Gemini 4 Argon vs GPT-6 Astra: Real Cost per Task Alternatives

Gemini 4 Argon vs GPT-6 Astra looks like a price mismatch, with Astra charging five times more per token. Independent tests show Astra using about a fifth of the output tokens per task, which already makes it cheaper per task and widens the gap after December 1, while long-document work still favors Argon.

Share this post X LinkedIn Email Facebook Copy linkTL;DR

OpenAI charges $10 per million input tokens and $50 per million output tokens for GPT-6 Astra through its application programming interface (API), according to its developer documentation, which is five times the $2 and $10 Google charges for Gemini 4 Argon during an introductory period that ends on November 30. On a rate card that looks decisive, yet when Artificial Analysis ran both models at high effort through its Intelligence Index, Astra cost $1.73 per task against $1.99 for Argon, because Argon produced about five times as many output tokens per task.

This comparison is for developers and companies choosing between OpenAI and Google for paid API work, where the bill depends on how many tokens a model burns rather than on what each token costs. Memeburn’s reading of the published data is that token efficiency already cancels most of Astra’s price premium on agent-style work and will reverse it once Argon’s rates double on December 1, while Argon keeps a structural price advantage on long documents, because input volume is fixed by the file rather than by how verbose the model is.

Gemini 4 Argon’s input list, taken from Google’s model page, is the clearest functional gap, since OpenAI’s model does not accept audio or video at all. The surcharge row matters as much as the headline prices, because any Astra request whose prompt crosses 272,000 tokens is billed at the higher rates in full, not just for the tokens above the line.

GPT-6 Astra at high effort averaged about 12,000 output tokens per task in Artificial Analysis’s suite, including about 5,000 reasoning tokens, while Gemini 4 Argon at high averaged about 62,000 output tokens and 36,000 reasoning tokens. Over the full index Argon generated 113 million output tokens against 26 million for Astra, so Memeburn calculated that Argon used about 4.3 times as many output tokens, which nearly offsets a fivefold difference in output price.

The result depends on which cost measure is used, which is why neither vendor’s framing tells the whole story. Astra at high was cheaper per average task, at $1.73 against $1.99, but cost more to run through the whole index, at $2,925 against $2,407, a 22% premium rather than a 400% one, and it scored 51 to Argon’s 53. When OpenAI’s model is pushed to its max setting, it matches Argon at 53 but uses 60 million output tokens and costs $5,324 for the full run, according to Artificial Analysis’s max-effort comparison.

Gemini 4 Argon’s rates double on December 1, according to Google’s API documentation, while Astra’s stay at $10 and $50, which narrows the per-token gap from five times to two and a half. Memeburn recalculated Artificial Analysis’s figures at the new prices, assuming each model uses the same tokens as in the published runs.

At those prices, Astra at high costs about 39% less than Argon for the full index and about 57% less per task, for a two-point lower score. At equal scores the comparison becomes close, since Astra at max costs about 11% more than Argon over the full index yet about 18% less per average task, so a team should measure its own task mix rather than assume either model is the budget option.

Gemini 4 Argon’s advantage survives where input dominates the bill, because a model can be thrifty with its reasoning but cannot shrink the contract it has been asked to read. Memeburn priced a hypothetical request with 500,000 input tokens and 20,000 output tokens at list rates without caching, which comes to about $11.50 on Astra once the 272K surcharge applies, against $1.20 on Argon today and $2.40 from December 1. Even if Argon used four times as many output tokens, the December bill would be about $3.60, or less than a third of Astra’s.

Accuracy does not obviously tip that calculation back toward OpenAI. Google’s model card puts Argon ahead at 84.2% against 71.8% on GraphWalks between 256,000 and 1 million tokens, OpenAI reports 96.3% for Astra on its own MRCR test at 512,000 to 1 million tokens without a Gemini score, and Artificial Analysis’s independent AA-LCR long-document test has the two within a point, at 80% for Argon against 81% for Astra at max. With accuracy roughly level on the one shared independent test, legal, financial and research teams feeding in whole document sets have little reason to pay the surcharge.

GPT-6 Astra’s strongest results are in software engineering and computer use, where Google’s own table shows it ahead on FrontierSWE v2 at 65.5% against 55.0%, on Terminal-Bench Science at 68.1% against 57.6% and on OSWorld 2.0 at 72.6% against 69.2%. Argon wins 14 of the 19 rows in that table, including DeepSWE v1.1 at 77.9% against 74.1%, while Terminal-Bench 4.0 is close to a tie at 58.2% for Astra against 57.4%. Artificial Analysis’s independent run gives Astra at max the edge on its AA-Briefcase workplace test, on GDP.pdf and on Terminal-Bench 4.0, while Argon leads AutomationBench-AA at 78% against 68% and GDPval-AA at 1,624 Elo against 1,573.

The vendor tables should not be mixed, because the same benchmark can return very different numbers for the same model. OpenAI’s launch post reports 59.3% for Astra on Agents’ Last Exam, while Google’s table lists Astra at 34.2% on what it labels the pass rate for the same benchmark, a gap large enough to reverse any ranking built on it. OpenAI’s post also claims Astra finishes OSWorld 2.0 tasks in about 40 minutes, roughly 47% faster than GPT-5.6 Sol, which fits the token data above, although Artificial Analysis lists no speed figures for Argon to compare against.

Based on published benchmarks and Memeburn’s arithmetic rather than a Memeburn test, the split looks like this:

Both models joined a crowded frontier tier within weeks of each other, which Memeburn tracked in six weeks of major AI launches, and the security side of Astra’s launch is covered in Memeburn’s look at what GPT-6 Astra means for crypto security.

Gemini 4 Argon is cheaper per token, at $2 input and $10 output per million until November 30 and $4 and $20 afterwards, against GPT-6 Astra’s $10 and $50. Per task the gap is smaller, because Artificial Analysis found Astra at high effort cost $1.73 per task against $1.99 for Argon.

GPT-6 Astra leads Gemini 4 Argon on FrontierSWE v2, Terminal-Bench Science and OSWorld 2.0 in Google’s table, but Argon wins 14 of 19 rows there. Artificial Analysis scores both at 53 when Astra runs at max effort and Argon at high.

Yes. GPT-6 Astra prompts above 272,000 input tokens are billed at twice the input rate and 1.5 times the output rate for the whole request, according to OpenAI’s API documentation. Google lists no comparable surcharge for Gemini 4 Argon.

Marko is a tech journalist covering AI, consumer technology, crypto, and digital innovation. His work focuses on clear, accessible reporting that helps readers understand how new technologies are shaping business, finance, and everyday life.