/chat/completions \
  -H "Authorization: Bearer $FLEXAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "Qwen3.8-Flash-Next", "messages": [{"role": "user", "content": "Hello"}]}'

​

Token Factory changelog

​

New models: GLM 5.3 Flash and Qwen3.8 27B

• GLM 5.3 Flash — $0.075 input, $0.25 output per million tokens, 1M context.
• Qwen3.8 27B — $0.42 input, $3.00 output per million tokens, 262K context.

curl https://api.flex.ai/v1/chat/completions \
  -H "Authorization: Bearer $FLEXAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "GLM-5.3-Flash", "messages": [{"role": "user", "content": "Hello"}]}'