Comparison of moonshotai/kimi-k2.7-code,
minimax/minimax-m3,
google/gemini-3-flash-preview, and
google/gemini-3.5-flash for Tegy's OpenRouter + Cloudflare AI Gateway path.
Prepared June 14, 2026. Prices and metadata are current OpenRouter listings unless otherwise noted. Anthropic and OpenAI models are intentionally excluded.
All four candidates support image input, so they should not produce the OpenRouter
No endpoints found that support image input failure that happened with
qwen/qwen3.7-max. For a production trial, I would test
minimax/minimax-m3 first if cost and context matter most, and
google/gemini-3.5-flash first if benchmark quality matters most.
| Model | OpenRouter listed / release date | Input modalities | Context | Output cap | OpenRouter pricing per 1M tokens | Supported agent parameters |
|---|---|---|---|---|---|---|
Kimi K2.7 Codemoonshotai/kimi-k2.7-code
|
June 12, 2026 | textimage | 262,144 tokens | 262,144 tokens |
Input $0.75 Output $3.50 Cache read $0.16 |
toolstool_choicereasoningstructured_outputs |
MiniMax M3minimax/minimax-m3
|
May 31, 2026 | textimagevideo | 1,048,576 listed; top provider shows 524,288 | 512,000 tokens |
Input $0.30 Output $1.20 Cache read $0.06 |
toolstool_choicereasoningstructured_outputs |
Gemini 3 Flash Previewgoogle/gemini-3-flash-preview
|
December 17, 2025 | textimagefileaudiovideo | 1,048,576 tokens | 65,536 tokens |
Input $0.50 Output $3.00 Image $0.50 Audio $1.00 |
toolstool_choicereasoningstructured_outputs |
Gemini 3.5 Flashgoogle/gemini-3.5-flash
|
May 19, 2026 on OpenRouter; Google model card published May 2026 | textimagefileaudiovideo | 1,048,576 tokens | 65,536 tokens |
Input $1.50 Output $9.00 Image $1.50 Audio $3.00 |
toolstool_choicereasoningstructured_outputs |
| Model | Artificial Analysis indices via OpenRouter | Other benchmark / model-size signal | Read |
|---|---|---|---|
| Kimi K2.7 Code | Not exposed in the OpenRouter model metadata at report time. | Cloudflare Workers AI docs list 1T total parameters, 262.1k context, vision input, tool calling, reasoning, and structured outputs. Kimi's release note reports +21.8% on Kimi Code Bench v2 over K2.6, but that is a vendor-family delta rather than a cross-model benchmark. | Newest and transparent on parameter scale. Benchmark coverage is thinner than Gemini/MiniMax in the public metadata. |
| MiniMax M3 |
Intelligence 54.7 Coding 43.4 Agentic 68.6 |
MiniMax reports SWE-Bench Pro 59.0%, Terminal-Bench 2.1 66.0%, SWE-fficiency 34.8%, KernelBench Hard 28.8%, and MCP Atlas 74.2%. | Best cost/context profile here, and the OpenRouter agentic index is better than Qwen Max's 66.6 baseline. |
| Gemini 3 Flash Preview |
Intelligence 46.4 Coding 42.6 Agentic 49.7 |
Google's Gemini 3 developer guide lists 1M / 64k context and $0.50 / $3 pricing. The Gemini 3.5 model card gives Gemini 3 Flash baseline scores including Terminal-bench 58.0%, SWE-Bench Pro 49.6%, MCP Atlas 62.0%, CharXiv Reasoning 80.3%, and MMMU-Pro 81.2%. | Capable and cheap, but mostly superseded by 3.5 Flash for this shortlist. |
| Gemini 3.5 Flash |
Intelligence 55.3 Coding 45.0 Agentic 70.3 |
Google model card reports Terminal-bench 2.1 76.2%, SWE-Bench Pro 55.1%, MCP Atlas 83.6%, OSWorld-Verified 78.4%, CharXiv Reasoning 84.2%, and MMMU-Pro 83.6%. | Best quality signal of the four. The tradeoff is price: output is 7.5x MiniMax M3 and 2.57x Kimi K2.7 Code. |
It is multimodal, cheap, long-context, explicitly oriented toward long-horizon agentic work, and has a stronger OpenRouter agentic index than the Qwen Max baseline. It is the most pragmatic candidate if the goal is to keep costs low while fixing image turns.
It has the best public benchmark signal in this comparison and broad multimodal input support, including files. The downside is cost, especially output tokens and internal reasoning.
Kimi is attractive because it is fresh, vision-capable, tool-capable, and large. It needs a real Tegy smoke test because its cross-model benchmark metadata is less complete.
It has the same broad modality family and lower pricing than 3.5 Flash, but the benchmark gap is clear enough that it should not be the default unless cost or availability forces it.
Image capability in OpenRouter metadata is necessary but not sufficient. Before switching Tegy, run the exact Claude Agent SDK path through Cloudflare AI Gateway with: text-only chat, image upload, PDF/DOCX Sprinta analysis, natural questionnaire creation, tool/plugin usage, and cache-status capture.
The prior crash was caused by a text-only endpoint rejecting image blocks. This report compares models that advertise image input; it does not prove their behavior under Tegy's complete Claude SDK loop.