Short answer
For a new Google Gemini API integration, don’t use gemini-3-pro-preview: Google says that identifier was shut down on March 9, 2026, and directs developers to Gemini 4 Argon (gemini-4-argon). This lifecycle notice concerns that specific identifier in the Google Gemini API; it does not establish that Gemini 3 Pro is unavailable through every product or access channel. (Google’s Gemini 3 Pro Preview documentation)
There is positive—but limited—coding-agent evidence for Gemini 3 Pro. Amp said it made Gemini 3 Pro its smart agent’s main model and praised its instruction-following, tool use, and codebase-pattern adherence. Amp also reported a 17-percentage-point increase in an initial internal Terminal-Bench 2.0 run after switching from Sonnet 4.5. That is Amp’s own preliminary result, not an independent or universal comparison; the article supplies no absolute scores or full test setup. (Amp’s account) For a current Google API integration, evaluate the documented current model against your own repository tasks rather than treating older reports as proof that Gemini 3 Pro remains available or is the best fit.
What the coding evidence says
Amp’s account is relevant because it describes a coding agent working with tools and an existing codebase—not just a model generating a snippet in a chat window. The team praised the model’s instruction-following, tool use, collaboration, and ability to follow established codebase patterns. Its Terminal-Bench result is worth noting, but only with the limits attached: it was an internal, explicitly not-even-optimized run, and no test configuration or absolute scores were provided. (Amp’s account)
Amp also reported problems that matter if an agent can run commands or alter a repository: occasional indefinite thinking, excessive thinking-like output, malformed or leaked output, reluctance to execute bash, unrequested commits, and non-absolute file paths. These are qualitative observations, not measured failure rates; they are reasons to test workflow behavior, not claims that every user will see these failures. (Amp’s account)
Other early user reports are not a firm tie-breaker. A 2025 Hacker News discussion contains conflicting opinions about Gemini, Claude, and GPT-5, and commenters questioned whether some circulating tests actually used Gemini 3. Those comments are anecdotal and sometimes refer to Gemini 2.5 or unspecified versions, so they should not be read as verified Gemini 3 Pro results. (Hacker News discussion)
The API decision: use a current model ID
Google identifies gemini-4-argon as its current model and explicitly recommends it as the migration destination for the shut-down gemini-3-pro-preview identifier. Google’s API version documentation says the Interactions API’s core features are generally available in stable v1; its REST reference also shows a v1beta endpoint and says stable v1 is available. For a new integration, use the stable version unless you need a beta-only feature. (Gemini 3 Pro Preview lifecycle notice; API versions; Interactions API REST reference)
A minimal direct-to-Google REST request using the stable v1 route is:
curl -X POST "https://generativelanguage.googleapis.com/v1/interactions" \
-H "x-goog-api-key: $GEMINI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "gemini-4-argon",
"input": "Review this function for correctness and propose a minimal fix: ..."
}'
The documented request uses the model and input fields and an API-key header. output_text is an SDK convenience property, not a REST JSON field. This example demonstrates the request, not how to extract text from the REST response; use Google’s Interactions API response schema for REST response handling. Google’s SDKs expose interaction.output_text as a convenience for accessing the last text blocks in the model’s response. (Interactions API REST reference; text generation documentation) Keep GEMINI_API_KEY in your deployment’s secret manager rather than hard-coding it.
For a direct Google API integration, confirm current model details and billing terms in Google’s documentation. For Token Forge Cloud, confirm the exact model, identifier, endpoint, and billing terms with Token Forge before building around it. Its public homepage lists Qwen, DeepSeek, GLM, and Seedance APIs; that does not confirm Gemini availability, nor does it rule out access through another channel. (Token Forge Cloud)
Cost and data-use terms for Google Gemini 4 Argon
Google’s pricing information lists introductory Gemini 4 Argon rates through November 30, 2026 of $2 per 1 million input tokens and $10 per 1 million output tokens; the listed rates afterward are $4 per 1 million input tokens and $20 per 1 million output tokens. These are token-based API rates, not a flat subscription price. Confirm the current pricing page before committing spend, especially if requests use separately billable features.
Worked arithmetic: a request with 100,000 input tokens and 20,000 output tokens costs about $0.40 at the introductory rates: (100,000 ÷ 1,000,000 × $2) + (20,000 ÷ 1,000,000 × $10). At the listed later rates, the same token counts would be about $0.80. This is arithmetic from the listed per-token rates, not a measured request; it excludes any separately priced features.
For Paid Services, Google says it does not use prompts or responses to improve its products. That is not a zero-retention promise: Google says Paid Services prompts and responses may be logged for a limited period solely to detect prohibited-use violations. The Interactions API also stores conversation state by default; setting store: false opts out of that default state storage, but should not be treated as disabling abuse-monitoring logging. Google requires Paid Services when making API clients available to users in the EEA, Switzerland, or the UK. Check the current terms, data-retention documentation, regional availability, and account eligibility before sending proprietary code.
Google describes the Paid tier as intended for production applications, with higher rate limits. Paid-tier access requires linking a billing account and prepaying. (Billing documentation)
How to choose for your coding workflow
Run the current model and your existing alternative against the same representative tasks. For a useful pilot, include a small bug fix, a change that must follow repository conventions, a multi-file task, and an agent task that uses tools. Keep the prompt, repository snapshot, tool permissions, and success criteria consistent.
Score each run on:
- Correctness: Do tests pass, and does the change solve the requested problem?
- Instruction and repository fit: Did it honor constraints and follow existing patterns?
- Tool behavior: Did it use permitted tools appropriately and avoid unauthorized changes?
- Recovery: If it made a mistake, could it correct course without looping or losing requirements?
- Cost and latency: What did successful tasks actually cost and how long did they take, including retries?
For an agent with shell or write access, inspect proposed diffs and require approval for consequential actions such as commits or deployments. Treat those as sensible evaluation controls, not as evidence that any particular model’s failure rate is known.
Bottom line: Amp’s reported Gemini 3 Pro coding-agent results explain why developers evaluated it, but Google’s gemini-3-pro-preview API identifier was shut down on March 9, 2026. For a new Google API integration, use the current documented model ID and stable route, then decide based on a consistent test of your own coding work, costs, and operational requirements.
Sources
Gemini 3 Pro Preview — Gemini API documentation
What's new in Gemini 4 Argon — Gemini API
API versions explained — Gemini API
Gemini Interactions API — REST reference
Text generation — Interactions API
Gemini API Additional Terms of Service
Zero data retention in the Gemini Developer API