On this page
Gemini 4 Argon costs half of Claude Opus 5.5 for 34 more days. As of October 27, 2026, Google's pricing page lists Argon at $2 in and $10 out through November 30, then Opus's $4 and $20 from December 1. Then pick on the work. On Google's card, Opus leads on terminal jobs and Argon wins two of the other three coding tests.
If your agents work in a terminal, pick Opus.
If the job is very long documents or finance and legal agent work, pick Argon.
If you picked Argon for its price, test again before December.
Key facts
| Gemini 4 Argon | Claude Opus 5.5 | |
|---|---|---|
| Maker | Anthropic | |
| Generally available | 13 Oct 2026 | 22 Sep 2026 |
| Model id | gemini-4-argon | claude-opus-5-5 |
| Price per million tokens (in / out) | $2 / $10 to 30 Nov, then $4 / $20 | $4 / $20 |
| How much it can read at once | 1M tokens, shared by input, thinking and output | 1M tokens |
| Longest single reply | 65,536 tokens (longer with long decode) | 128K tokens |
| Free tier | No | No |
| Download and run it yourself | No | No |
A token is about three quarters of a word.
Why is there a deadline?
Argon itself has not changed since launch; the clock on its price is what matters now. Google's changelog marks Argon generally available on October 13, 2026, on the paid API only, with no free tier.
As of October 27, every Argon rate on the pricing page carries the same end date: standard, batch and cached reads all run at the lower price through November 30 and double on December 1. Counting from October 28, teams that chose Argon on cost have 34 days to test it against Opus on their own work.
Which is better for coding, Gemini 4 Argon or Claude Opus 5.5?
Opus is ahead on terminal work, Argon is ahead on one long-task test, and they split the rest. Every number below comes from the model card Google published on October 13, so both models were scored on the same version of each test. These are Google's figures, not an independent run.
| Coding task (Google's figures) | Gemini 4 Argon | Claude Opus 5.5 |
|---|---|---|
| Finishing multi-step jobs in a command-line terminal | 57.4% | 66.4% |
| Long, multi-file software engineering tasks (DeepSWE) | 77.9% | 74.2% |
| Long, multi-file software engineering tasks (FrontierSWE) | 55.0% | 62.3% |
| Building small apps from a description | 91.9% | 90.3% |
The terminal gap is the one that matters most for agents. Opus finished about two in three terminal jobs; Argon finished closer to four in seven. Anthropic's own launch page reports the same 66.4% on that test, so the two vendors agree on that figure.
The two long-task tests disagree with each other. Argon wins one by under 4 points and loses the other by 7. Read that as a draw that depends on the kind of task, and test on your own repository. A score from one version of a test never compares with another version; we wrote about why.
Outside coding, the same card puts Argon ahead on office-style work. On a finance agent test it scores 65.4% to Opus's 58.6%. It also leads on automating business workflows end to end.
What does each cost before and after December 1?
Until November 30, Argon costs half of Opus on every line. From December 1, the two list prices match on every line except caching.
| Per million tokens | Argon to 30 Nov | Argon from 1 Dec | Opus 5.5 |
|---|---|---|---|
| Input | $2.00 | $4.00 | $4.00 |
| Output (thinking counts as output) | $10.00 | $20.00 | $20.00 |
| Reading a cached prompt | $0.10 | $0.20 | $0.20 |
| Writing to the cache | not listed | not listed | $5 (5 min) or $8 (1 hour) |
| Batch input / output | $1 / $5 | $2 / $10 | $2 / $10 |
On Google's pricing page a cached read costs 5% of a normal input token, before and after the change. Anthropic's model page lists the batch discount as 50% and the cache read at $0.20.
A worked month: an agent reads 100 million tokens and writes 10 million, and 80 million of the reads hit the cache.
| Worked month | Uncached input (20M) | Cached input (80M) | Output (10M) | Total |
|---|---|---|---|---|
| Argon in November | $40 | $8 | $100 | $148 |
| Argon in December | $80 | $16 | $200 | $296 |
| Opus | $80 | $16 | $200 | $296 plus cache writes |
Anthropic's cache-write charge depends on how often your prompt prefix changes, so we left it out of the total.
The list price stops deciding this on December 1. What decides it is how many tries each model needs to finish a task, which is the argument in cost per solved task.
How do context and caching differ?
The two models both read up to a million tokens, but the Google model counts its own thinking and reply inside that million. Google's model page gives one 1,048,576-token window shared by input and output, and a normal reply stops at 65,536 tokens unless you use long decode. Opus replies up to 128K tokens, or 300K through the batch API with a beta header.
On long documents Argon reads better by Google's numbers. On a test of following links through very long text, Argon scores 84.2% and Opus 66.8%.
Caching works differently:
- Argon caches automatically. You cannot create a cache yourself; the pricing page says explicit caching "is not supported".
- Opus caches what you mark, from 512 tokens, for five minutes or an hour, and charges to write it.
Automatic caching needs no code. Explicit caching lets you decide which prefix stays warm, which matters when a long system prompt must stay cached across a slow human review step.
When should you pick each?
Pick Opus for agents that run commands, edit files and recover from errors in a terminal. That is its clearest lead, and both vendors' numbers agree on it.
Pick Argon for very long inputs, finance and legal agent work, and anything where the cost under November's price lets you run a bigger test before December 1. Argon supports the batch API, which halves the price again: $1 in and $5 out until November 30.
