Gemini 3 Flash Preview vs Qwen3 235B A22B Instruct 2507

Pricing math

Gemini 3 Flash Preview costs $0.50/M input and $3.00/M output at the provider; on OpenKey that's $0.515/M input and $3.09/M output ($0.50 x 1.03, $3.00 x 1.03). Qwen3 235B A22B Instruct 2507 costs $0.09/M input and $0.10/M output at the provider, or $0.0927/M and $0.103/M on OpenKey (same 3% fee applied). Run a 10M-input/2M-output batch job and the totals are $11.00 for Gemini versus $1.10 for Qwen — a 10x gap. Gemini's input price alone is 5.56x Qwen's. If your workload is mostly text and doesn't need reasoning traces or multimodal input, Qwen is the cheaper default by a wide margin.

Coding and agent benchmarks

On Design Arena's model-category benchmarks, Gemini 3 Flash Preview scores an elo of 1239 in codecategories (rank 33, 57.6% win rate) and 1232 in gamedev (rank 38, 58.3% win rate). Qwen3 235B A22B Instruct 2507 scores 1088 in codecategories (rank 81, 42.7% win rate) and 1018 in gamedev (rank 94, 35.2% win rate). Gemini also has agent-arena data Qwen lacks: 1073 elo in agenticslides (rank 9) and 1218 elo in godotgamedev (rank 8), both top-10 finishes. If the task involves autonomous coding agents or full-stack app generation, Gemini's benchmark lead is consistent and Qwen has no comparable agent-arena scores to counter it.

Context and modality

Gemini's context window is 1,048,576 tokens against Qwen's 262,144 — a 4x ratio. Gemini also accepts text, image, file, audio, and video input, while Qwen is text-in/text-out only. Max output differs too: Gemini caps at 65,535 completion tokens, Qwen at 16,384. For long-document analysis, multi-file codebases, or any pipeline that ingests audio or video, Gemini is the only option of the two. For single-document or chat-length text tasks, Qwen's smaller window is not a practical constraint.

Reasoning and tool support

Gemini 3 Flash Preview supports explicit reasoning effort levels (high, medium, low, minimal, default medium) via `reasoning` and `include_reasoning` parameters, plus `tool_choice` and `tools` for agentic calls. Qwen3 235B A22B Instruct 2507 has no reasoning parameter exposed and is not marked as a reasoning model, but it supports a broader sampling parameter set — `top_k`, `min_p`, `repetition_penalty`, `logprobs`, `top_logprobs` — useful for fine-grained decoding control in batch or research pipelines. If you need the model to expose its thinking process or tune effort per request, Gemini is built for that. If you need low-level sampling control at scale, Qwen gives you more knobs.

Questions

Which model is cheaper for large batch jobs?

Qwen3 235B A22B Instruct 2507, by a wide margin. A 10M-input/2M-output job costs $1.10 on Qwen versus $11.00 on Gemini 3 Flash Preview — a 10x difference driven mostly by Gemini's 5.56x higher input price per token.

Does either model handle images or audio?

Only Gemini 3 Flash Preview. It accepts text, image, file, audio, and video input (output is text-only). Qwen3 235B A22B Instruct 2507 is text-in/text-out only, with no other input modalities listed.

Which one wins on coding benchmarks?

Gemini 3 Flash Preview, on every Design Arena code-related metric available: 1239 elo (57.6% win rate, rank 33) in codecategories versus Qwen's 1088 elo (42.7% win rate, rank 81). Gemini also has agentic benchmark data — rank 9 in agenticslides — that Qwen lacks entirely.

How much bigger is Gemini's context window?

4x. Gemini 3 Flash Preview supports 1,048,576 tokens of context against Qwen3 235B A22B Instruct 2507's 262,144. Gemini's max output is also larger: 65,535 tokens versus Qwen's 16,384.