Together AI announced hosted access to Qwen4 Max on November 21, 2026, giving small app teams another route to large-document and coding workloads. A new prepaid organization must buy at least $15 of credits to start; that money funds usage, rather than adding an activation fee. For an illustrative 100 calls with 500,000 uncached input tokens and 20,000 output tokens each, reasoning included, consumption is $84 at Together’s published list rates or $58.80 at its advertised promotion.
Those estimates use Together’s pricing page as checked on November 21. Public pages disagree on tariffs and do not establish a precise promotion cutoff, so confirm the applicable rate before committing a larger budget.
What must you pay before starting?
Together’s credit rules say there is no general free trial, and free credits do not satisfy the initial $15 purchase requirement. Until that purchase, the console restricts playground access, API-key creation and teammate invitations. Later purchases have a $5 minimum. Prepaid use needs a positive balance; customers with contracts follow their agreed terms.
A deposit is not an exact spending limit. Together’s spend-control documentation says delayed metering can allow charges beyond a zero balance, while auto-recharge keeps buying credits. For a bounded pilot, inspect existing recharge settings and limit calls in the application. The organization’s balance is shared, so a separate project does not isolate its spending.
If your priority is a free evaluation, see Qwen4-Max’s QwenCloud free API quota. That is a separate provider’s eligibility-based allowance, not a Together entitlement.
Which prices should you budget with?
Tokens are the units used to count model input and output. The following published rates are in US dollars per million tokens. Use the list-price column for planning beyond the short promotion, subject to confirmation of the rate applicable to your account.
Usage | Published list rate | Advertised promotion, November 21 |
|---|---|---|
Uncached input | $1.50 | $1.05 |
Cached input | $0.075 | $0.0525 |
Output, including reasoning | $4.50 | $3.15 |
There are two qualifications. The announcement says the 30% offer runs through November 22, 2026; the pricing page says it is valid until November 23. Neither specifies the hour or timezone. Do not assume the latter wording guarantees an extra day.
Separately, the serverless model catalogue lists $1.425 for input, $0.07125 for cached input and $4.275 for output per million tokens. These differ from the pricing page; the public sources do not resolve which rate an account will actually be charged. The calculations here use the explicitly labelled list and promotional columns above, not a blend of the two pages.
What would 100 calls cost?
These are illustrative text-only calculations, not measured bills. Assume exactly 100 calls, with no cached input. Count all instructions, document text and conversation history as input, and reasoning plus the answer as output.
Short-request pilot: 5,000 input and 1,000 total output tokens per call consume $1.20 at list rates, or $0.84 at the advertised promotional rates.
Large-document pilot: 500,000 input and 20,000 total output tokens per call consume $84 at list rates, or $58.80 at the advertised promotional rates.
The large pilot totals 50 million input and 2 million output tokens. Using the published tariffs, its list-price calculation is 50 × $1.50 + 2 × $4.50 = $84. At the advertised promotion it is 50 × $1.05 + 2 × $3.15 = $58.80.
These totals exclude retries, extra agent turns, external tools, hosting, human review, taxes and currency conversion. One call is not necessarily one completed customer task. The $15 purchase funds consumption: do not add it again to the totals. Conversely, a fresh account still needs that purchase even if its short pilot consumes only $0.84.
When does caching change the budget?
Prompt caching reuses an unchanged beginning of a request. Put stable instructions and reference documents before the changing question. Together applies caching automatically, but repeated content is not a guaranteed cache hit. Entries typically last about an hour and can be removed sooner under load; check actual cached-token usage.
As a sensitivity calculation, suppose 80% of the large pilot’s input really receives cached pricing across all 100 calls, including cold calls. That means 40 million cached input tokens, 10 million uncached input tokens and the same 2 million output tokens. Consumption becomes $27 at list rates, or $18.90 at the advertised promotional rates.
In that assumed scenario, reuse saves $39.90 against the $58.80 promotional uncached budget—more than the $25.20 promotion saving on the uncached workload itself. This shows why actual reuse matters to the budget; it does not predict an app’s cache-hit rate. Start with the uncached estimate until usage supports a lower one.
Which limits could stop your pilot?
Streaming is mandatory. Together’s model instructions require
stream: true; a non-streaming request returns HTTP 400 withstreaming_required. Streaming delivers the response in pieces, so a client that only accepts a complete response needs adapting.Thinking uses the output budget. Together says it is on by default and can be disabled with
reasoning: {"enabled": false}. Its reasoning documentation describesmax_tokensas a ceiling for reasoning plus the answer. Too little room can truncate a response. Do not add a reasoning breakdown again if it is already included in total output usage.Capacity is shared. Serverless performance is best-effort, with possible 429 or 503 errors and model-specific restrictions during high demand. The documentation does not give a fixed Qwen4 Max requests-per-minute entitlement. Spread calls out and use a limited number of retries with progressively longer waits.
Dedicated hosting and fine-tuning are not available for Qwen4 Max in this offer, according to the announcement. It is also absent from the published reserved-throughput model list. Contacting sales about an unlisted model is an enquiry, not confirmation that reserved capacity is available.
The model page lists a one-million-token context, text and image input, text output and the identifier Qwen/Qwen4-Max. That context figure is not an output allowance. Together-specific maximum output, regional eligibility and batch eligibility are not established by the sources above; do not import another provider’s limits or assume a batch discount.
Is Together a fit for your evaluation?
This route fits a funded pilot with a streaming-capable client and tolerance for shared capacity. Before expanding it, compare actual billed tokens and costs with the number of usable results. A low per-call estimate alone does not establish low cost per finished task.
If you are comparing long-context services, the Gemini 4 Neon API cost and output-limit guide covers another provider’s reasoning and cache budget. Compare the same workload and required features; token prices alone do not establish which model produces better results.