What we did not test. We have not run Argon or Opus 5.5 on our own coding tasks. Every score in this post is a vendor figure.
Can you use Gemini 4 Argon or Claude Opus 5.5 with data kept in Canada?
No. Canada has no documented route for either model, so your prompts are processed on the vendor's infrastructure outside the country.
Google's data residency page lists Argon for the US, the EU, the UK and Singapore, with no Canadian region. Global endpoints "don't provide any data residency guarantees". Google's API pages for the model say nothing about region at all.
Anthropic's data residency controls offer two inference settings, global or US only, and US only costs 1.1 times the standard rate. Workspace storage is in the US.
Bedrock launched a Canada profile on October 27, 2026 that keeps requests inside the country. Claude Haiku 5.5, a small model, is the only one it carries. Google's residency page has no Canadian entry for any Claude model.
For inference in Canada at this level, the route is an open-weight model on dedicated GPUs in a Canadian region. Kimi K3 is the strongest one you can download today; its weights have been on Hugging Face since July 2026. Its model card lists 2.8 trillion parameters, of which 104 billion work on each token. The download is about 1.56 terabytes.
Moonshot, the model's maker, reports K3's terminal score on Terminal-Bench 2.1, an older version of the test than the 4.0 figures above, so the two cannot be set side by side. Our comparison of K3 with Claude Opus 4.8 goes through what Moonshot did publish.
The licence is permissive with conditions. A company that sells model access as a service needs a separate agreement with Moonshot once its revenue passes $20 million over 12 months. Very large products, above 100 million monthly users or $20 million in monthly revenue, must show the model's name in the interface. Internal use is exempt from both.
The weights can run in AWS ca-central-1 (Montreal) or ca-west-1 (Calgary), or on hardware you own. Our Canadian residency guide lists three trade-offs. GPUs are scarcer in these regions, you pay for them while idle, and you own the serving.
XY Space deploys open-weight models on dedicated GPUs in the client's own Canadian cloud account or on-premises. Prompts and outputs stay in the country, and we build the workflow on top. How we deploy covers the build.
If you need this model's capability with inference kept in Canada, bring one job to a free 30-minute call.
FAQ
Is Gemini 4 Argon cheaper than Claude Opus 5.5?
Until November 30, 2026, yes: Argon costs half as much per token. From December 1, the pricing page lists it at $4 per million input tokens and $20 per million output, the same as Opus. Batch prices also match from that date. Cached reads match at $0.20, though Opus also charges to write the cache.
Is Gemini 4 Argon better than Claude Opus 5.5 for coding?
Argon is better on some coding tasks and worse on others. On the vendor's own model card, Opus leads on finishing multi-step terminal jobs, 66.4% to 57.4%. It also leads on FrontierSWE, 62.3% to 55.0%. Argon leads on DeepSWE, 77.9% to 74.2%, and on building small apps from a description. These are vendor numbers, so run both on a sample of your own tickets before choosing.
Can I use Gemini 4 Argon for free?
No. Google's pricing page states that Argon is not available on the free tier. You need a paid tier, and the price per million tokens doubles on December 1, 2026. Test it on a small paid budget before committing a workload, and do the test before the price changes.
Can I run Gemini 4 Argon or Claude Opus 5.5 in Canada?
Not with documented in-country inference. Neither maker lists a Canadian processing option for these two models, and the one Canadian route on Bedrock carries only a smaller Claude model. For in-country inference at this level, run an open-weight model such as Kimi K3 on dedicated GPUs in a Canadian region.
Sources
- Google, Gemini API pricing, as of 27 Oct 2026 (Wayback snapshot): aistudio.google.com/docs/pricing
- Google, Gemini API changelog (Argon GA 13 Oct 2026): aistudio.google.com/docs/changelog
- Google, Gemini 4 Argon model page, as of 19 Oct 2026 (Wayback snapshot): aistudio.google.com/docs/models/gemini-4-argon
- Google DeepMind, Gemini 4 Argon model card, 13 Oct 2026 (Terminal-Bench 4.0, DeepSWE v1.1, FrontierSWE v2, Vibe Code Bench, Vals Finance Agent v2, AutomationBench, GraphWalks): deepmind.google
- Anthropic, Claude Opus 5.5 model overview, released 22 Sep 2026: platform.claude.com
- Anthropic, Claude Opus 5.5 announcement, 22 Sep 2026 (Terminal-Bench 4.0): anthropic.com
- Anthropic, Data residency: platform.claude.com
- Google Cloud, Gemini Enterprise Agent Platform data residency: docs.cloud.google.com
- AWS, Claude Haiku 5.5 Canada cross-region inference, 27 Oct 2026: aws.amazon.com
- Moonshot AI, Kimi K3 model card and licence (weights July 2026; Terminal-Bench 2.1): huggingface.co/moonshotai/Kimi-K3
Written by
Cho Yin Yong
Principal AI Solutions Engineer, XY Space
Principal AI Solutions Engineer at XY Space. University of Toronto lecturer for five years, co-author of two patents, winner of two competitive AI awards, and nine years of regulated engineering leadership.
More from Cho Yin YongShare this article
Work with us
We build the systems these posts describe, and we'll tell you in the first call whether yours is worth building.
Start a project