GPT-6 Astra is sold on one capability: it can drive a computer for hours without losing the thread. The headline price is $10 per million input tokens and $50 per million output. Both numbers are close to useless on their own.
The first is useless because the comparison everyone made was the wrong one. The second is useless because the effective rate moves by a factor of a hundred depending on how you call it.
Strip the launch language and Astra is one capability sold hard: it can drive a computer for hours without losing the thread. Everything else in the announcement is downstream of that.
The pricing is $10 per million input tokens and $50 per million output. Two comparisons matter and most coverage only made the first:
Against GPT-5.6 Sol, it is 2.5x. That is the number your CFO will ask about, and the capability gain is real but not obviously 2.5x real for general work.
Against Claude Fable, it is identical. That is the number that should drive your decisions. At the same price you are choosing between a model that is ahead on operating software and one that is ahead on reasoning quality. Price has stopped being the tiebreaker, which means task shape is the only tiebreaker left.
The spec that matters for long-horizon work, because it is what lets a run survive a full working day:
Note what is missing, because it constrains architecture: no fine-tuning, no embeddings, no realtime, no audio. If your plan involved tuning a model on your own corpus, that plan is not available here. The intended path is context and tools, not weights.
The price is not $10/$50
The headline rate is the least useful number in the pricing table. The effective rate moves by a factor of a hundred depending on how you call it, and long agentic runs sit at both extremes at once.
Four consequences worth acting on:
Cached input is $1 per million against $10 standard. Cache writes bill at 1.25x uncached, so the write pays for itself on the second read. A long-running agent that re-reads the same codebase, policy set or corpus every turn should be hitting cache on nearly all of it. If your bill looks like standard input all the way down, you are paying ten times more than the workload requires.
The 272K cliff is real. Cross it and input bills at 2x and output at 1.5x. A million-token context is available, not free, and drifting over the line mid-run is a silent doubling. Compaction is the countermeasure, and it is the same mechanism that bought the 37 points in section 3.
Batch and Flex are half price. Overnight QA sweeps, corpus assembly and nightly regression runs have no latency requirement, which makes them exactly the workloads that should never run at standard rates. Fast mode is 2x and is for the handful of paths where a human is waiting.
Tool calls bill per call on top of tokens. Search and computer use are metered separately, and computer use is the thing you bought the model for, so this line grows with adoption rather than shrinking.
Rate limits scale hard by tier, from 500 requests and 500K tokens per minute at Tier 1 to 15,000 requests and 40M tokens per minute at Tier 5. A fleet plan that works in a pilot can hit a wall on the day you scale it, and moving tiers is a spend history, not a support ticket.
Put the four levers together and the same workload can differ by an order of magnitude on the invoice with no change at all to what the agent does. Caching is the biggest lever and the least discussed.
What would your agent workload actually cost?
Book a free Diagnostic: 30 to 45 minutes, no deck, no pitch. We price one real workflow against cached, batch and standard rates before you commit to a model.
Book the Diagnostic →Part 1 of Working with GPT-6 Astra: The Operator Guide.
