GPT-6 Astra Pricing on Quanta: $0.8333 per 1M Input Tokens

01 · What does a large classification run cost on Quanta?

A large classification run is rarely a single request. It is usually a sustained sequence: many records to label, each record turned into a prompt, each answer checked, and the whole job repeated as the dataset grows. That makes cost predictability and request handling more important than a headline price alone.

Quanta is an AI API relay designed for teams with heavy, sustained workloads. You send requests through one balance, and every call is metered on real tokens. Input, output, cache writes and cache reads are priced separately, so the dashboard can show the tokens and charge for each individual call rather than hiding the details inside an aggregated report.

For example, GPT-6 Astra is listed at $0.8333 per 1M input tokens, $4.19 per 1M output tokens and $0.0833 per 1M cache-read tokens. There is no monthly fee and no minimum spend. Quanta uses a prepaid balance, and a new account starts with $0.50 in trial credit.

That structure suits classification work in particular. If a run uses a fixed prompt template, the input side can be estimated before the job starts. If the job benefits from repeated context, cache-read pricing gives you a second, visible lever. And because the figures are published per 1M tokens, a team can build a working budget without waiting for an invoice to explain where the money went.

02 · How does token metering translate into a classification budget?

Start with the shape of the run. Suppose, as an example, your team classifies 1 million records and each record produces 100 input tokens and 20 output tokens. That is 100 million input tokens and 20 million output tokens. At GPT-6 Astra rates on Quanta, the input portion would be $83.33 and the output portion would be $83.80, for an illustrative total of $167.13 before any cache effects.

These are examples, not quotes for your own dataset. The real figure depends on the prompt you assemble, the schema you ask for, whether answers are trimmed, and how much of the prompt is served from cache. Quanta prices those four categories separately: input, output, cache write and cache read. Models without a separate cache-write rate show a dash for that category.

It is worth comparing model families before committing the whole run to one route. The same worked-example basis of 3:1 input to output with no cache hits gives a useful reference point:

• › Claude Opus 5.5: 10 million tokens a month is approximately $34.58, and 100 million is approximately $345.83.
• › Claude Sonnet 5.5: 10 million tokens a month is approximately $25.07, and 100 million is approximately $250.69.
• › GPT-6.1 Sol: 10 million tokens a month is approximately $4.10, and 100 million is approximately $40.97.
• › Gemini 2.5 Flash: 10 million tokens a month is approximately $6.49, and 100 million is approximately $64.93.

These are examples on Quanta, not a recommendation for every workload. A classification pipeline may benefit from a fast model for routine categories and a stronger model for ambiguous ones. Quanta gives you a single catalogue and a single balance, so routing decisions can be made against visible per-token prices.

For teams that share budget, the referral programme adds another line to the account maths. A referred account earns you 19% of its top-ups, and 5% applies one level further. Rewards can be moved into the balance, or withdrawn once they reach $27.78.

03 · Can Quanta handle concurrency and sustained request volume?

Yes, and the limits are published rather than implied. Quanta enforces 120 requests per minute per key, with 8 requests in flight per account. A request body may be up to 25 MB, and an account can hold up to 20 live keys.

Those numbers are useful when you are designing a queue. If a classifier sends several requests in parallel, the in-flight ceiling of 8 is the constraint to plan around. If you run several workers against different keys, the 120 requests per minute per key limit applies to each key. Nothing about the limit requires a monthly commitment; Quanta is pay-as-you-go.

Long-running streams are also treated explicitly. There is no total-time cap on a stream, and a stream ends only after 300 seconds of upstream silence. For a bulk job, that means a slow response is not cut off simply because it crossed an arbitrary wall-clock threshold.

Quanta supports OpenAI-compatible access at /v1, Anthropic native and Gemini native, all under one key and one balance. GPT-6 Astra sits in the OpenAI-compatible route, alongside DeepSeek, Qwen, Hunyuan, Grok, GLM and Kimi models. Claude models work only from the official Claude Code and Claude Desktop clients, with Claude Code 2.1.285 or newer required. Gemini models use the native protocol only, through Gemini CLI. Codex, Cursor, Cherry Studio, Cline, OpenCode, OpenAI SDKs and cURL pair with GPT, DeepSeek and the other non-Claude, non-Gemini models.

04 · How do billing, invoices and refunds work for a large run?

Quanta bills from a prepaid balance, and the dashboard lists the tokens and charge of every call. That granularity matters when a classification run spans thousands of records: you can inspect the expensive calls, confirm the token counts, and see which stage of the pipeline consumed the balance.

Top-ups are available at $1.39, $6.94, $13.89, $27.78, $69.44 and $138.89. A small run can begin with a modest top-up, while a larger campaign can be funded in fewer steps. Unused balance can be refunded within 15 days of the top-up, which keeps the prepaid model from becoming a one-way commitment.

For a company, invoices and advance quotes are arranged through the Contact page or the Messages page of the dashboard, and payment is made by bank transfer from a company account. There is no monthly fee, no minimum spend and no subscription tier to unwind later. The service is metered on what the run actually uses.

One operational point is worth stating plainly. The text you send is retained for 30 days for abuse handling and billing disputes, and then deleted. Model answers are not stored. If your classification pipeline handles sensitive records, that retention window is part of the design conversation, not a footnote to discover later.

05 · What does a practical Quanta setup look like?

For a large classification run, the practical sequence is straightforward. Choose the model family that matches your client, fund the prepaid balance, and route the job through the relevant protocol. Keep an eye on the published limits: 120 requests per minute per key, 8 requests in flight per account, 25 MB per request body and 20 live keys per account.

Then work from the dashboard rather than from an estimate. Each call shows its tokens and its charge, split across input, output, cache writes and cache reads. When a run is finished, the totals tell you what the classification actually cost and which model choices to keep for the next pass.

Quanta combines a clear price list, prepaid pay-as-you-go billing, published concurrency limits and company invoicing through the dashboard. For teams running large classification jobs, that is a practical foundation: open an account, claim the $0.50 trial credit, top up what the pilot needs, and let the per-call records show you where the work goes.