Short answer: Gemini 3 Pro Preview showed promising but mixed coding benchmark results. It is no longer a choice for a new Google Gemini API integration: Google says gemini-3-pro-preview was shut down on March 9, 2026. Google’s notice does not establish whether a third-party provider still offers access, so check that provider’s current listing and terms separately. (Google’s Gemini 3 Pro Preview model page)
First, distinguish the model from the coding tools
“Gemini 3 Pro” here means the original Gemini 3 Pro Preview model, identified by Google as gemini-3-pro-preview. A third-party Kilo Code listing used the provider-style identifier google/gemini-3-pro-preview; that listing is not confirmation of current availability through Google’s API. (Kilo Code listing)
Do not confuse the retired model with Gemini 3.1 Pro Preview, a separate model, or Google Antigravity, a coding interface. TechCrunch described Antigravity as an agentic workflow spanning an editor, terminal, and browser. That description is not an API specification or a coding-quality evaluation. (TechCrunch’s Antigravity coverage)
What the reported coding benchmarks showed
A Roboflow article published in December 2025 reported these Gemini 3 Pro results: (Roboflow’s benchmark article)
| Benchmark | Reported Gemini 3 Pro result | Evaluation note |
|---|---|---|
| LiveCodeBench Pro | 2,439 Elo | Higher is better, according to the article’s table |
| Terminal-Bench 2.0 | 54.2% | Tested with the Terminus-2 agent |
| SWE-Bench Verified | 76.2% | Single attempt |
| τ2-bench | 85.4% | Listed as an agentic tool-use result |
In that article’s comparison, Gemini 3 Pro led the listed LiveCodeBench Pro and Terminal-Bench 2.0 results. Its SWE-Bench Verified score was below Claude Sonnet 4.5 at 77.2% and GPT-5.1 at 76.3%. That is evidence of mixed results by task, not proof that Gemini 3 Pro was universally the best coding model. These are third-party reported figures; the article does not provide enough detail to reproduce every evaluation, and the results do not predict performance on your repository.
A separate Kilo Code page reported scores for different tests, including LiveCodeBench and TerminalBench Hard. Do not merge those numbers with the table above: the benchmark names, attribution, and evaluation details differ. Test models on representative tasks from your own workload rather than averaging results from unrelated evaluations. (Kilo Code’s separate benchmark listing)
Can you still use Gemini 3 Pro through a production API?
Not through Google’s Gemini API. Google says gemini-3-pro-preview was shut down on March 9, 2026. Its deprecations page lists Gemini 3.1 Pro Preview as a recommended replacement; that is a different model, and its specifications should not be transferred to Gemini 3 Pro. (Google’s deprecations page)
Older third-party guides describe selecting Gemini 3 Pro inside Kilo Code or using a Google API key in a Roboflow workflow. Those are descriptions of third-party product routes. They do not override Google’s shutdown notice or establish that the model remains callable through those services. Confirm availability directly with the provider before relying on it.
A documented Google API alternative: Gemini 4 Argon
Google’s Gemini API documentation provides an Interactions API example using Gemini 4 Argon (gemini-4-argon). This is a different model, not a renamed Gemini 3 Pro. The example below shows the documented call pattern; Google labels the Interactions API as Beta, so do not treat this example alone as confirmation of a stable production route. (Google’s Argon guide; Interactions API quickstart)
For the Python example, Google’s documentation lists the google-genai SDK. Set GEMINI_API_KEY in the environment; Google says its client libraries detect that variable automatically. (SDK installation documentation; API-key documentation)
pip install -U google-genai
export GEMINI_API_KEY="YOUR_API_KEY"
from google import genai
client = genai.Client()
interaction = client.interactions.create(
model="gemini-4-argon",
input="Review this function for bugs and suggest a minimal patch: ...",
)
print(interaction.output_text)
The model identifier in this request is Argon—not Gemini 3 Pro. Keep credentials out of source code, and verify the model’s current API support and deployment requirements before using this Beta API route in production.
Cost to budget for
As of November 9, 2026, Google lists Gemini 4 Argon at $2 per 1 million input tokens and $10 per 1 million output tokens through November 30, 2026. The listed rates change to $4 per 1 million input tokens and $20 per 1 million output tokens on December 1, 2026. Recheck Google’s pricing before budgeting, especially for usage after November 30. (Google’s Argon guide and pricing)
At the introductory rates, 40,000 input tokens and 10,000 output tokens would cost about $0.18: $0.08 for input plus $0.10 for output. This is arithmetic based on those token counts and rates, not a forecast for a coding-agent session. Actual cost depends on measured token use and task success.
A practical coding-model pilot
Before adopting Argon—or another currently available model—compare it with your baseline on a small, representative set of tasks:
- Choose work that reflects your team’s needs: a bug fix, test-writing, a refactor, and a task using your actual repository conventions or tools.
- Use the same task instructions, repository snapshot, and tool permissions for each model.
- Define acceptance checks in advance: tests pass, behavior is correct, patch size is acceptable, and there are no new security or compatibility issues.
- Record task completion, human correction time, latency, and actual token cost. Review changes before merging; benchmark scores do not replace code review.
- Decide based on your workload. Results on terminal-agent tasks may not predict performance on your team’s maintenance work or cost target.
Bottom line: Gemini 3 Pro’s historical results help explain interest in the model, but Google shut down its Gemini API endpoint. For a new Google coding integration, evaluate an available model on representative tasks and verify the API’s current status before production adoption.
Sources
Gemini 3 Pro Preview model page
Gemini 3 Pro benchmark and usage article
Google: Gemini 3 Pro Preview Coding Benchmark | Kilo Code
Google launches Gemini 3 with new coding app and record benchmark scores