GPT-6 Astra pricing

Price per 1M tokens

Batch factor: 50% of standard token pricing where the provider Batch API applies.

Worked monthly example

Consider a planning workload with 10 million input tokens and 2 million output tokens in one month. At the published token rates used here, the token component is about $200 before taxes, storage, tool calls, retrieval, networking, or other product-specific charges. This example is intentionally simple: it shows why both sides of the token equation matter. A model with inexpensive input can still become costly if output is verbose or if an agent repeatedly calls the model.

If the same workload is used by a team, multiply the request volume rather than simply multiplying a per-seat subscription assumption. API spend follows actual traffic. The calculator on the homepage lets you vary users, requests, input size, output size, caching, and agent multipliers instead of relying on this fixed example.

Caching scenario

Suppose 60% of those 10 million input tokens are eligible for the published cached-input rate while 40% remain regular input. With the same 2 million output tokens, the simplified token component becomes about $146. The difference is meaningful only when your application actually reuses cacheable prefixes and the provider applies the cached rate. Do not assume every repeated request qualifies.

Caching can improve economics for stable system prompts, repeated documents, or shared context. It is less useful when prompts change constantly or when the expensive part of the workload is output generation. Measure cache-hit behavior in production before turning a theoretical discount into a budget forecast.

Batch and asynchronous work

If the provider's published Batch API applies to this model and your workload can wait for asynchronous completion, the simplified example above would be about $100 at the recorded batch factor. Batch processing is usually a better fit for offline enrichment, document processing, evaluation, or back-office jobs than interactive chat. Latency requirements can therefore matter as much as token price.

What changes the result

For inputs above 272,000 tokens, this provider publishes long-context multipliers. That means a request that looks economical at normal context can cost materially more once the threshold is crossed. Model the long request separately instead of averaging it into short traffic.

Agent loops, retries, tool use, retrieval, image or audio processing, regional pricing, rate tiers, and minimum infrastructure commitments can all move the real bill. AI Cost Lab therefore treats the published token rate as an input to a scenario, not as a quote. The safest comparison is to estimate your own traffic pattern, then verify the final commercial terms with the provider before purchase.