Short answer
There is no verified overall coding winner—and, for the likely comparison of Gemini 3 Pro Preview versus Claude Sonnet 4.5, there is no longer a live production API head-to-head to run. Google says Gemini 3 Pro Preview (gemini-3-pro-preview) was shut down on March 9, 2026. Anthropic says Claude Sonnet 4.5 is deprecated but still functional, and is scheduled to retire on November 30, 2026. As of November 12, 2026, Sonnet 4.5 is therefore a short-term legacy option, not a sound default for a new integration. (ai.google.dev)
The coding evidence captured for these model families does not establish a fair, reproducible winner for that exact pair. Some published figures concern successor models or different benchmarks; user reports are anecdotal. For a production decision, choose model versions that are currently supported and compare them on your own representative coding tasks rather than carrying a historical Gemini 3-versus-Claude 4.5 verdict into a new build.
First clarify which models “Gemini 3” and “Claude 4.5” mean
The family names are ambiguous. Google’s API documentation identifies Gemini 3 Pro Preview as gemini-3-pro-preview; the page says it accepted up to 1,048,576 input tokens and produced up to 65,536 output tokens, but also explicitly marks it shut down. (ai.google.dev)
Anthropic’s Claude Sonnet 4.5 documentation lists the API ID claude-sonnet-4-5-20250929 (and the alias claude-sonnet-4-5), a 200,000-token context window, and a 64,000-token maximum output. It is deprecated, still functional according to Anthropic, and scheduled to retire on November 30, 2026. (platform.claude.com)
“Claude 4.5” might instead mean Claude Opus 4.5. That is a different model tier, so results or costs for Sonnet should not be applied to it. Anthropic’s current pricing page lists Opus 4.5 separately from Sonnet 4.5. (platform.claude.com)
| Model named in this comparison | API status and identifier | Documented token limits |
|---|---|---|
| Gemini 3 Pro Preview | Shut down March 9, 2026; gemini-3-pro-preview |
1,048,576 input; 65,536 output |
| Claude Sonnet 4.5 | Deprecated; still functional, retirement scheduled November 30, 2026; claude-sonnet-4-5-20250929 |
200,000 context; 64,000 maximum output |
These specifications clarify the historical pairing; they do not imply that either model is the right choice for a new production integration. (ai.google.dev)
Which is better at coding?
The available evidence does not justify declaring Gemini 3 Pro or Claude Sonnet 4.5 the overall coding winner. The captured comparison material does not provide a controlled, reproducible head-to-head test of those exact versions, with the same repository, task, tools, and scoring method.
A later comparison article reports results for Gemini 3.1 Pro and Claude Opus 4.7, not the models in this comparison. It gives scores from different benchmarks—SWE-Bench Verified and Terminal-Bench 2.0 for Gemini 3.1 Pro, and SWE-Bench Pro for Opus 4.7. Those figures cannot be used as a direct score comparison, or as evidence of how Gemini 3 Pro compares with Claude 4.5. (ai.google.dev)
Likewise, one user’s experience with Gemini Flash and Claude Sonnet or Opus does not settle a Pro-versus-Sonnet comparison: Flash is a different Gemini tier, and the reports describe particular tasks rather than a controlled evaluation. Treat such accounts as anecdotes, not a ranking.
For specific workloads, the honest answer is therefore:
- Repository-level changes, debugging, and refactoring: no supported winner for the exact pair from the evidence available here.
- UI or front-end implementation: no supported winner for the exact pair; results from Gemini Flash anecdotes should not be generalized to Gemini 3 Pro.
- Production API choice today: Gemini 3 Pro Preview is shut down, while Sonnet 4.5 is near its announced retirement. That lifecycle difference matters more than a historical coding preference when planning a new integration. (
ai.google.dev)
API pricing and production trade-offs
Anthropic lists Claude Sonnet 4.5 at $3 per million input tokens and $15 per million output tokens (USD). That is a token rate—not a guaranteed cost for completing a coding task. Retries, longer prompts, generated code, and any tools used can change the total. Anthropic’s page also lists prompt-caching rates and says Batch API input and output receive a 50% discount. (platform.claude.com)
There is no current Gemini 3 Pro API rate to compare against: Google marks that model shut down. Historical pricing for Gemini 3 Pro, or prices for Gemini successor models, should not be presented as the current cost of calling gemini-3-pro-preview. (ai.google.dev)
If by “Claude 4.5” you meant Opus 4.5, Anthropic’s pricing page lists it at $5 per million input tokens and $25 per million output tokens (USD). That does not revive the Gemini 3 Pro comparison: the Google model remains shut down, and Opus is not Sonnet. (platform.claude.com)
Practical choice: avoid starting a new integration on a model with a stated retirement date. If you already use Sonnet 4.5, account for migration before November 30, 2026. For Google, select a currently supported Gemini API model rather than Gemini 3 Pro Preview. Provider model pages and lifecycle notices should be the source of truth for current availability and migration planning. (ai.google.dev)
How to compare supported models for your codebase
Use this as a proposed evaluation—not a test already run for this article:
- Choose the exact model versions and API routes. Record the pinned model ID, provider, and any coding agent or tool setup. Do not compare Gemini Pro with Flash or Claude Sonnet with Opus and describe the result as a single family-versus-family verdict.
- Select representative work. Include real tasks from your queue, such as a bug fix in an existing repository, a multi-file change, a test-writing task, and a UI adjustment if that reflects your workload.
- Hold conditions steady where possible. Give each model the same task description, repository snapshot, relevant context, tool permissions, and time or token budget. Keep acceptance criteria fixed.
- Score outcomes, not impressions alone. Record whether tests pass, whether the patch meets the request, how many corrections or retries were needed, and how much human editing was required. Also record latency and API cost if they affect the product.
- Repeat before deciding. A single success or failure can be noisy. Compare multiple tasks and review the resulting patches against the same criteria.
This method can tell you which currently supported model fits your workload. It cannot recreate a Gemini 3 Pro API comparison now that Google has shut that endpoint down.
Bottom line
For Gemini 3 Pro Preview versus Claude Sonnet 4.5, the evidence does not establish a coding winner, and the pair is no longer suitable for a new, ongoing production API comparison: Gemini 3 Pro Preview shut down on March 9, 2026, and Sonnet 4.5 is scheduled to retire on November 30, 2026. Treat older benchmark claims and user experiences as historical context. For a new integration, compare currently supported model versions on matched tasks from your own codebase, and include lifecycle, reliability, and cost alongside code quality. (ai.google.dev)
Sources
Gemini 3 Pro preview — Gemini API | Google AI for Developers
Claude Sonnet 4.5 — Claude Platform Docs
Pricing — Claude Platform Docs
Gemini 3 Flash better than Claude Sonnet 4.5? — Cursor Community Forum
GPT 5.1 vs Claude 4.5 Sonnet vs Gemini 3 Pro vs DeepSeek-V3.2