Gemini 4 Argon vs Claude Opus 5.5 vs GPT-6 Astra: price, limits and which to pick

Gemini 4 Argon, Claude Opus 5.5 and GPT-6 Astra are the 3 makers' top general models in late 2026, and on Epoch AI's capability index they score within each other's range of doubt, so none is measurably ahead. The choice comes down to price and limits: Gemini 4 Argon is half price until 30 November 2026 and then costs exactly what Claude Opus 5.5 costs, $4 per million input tokens and $20 per million output, while GPT-6 Astra costs 2.5 times that.

Google made Gemini 4 Argon generally available on 13 October 2026, for "complex coding, agentic, math, and science tasks". Anthropic tells most teams to start with Opus 5.5, and OpenAI calls Astra its model "for the most demanding work". Below: the specs side by side, the month of the same job on each, and the changes that break a swap.

On 28 October Google announced Gemini 4 Carbon, which it calls its strongest model and which builds on Argon. Only a subset of cyber defenders get it at first; developers get it later, on no date yet, and it is not in the Gemini API release notes. Until it is, Argon is the Gemini 4 model you can call. What Carbon is and how it compares with Argon.

Side by side

The 3 have about the same context window; they differ in output length, caching and what they accept as input.

Scores are the Epoch Capabilities Index by Epoch AI (CC BY 4.0), which combines over 50 benchmarks into one number. A gap of 2 points with ranges 7 points wide says the 3 are level, not that Opus 5.5 wins.

The price that changes on 1 December

Gemini 4 Argon's launch price lasts until 30 November 2026; from 1 December every Argon rate doubles, and Standard input, output and cache-read prices match Claude Opus 5.5 to the cent. Batch, Flex and Priority double too: Batch goes from $1 and $5 to $2 and $10, the same as Opus 5.5 on its Batch API.

So a cost comparison made in October with Argon at $2 is out of date in 5 weeks. A job that looks half price on Argon now costs the same as on Opus from December, unless caching or the long-prompt rules tip it.

Take the same job on all 3 and switch the price month:

Try itThe same job on Gemini 4 Argon, Claude Opus 5.5 and GPT-6 Astra

Job

·Coding agent turn·Question on a long document·Chat answer

Requests a month

20,000

Input tokens, per request

120K

Output tokens, thinking included

6K

NoneHalf90%

To 30 NovFrom 1 Dec

Right awayAs a batch

The month, cheapest first

• Gemini 4 Argon$1,896 ($0.0948 each)

• Claude Opus 5.5$3,792 ($0.1896 each)

• GPT-6 Astra$10,560 ($0.528 each)

Capability score, with its range of doubt

• Gemini 4 Argon
• Claude Opus 5.5
• GPT-6 Astra

158166174

3 jobs at the start prices, per request:

3 things the table shows:

• Long documents are where Astra hurts. A 400K-token prompt is over its 272K step, so the whole request is billed at $20 in and $75 out: 5 times Opus 5.5. The context window is about the same on all 3; the price of filling it is not.
• Short prompts miss Argon's cache. Implicit caching only starts at 4,096 tokens, so the 3K chat prompt is billed in full on Argon, and from December Opus 5.5 is cheaper for it. Claude caches a marked prefix from 512 tokens; prompt caching covers how prefixes work, and the prompt caching calculator prices it on every Claude model.
• Agent turns tie from December. A turn that re-reads a long, cached history costs the same on Argon and Opus once Argon's launch price ends, and close to 3 times as much on Astra.

These are list prices for the same token counts. Each maker's tokenizer turns the same text into a different number of tokens, and each model spends its own amount on thinking, which all 3 bill as output. Count your real prompts in the token counter and treat a gap under about 20% as a tie until you have measured it.

Which to pick for which job

Pick by what the job needs from the model's limits, then test the 2 cheapest that fit on your own prompts.

• Agents and coding: Opus 5.5 or Argon. Both are priced the same from December, and each has its maker's agent product: Claude Code and the Claude Managed Agents runtime, and Google's Antigravity agent, which runs Argon in Gemini Managed Agents. Claude's explicit cache control makes the bill of a long agent session easier to predict.
• Very long answers: Argon. Long decode writes up to 1 million tokens in one response, run on Google's servers in hops of 64K with background execution. The other 2 stop at 128K in a normal request.
• Video, audio and PDF in the prompt: Argon takes all 3 directly. Opus 5.5 and Astra take text and images.
• Data that must stay on your cloud: Opus 5.5 is sold through Amazon Bedrock, Google Cloud and Microsoft Foundry under the same model id, so it can run inside the account you already have. US-only processing on the Claude API adds 10%.
• Astra: only where your own evals show it ahead on your work. It costs 2.5 times Opus 5.5, and OpenAI's own GPT-6.1 Sol scores 166.3 against Astra's 166.4 at a fifth of the price; GPT-6 Astra vs Sol vs Luna covers that choice.

If a small model would do, none of the 3 is the answer: model routing sends the easy requests to a cheaper one, and which model finds the cheapest model that scores high enough for a job.

What breaks when you switch

Each model has its own request settings, so moving a working prompt to another of the 3 needs more than a new model id.

Moving to Gemini 4 Argon, from Google's migration checklist:

• Remove explicit caching. cachedContents and caches.create are not supported; implicit caching takes over for prompts of 4,096 tokens and up.
• Remove temperature, top_p and top_k. Argon ignores them, so output that depended on a low temperature may vary more.
• Raise max_output_tokens to at least 32,768. Argon can spend more of the limit on thinking than earlier Gemini models, and the response stops when thinking and output together reach it.
• Set client timeouts to at least 15 minutes, or use background execution, because hard prompts can take several minutes.

Moving to Claude Opus 5.5: thinking is adaptive and always on, steered by effort (default medium). The manual budget_tokens setting of earlier Claude models is not accepted. Cache what repeats with cache_control; a write costs 1.25 times input for 5 minutes, 2 times for an hour.

Moving to GPT-6 Astra: keep prompts under 272K input tokens if you can, because one token over bills the whole request at the higher rate, and set reasoning.effort from low to max instead of a token budget.

Run the same set of real tasks on each before you switch; how to evaluate an AI agent covers building one, and an eval set of a few hundred cases shows a difference the index cannot.

Every price, kept current

The model pages carry every rate, the cached and batch prices and what 8 common jobs cost on each: Gemini 4 Argon, Claude Opus 5.5 and GPT-6 Astra. The 2-model pages set them against each other with your own token counts: Opus 5.5 vs Argon, Opus 5.5 vs Astra and Argon vs Astra. Prices there are read from each maker's own page every week, so Argon's December rise shows up when it happens.