GPT-6 Astra vs Claude Opus 5.5 vs Gemini 4 Argon: cost per task

OpenAI, Anthropic and Google each have a flagship model: GPT-6 Astra, Claude Opus 5.5 and Gemini 4 Argon. Their price per million tokens differs by up to five times today, but the price per million tokens is not what you pay. You pay per task, and a task's cost depends on how many tokens each model reads and writes to finish it, including thinking tokens you never see. This page puts the list prices side by side, works through four typical tasks and shows where the real differences come from. All prices are list prices in US dollars per million tokens on the standard pay-as-you-go tier, checked on each provider's own pricing page on 26 October 2026.

List prices side by side

All three bill thinking or reasoning tokens at the output rate, and all three offer a batch option at half the standard price. Google's 1 December increase is covered in detail in our Gemini price increase guide.

Cost per task: four worked examples

The token counts below are illustrations, not measurements. They assume each model uses the same number of tokens for the same task, which is rarely true; the next section explains why. Cache write charges are left out because, at a high hit rate, they are a small share of the total.

• Support answer: 6,000 input tokens, of which a 4,000-token system prompt and help text are served from cache, and 500 output tokens.
• Document extraction: a 20,000-token document in, 1,000 tokens of structured fields out, no caching. Shown on standard and on batch.
• Agent run: a multi-step task that reads 400,000 input tokens across all its steps, 90% of them from cache, and writes 20,000 output tokens. No single step is over 272K tokens.
• Long-document question: one 300,000-token prompt and a 2,000-token answer.

With equal token counts the ranking never changes: Argon is cheapest until 30 November, Opus 5.5 and Argon tie from 1 December, and Astra costs 2.5 times Opus 5.5 on ordinary prompts and about five times on the long-document question, where its surcharge applies.

Why token counts matter more than rates

Equal token counts are the exception. Three things make the same task use different numbers of tokens on different models.

1. Thinking tokens

All three models think before they answer, and you pay for that thinking at the output rate even when you see only a short reply or a summary. Anthropic's documentation says plainly that the billed output count does not match the visible text. Each provider lets you set how much the model thinks: Astra's reasoning effort runs from low to max, Opus 5.5's effort setting defaults to medium, and Argon's thinking level is low, medium or high with medium as the default. On a simple task, the setting can change the output tokens, and so the bill, more than the choice of model.

2. Different tokenizers

Each provider splits text into tokens its own way, so the same prompt is a different number of tokens on each model. Compare token counts from your own usage data rather than assuming one provider's count applies to another.

3. Caching rules

Caching is where most input savings come from, and the rules differ. Opus 5.5 caches prompts from 512 tokens, at points you mark in the request, with a one-off write charge. Argon caches automatically, but only prompts of at least 4,096 tokens, and does not support explicit caching. Astra lists a cache write price of 1.25 times its input rate. In every case, the start of the prompt has to be identical between requests for the cache to apply.

The break-even rule

Because Astra's rates are exactly 2.5 times Opus 5.5's for input, cached input and output, Astra is cheaper per task only if it completes the task in less than 40% of the tokens Opus 5.5 needs. The same reasoning applies between Opus 5.5 and Argon from 1 December: with identical rates, whichever uses fewer tokens on your task is cheaper. Until 30 November, Argon's rates are half of Opus 5.5's, so Opus 5.5 would need to use fewer than half of Argon's tokens to cost less.

Long prompts

If you send whole contracts, codebases or long chat histories, check the long-prompt rules first. OpenAI's model page says prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request, which makes Astra $20 input and $75 output per million tokens on those requests. Anthropic says Opus 5.5's full 1M-token window is priced at standard rates. Google lists no long-prompt tier for Argon; on Google's page, only Gemini 3.1 Pro (preview) has one, at 200K tokens. Keeping each Astra request under 272K tokens, for example by retrieving only the relevant parts of a document, avoids the surcharge entirely.

Ways to lower the cost on any of the three

• Batch work that can wait. All three charge half price for batch. On Argon after 1 December, batch costs what standard costs today.
• Lower the thinking setting on simple tasks. Test the low setting for classification, extraction and short factual answers on your own examples.
• Keep the start of the prompt fixed. Instructions, examples and reference text first; dates, names and the question last.
• Send simple requests to a smaller model. Each provider's mid-size model, such as GPT-6.1 Sol, Claude Sonnet 5.5 or Gemini 3.8 Flash, lists at a fraction of its flagship's price. Use the flagship for the requests that need it.
• Cap output. Set a maximum output length that fits the task, so a runaway answer or long chain of thought cannot run up the bill.

Measure it on your own workload

The only reliable cost-per-task figure is one from your own traffic. Start with the free API pricing calculator: enter your request volume and average input, output and cached tokens to see all three flagships, and every smaller model, side by side, with announced price changes. To see what you actually spend by model today, run your usage export through the free Bill Analyzer. And if you want a tested answer to which of your requests can move to a cheaper model without losing quality, the LLM Cost Report prices your real requests on every current model and gives you a ranked savings plan.