Google opened Gemini 4 Argon on the API on October 13, 2026, with a launch price of US$ 2 per million input tokens, half of Claude Opus 5.5. Google's own benchmark table puts Argon ahead in 13 of 19 tests. But in the two that look most like a coding agent working in a terminal, Opus 5.5 is still ahead.
This guide puts the numbers side by side, says where each one comes from and shows how to call both models so you can test them on your own repository. Everything checked on October 20, 2026.
Which is better for coding, Gemini 4 Argon or Claude Opus 5.5?
It depends on the task. For an agent that works in the terminal, runs commands and iterates until the test passes, Opus 5.5 leads comfortably: 66.4% vs 57.4% on Terminal-bench 4.0. For a task that requires reading a huge codebase at once, Argon leads: 84.2% vs 66.8% on GraphWalks from 256K to 1M tokens.
On the Artificial Analysis Intelligence Index, Opus 5.5 ranks 1st among 217 models (score 58) and Argon ranks 8th (score 53). Argon costs US$ 1.99 per index task; Opus 5.5, US$ 5.98.
What do the coding benchmarks say?
The table below gathers the coding tests from the launch table published by Google, with both models at maximum reasoning effort. These are numbers from Argon's maker, and the notes column shows who ran each test.
| Benchmark | Gemini 4 Argon | Claude Opus 5.5 | Who ran it |
|---|---|---|---|
| Terminal-bench 4.0 | 57.4% | 66.4% | Argon: Google. Opus: public leaderboard |
| FrontierSWE v2 | 55.0% | 62.3% | Proximal leaderboard |
| DeepSWE v1.1 | 77.9% | 74.2% | Argon: Google, mini-swe-agent harness. Opus: system card |
| Vibe Code Bench | 91.9% | 90.3% | Vals AI |
| PostTrainBench | 45.3% | 49.3% | Google, for all models |
| GraphWalks 256K to 1M (long context) | 84.2% | 66.8% | Google, for all models |
Source: Gemini 4 Argon launch table (Google DeepMind, September 30, 2026), reproduced by DataCamp and Apidog.
Two caveats before using these numbers. First, on DeepSWE Argon was measured by Google on its own harness, while Opus 5.5 came from Anthropic's system card. A different harness changes the result, so the 3.7-point lead is not a clean comparison. Second, on Terminal-bench 4.0 the Argon number is also self-reported. The two tests run by third parties with the same method (FrontierSWE and Vibe Code Bench) give one point to each side.
How much does each one cost?
Today Argon costs half. Starting December 1, it costs the same per token. Google announced that the launch price runs until November 30, 2026.
| Attribute | Gemini 4 Argon | Claude Opus 5.5 |
|---|---|---|
| API ID | gemini-4-argon | claude-opus-5-5 |
| Input (per 1M tokens) | US$ 2 until Nov 30, then US$ 4 | US$ 4 |
| Output (per 1M tokens) | US$ 10 until Nov 30, then US$ 20 | US$ 20 |
| Cached input | 95% off (implicit caching) | reads at 5% of the input price |
| Context window | 1,048,576 tokens, input and output combined | 1M tokens |
| Max output per call | 65,536 tokens (default) | 128K tokens |
| Reasoning level | low, medium, high (default medium) | effort, default medium |
| Output speed (Artificial Analysis) | 83.9 tokens/s | 95.1 tokens/s |
| Intelligence Index (Artificial Analysis) | 53, 8th of 217 | 58, 1st of 217 |
| Cost per index task | US$ 1.99 | US$ 5.98 |
Sources: Google blog (October 13, 2026), Gemini API docs, Anthropic models docs and both model pages on Artificial Analysis, all checked on October 20, 2026.
Our own math, from these numbers: at launch price, each index point costs about US$ 0.04 on Argon and US$ 0.10 on Opus 5.5. If Argon spends the same tokens per task after the price doubles in December, a point goes to about US$ 0.075, and the advantage drops from 2.7x to 1.4x. It stays cheaper per point because it spends fewer tokens than Opus 5.5 to reach the answer.
When should you choose Gemini 4 Argon?
- When the task is understanding a large monorepo at once. The 84.2% on GraphWalks 256K to 1M is the widest gap between the two in the whole table.
- When volume is high and the launch price window helps. Until November 30 you pay half per token.
- When the work mixes code with image, video, audio or PDF input. Argon accepts all of these formats.
- When you are already in the Google ecosystem: AI Studio, Android Studio or Antigravity (on the Google AI Ultra plan).
When should you choose Claude Opus 5.5?
- When the coding agent lives in the terminal: runs the test, reads the error, fixes, runs again. That is a 9-point lead on Terminal-bench 4.0.
- When the task is long, open-ended software engineering, as FrontierSWE v2 measures (62.3% vs 55.0%).
- When the answer needs to be large. Max output is 128K tokens per call, against 65,536 for Argon in default mode.
- When you already use Claude Code and do not want to switch tools.
How do you test both on your code?
Benchmarks show the trend; your repository tells you the rest. Run the same request on both and compare.
Gemini 4 Argon through the Gemini API, with reasoning at the highest level (requires a paid tier key):
curl -X POST "https://generativelanguage.googleapis.com/v1beta/interactions" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-H 'Content-Type: application/json' \
-d '{
"model": "gemini-4-argon",
"input": "Find the bug in this function and propose the failing test: ...",
"generation_config": {
"thinking_level": "high"
}
}'Claude Opus 5.5 through the Anthropic API:
curl https://api.anthropic.com/v1/messages \
-H "x-api-key: $ANTHROPIC_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "content-type: application/json" \
-d '{
"model": "claude-opus-5-5",
"max_tokens": 8192,
"messages": [
{"role": "user", "content": "Find the bug in this function and propose the failing test: ..."}
]
}'To test Opus 5.5 as a terminal agent, inside your own repository:
claude --model claude-opus-5-5One practical caveat: as of October 20, 2026, Argon does not yet appear in the model list of the Gemini CLI source code (packages/core/src/config/models.ts). To use Argon as an agent today, the official path is Antigravity, or the API directly in your own harness.
Summary: which one to choose?
| If your case is... | Choose |
|---|---|
| Terminal agent, test and fix loop | Claude Opus 5.5 |
| Reading a huge repository at once | Gemini 4 Argon |
| High volume until Nov 30, 2026 | Gemini 4 Argon |
| Very long answer in a single call | Claude Opus 5.5 |
| Video, audio or PDF input | Gemini 4 Argon |
Both charge per token, and in a coding agent the bill grows with every attempt. If you do not want that headache anymore, Verboo Code offers unlimited tokens.
