Gemini 4 Argon vs GPT-6 Astra vs Claude Opus 5.5 Benchmarks

Update (9 Nov 2026): Claude Fable 5.2 and Mythos 5.2 are out - prices, access and benchmarks are in my Claude Fable 5.2 guide.

👁 127 reads

Gemini 4 Argon beats GPT-6 Astra and Claude Opus 5.5 in 13 of 19 tests published by Google, and at launch pricing costs one-fifth of Astra’s rates. However, it loses in two of four programming tests and is the only one of the three not available today. We’ve compiled data published by the manufacturers and calculated the cost of three typical tasks.

Three Models Compared

Astra and Opus prices are standard rates from manufacturer price lists, excluding tax. Argon prices come from Google’s announcement. The model is not yet in the official Gemini API price list. We cover each model separately: Gemini 4 Argon, GPT-6 Astra, and Claude Opus 5.5.

Benchmarks: Who Wins What

Below is the complete table from Google DeepMind’s methodology document. Google calculated Argon’s results itself at the highest reasoning setting; Astra and Opus results come from public leaderboards or manufacturer model cards. We’ve bolded the best result in each row.

Office Work

Argon has its clearest advantage here. In the Harvey legal test, it scores five times higher than Opus 5.5, though all three models are far from the halfway mark.

Programming and ML Engineering

Effectively a tie. Argon wins in DeepSWE and Vibe Code Bench, but loses to Astra by 10.5 points in FrontierSWE v2 and to Opus by 9 points in Terminal-Bench 4.0. Additionally, Bloomberg, cited by Techmeme, reported that according to some Google employees, the model performs worse on real programming tasks than in benchmarks. Google disputes this. For terminal work and programming agents, Opus 5.5 currently has the strongest numbers in this trio.

Science and Mathematics

Long Context, Computer Use, Multimodal

For OSWorld-2.0, Google notes that Anthropic only reports a combined score for two parts of the test, so Opus is not included in that row. Google calculated GraphWalks results itself for all models.

Cybersecurity

We’re citing the Gray Swan result (resistance to hidden instructions in content the model reads) from VentureBeat, which quoted data from Google’s materials. In CWE-bench, Google declared a tie with Astra in first place.

Bottom Line

Out of 19 rows: Argon leads in 13, ties in 1, Astra wins 3, Opus 5.5 wins 2. Two caveats: Google assembled this table, so it chose the tests, and popular benchmarks like SWE-bench, GPQA, Humanity’s Last Exam, and ARC-AGI are missing. Also, Argon’s results haven’t been independently verified yet because the model isn’t publicly available.

Cost of Typical Tasks

We calculated the cost using list prices only, excluding tax, cache write fees, tools, and batch modes. This is a calculation based on rates, not a measurement. The actual bill depends on how many reasoning tokens the model consumes, and no one has measured that for Argon yet.

Task 1: Long Document Analysis

100k input tokens (contract or report), 5k output tokens.

Task 2: Programming Agent Session

Total 2M input tokens, of which 1.6M read from cache and 0.4M without cache, plus 100k output tokens. We assume no single query exceeds 272k tokens.

Task 3: Very Long Output

50k input tokens and 400k output tokens, e.g., a large code migration. The limit matters here: only Argon can do this in a single response. Astra and Opus 5.5 each have 128k output token limits, so they would need at least four calls, with each subsequent one reloading the context.

Conclusion from the numbers: at launch pricing, Argon is the cheapest of the three flagship models for every task. After the launch period, it matches Opus 5.5, while Astra remains about 2.5 times more expensive than both. Google hasn’t specified how long the launch pricing will last.

Which Model to Choose Today

• Office work, finance, legal, long documents - on Google’s numbers, Argon, but you can’t get it today. Of those available, Opus 5.5 has higher scores than Astra in Vals Index and Finance Agent at a lower price.
• Terminal agents and programming - Opus 5.5 (Terminal-Bench 4.0, PostTrainBench), and for difficult code from FrontierSWE, also Astra.
• Science with terminal use and computer control - Astra (Terminal-Bench Science, OSWorld-2.0).
• Very long outputs in a single run - only Argon has a limit above 128k tokens.

These are conclusions from manufacturer tables. We run our own comparisons in the Lab: Astra and Opus 5.5 are already tested; we’ll add Argon as soon as it reaches the API. You can find all models in one place on the AI models page, and we cover the rest of Google’s family in our Gemini guide.

Frequently Asked Questions

Is Gemini 4 Argon better than GPT-6 Astra?

In Google’s table yes, in 13 of 19 tests, but Astra wins in FrontierSWE v2, Terminal-Bench Science, and OSWorld-2.0. Independent measurements aren’t available yet.

Is Gemini 4 Argon better than Claude Opus 5.5?

In office work, long context, and video yes. In Terminal-Bench 4.0 and PostTrainBench, Opus 5.5 wins.

Which of the three models is cheapest?

At launch pricing, Argon: $2 / $10 per million tokens. After the launch period, Argon and Opus 5.5 cost the same, $4 / $20. Astra is most expensive at $10 / $50.

Why isn’t GPT-6.1 Sol in the comparison?

Google didn’t compare Argon to it in their table, so we don’t have shared, published results. Sol appears only in the cost calculation because it has identical pricing to launch Argon.

Kinetyra