Short answer
For a coding or agent workload, Claude Opus 4.5 is the first model to pilot if you specifically need to compare these two exact API endpoints: Anthropic documents claude-opus-4-5-20251101 as Active (legacy) and still available. But it is a short-horizon choice: Anthropic says retirement will be no sooner than November 24, 2026, and recommends considering migration. Google’s gemini-3-pro-preview endpoint on the Gemini API was shut down on March 9, 2026, so it is not a candidate for a new integration through that API. The Google finding applies to that model ID and API channel, not every Gemini 3 model or hosting route.
The performance case for trying Opus on coding and engineering-agent tasks is a pilot hypothesis, not a general verdict: a competitor comparison recommends it for difficult coding and agents, while one hands-on report found mixed results across two coding-agent tasks. For a durable production decision, evaluate currently available endpoints rather than treating this retired Gemini endpoint as a live alternative. (platform.claude.com)
Which models and API paths are being compared?
This comparison refers to Claude Opus 4.5 on Anthropic’s direct Claude API (claude-opus-4-5-20251101) and Gemini 3 Pro Preview on Google’s Gemini API (gemini-3-pro-preview). Anthropic’s model reference labels Opus 4.5 “Active (legacy),” says it remains available, and gives a retirement commitment of no sooner than November 24, 2026. That is a documented candidate for a near-term direct-API pilot, but not a reason to assume a long-term production runway. Status and retirement timing can differ on partner-operated platforms. (platform.claude.com)
Google’s model documentation says gemini-3-pro-preview was shut down on March 9, 2026. Do not send new Gemini API requests to that endpoint. This does not establish the status of other Gemini models, versions, or partner-hosted offerings.
Capability and workload trade-offs
| Decision factor | Claude Opus 4.5 | Gemini 3 Pro Preview | What the evidence establishes |
|---|---|---|---|
| Coding and agents | A competitor comparison recommends it for difficult coding, deep reasoning, and engineering agents. | The same comparison emphasizes multimodal and search-oriented work. | Attributed recommendations, not a controlled head-to-head result. |
| Context and modalities | A competitor article lists a 200,000-token context window. | A competitor article lists a 1,000,000-token window and broader modalities. Google’s API model page lists a 1,048,576-token input limit for this preview model. | Historical or attributed specifications; the Google Gemini API endpoint is shut down. Do not treat these as current-model specifications. |
| API availability | Anthropic documents the exact model as Active (legacy), still available, with retirement no sooner than November 24, 2026. | Google says the exact Gemini API preview model was shut down March 9, 2026. | The endpoints do not have equal availability for a new direct-API deployment. |
The workload positioning and the 200,000-token Opus figure above come from a competitor article, not a current head-to-head evaluation. It recommends Opus for difficult coding and long-horizon agents, and Gemini 3 Pro for multimodal, long-context, or search-grounded work. (acecloud.ai)
What the hands-on coding report found
A Composio author-run comparison used Claude Code and Gemini CLI on the same Next.js Kanban repository. On a feature-build task involving a fallback, a 10-minute cache, UI behavior, and tests, the author reported that Gemini’s implementation worked; Opus’s implementation ran but had a cache-related UI defect. The author selected the best of three responses for each model on this task, so this is not a typical single-run result.
On a separate GitHub-triage agent task using Composio Tool Router, Opus produced a working demo with flaws, while Gemini repeatedly looped and did not produce a usable final build. The author reported costs of about $0.45 for Gemini versus $2.21 for Opus on the feature task, and approximately $6.30 for Gemini versus $2.88 for Opus on the agent task. These are costs from those particular runs—not API rate cards or general cost estimates. The author explicitly cautions against treating two tasks in one repository as a hard rule.
Taken together, the report is mixed: Gemini did better in the reported feature task, while Opus had the better reported outcome on the separate agent task. It does not establish a universal winner. The Gemini results are retrospective because the tested Gemini API preview endpoint is shut down. Composio hands-on comparison
API pricing
Anthropic’s Claude Platform documentation lists Opus 4.5 at $5 per million input tokens and $25 per million output tokens for standard pricing. It also lists prompt-cache writes at $6.25 per million tokens for 5-minute writes and $10 per million tokens for 1-hour writes, and cache reads at $0.50 per million tokens. Batch API requests receive a 50% discount on input and output. These prices help estimate usage; they do not extend the model’s retirement commitment. (platform.claude.com)
A competitor comparison gave historical Gemini 3 Pro Preview estimates of about $2 per million input tokens and $12 per million output tokens for standard API usage at prompts up to 200,000 tokens. Since Google says the compared Gemini API endpoint was shut down, do not use those figures as pricing for a currently usable Gemini API model. (acecloud.ai)
Token rates alone do not determine cost per successful task. Retries, tool use, input and output volume, caching, and human correction all affect the total. The author-run costs above show how task outcomes can differ, but should not be reused as estimates.
Which should you pilot?
- Coding or engineering-agent evaluation: If you need to test one of these exact endpoints, start with Opus 4.5 as a near-term pilot hypothesis. Anthropic documents it as still available, and the attributed workload guidance favors it for difficult coding and agent work. Treat the retirement commitment—no sooner than November 24, 2026—as a real planning constraint, not a guarantee of long-term availability.
- New integration with Google’s Gemini API: Do not use
gemini-3-pro-preview; Google says that endpoint was shut down. The evidence here does not establish which other Gemini 3 endpoint, if any, is suitable for your deployment. - Before production: Compare currently available model endpoints on the same representative prompt, data, tools, and acceptance criteria. Track correctness, failure and retry rate, end-to-end latency, and actual cost per accepted result. Set quality and cost thresholds before testing.
Bottom line
For the exact direct-API pair in this comparison, Opus 4.5 is the documented near-term pilot candidate for coding and agent workloads; Gemini 3 Pro Preview is not available on Google’s Gemini API. Opus is marked Active (legacy), with retirement no sooner than November 24, 2026, so it should not be treated as a long-term default without a migration plan. The captured hands-on evidence is task-specific and mixed; it does not prove a universal winner. For a production commitment, test currently available endpoints against your own workload.
Sources
Claude Opus 4.5 — Claude Platform Docs
Model deprecations — Claude Platform Docs
Gemini 3 Pro Preview — Gemini API
Claude Opus 4.5 vs Gemini 3 Pro vs Sonnet 4.5 — AceCloud
Claude Opus 4.5 vs Gemini 3 Pro: hands-on coding comparison — Composio