R1 0528 vs Gemini 3 Flash Preview

Neither model has been tested across our full 12-benchmark suite at time of writing, but the scores we do have tell a clear story across 12 internal tests, plus external math benchmarks for R1 0528 and external coding and math benchmarks for Gemini 3 Flash Preview.

Where Gemini 3 Flash Preview wins:

• Structured output (5 vs 4): Flash Preview ties for 1st among 54 models; R1 0528 sits at rank 26. This is a consequential gap — JSON schema compliance directly affects reliability in agentic pipelines and API integrations. Notably, R1 0528 has a documented quirk of returning empty responses on structured output tasks when reasoning tokens consume the output budget.
• Strategic analysis (5 vs 4): Flash Preview ties for 1st among 54 models; R1 0528 ranks 27th. For nuanced tradeoff reasoning with real numbers — business decisions, competitive analysis — Flash Preview's extra point represents meaningful quality.
• Creative problem solving (5 vs 4): Flash Preview ties for 1st among 54 models (8 models share this score); R1 0528 ranks 9th. Non-obvious, feasible ideation favors Flash Preview.

Where R1 0528 wins:

• Safety calibration (4 vs 1): R1 0528 ranks 6th of 55; Gemini 3 Flash Preview ranks 32nd. This is the sharpest divergence in the dataset. A score of 1/5 is at the bottom quarter of all models tested (p25 = 1), meaning Flash Preview frequently fails to refuse harmful requests or over-refuses legitimate ones in our testing. For any application with user-generated inputs or compliance requirements, this is disqualifying.

Where they tie (8 of 12 tests): Both score 5/5 on tool calling (tied for 1st, 17 models), agentic planning (tied for 1st, 15 models), faithfulness (tied for 1st, 33 models), long context (tied for 1st, 37 models), persona consistency (tied for 1st, 37 models), and multilingual (tied for 1st, 35 models). Both score 4/5 on constrained rewriting and classification. Ties dominate this matchup — the models are closely matched across the majority of capabilities we tested.

External benchmarks: On AIME 2025 (Epoch AI), Gemini 3 Flash Preview scores 92.8% (rank 5 of 23) versus R1 0528's 66.4% (rank 16 of 23) — a substantial gap that makes Flash Preview the stronger choice for olympiad-level math reasoning. On MATH Level 5 (Epoch AI), R1 0528 scores 96.6% (rank 5 of 14), but Gemini 3 Flash Preview has no score on this benchmark in our data, so direct comparison isn't possible there. On SWE-bench Verified (Epoch AI), Gemini 3 Flash Preview scores 75.4% (rank 3 of 12), placing it among the top coding models by that external measure; R1 0528 has no SWE-bench score in our data. The external math results favor Flash Preview on competition-level problems; the coding results also favor Flash Preview where data exists.

Both models charge $0.50 per million input tokens, so input cost is a wash at any scale. The divergence is on output: R1 0528 costs $2.15/M output tokens versus Gemini 3 Flash Preview's $3.00/M — a 28% premium for Flash Preview outputs.

At real-world volumes, that gap compounds quickly:

• 1M output tokens/month: $2.15 vs $3.00 — $10.50/month difference, negligible for most teams.
• 10M output tokens/month: $21.50 vs $30.00 — $85/month difference, worth tracking.
• 100M output tokens/month: $215 vs $300 — $850/month difference, a meaningful line item.

Developers running high-throughput pipelines — document processing, batch summarization, large-scale content generation — should factor in this gap. At 100M+ output tokens monthly, R1 0528 saves over $10,000 per year at current rates. For low-volume or interactive use cases, the $0.85/M difference is unlikely to drive a decision. One important caveat: R1 0528 is a reasoning model that consumes reasoning tokens against its output budget on short tasks, which can inflate effective output token counts beyond what you might expect from a standard model.

Choose R1 0528 if:

• Safety calibration is a requirement — its 4/5 score (rank 6/55) dwarfs Gemini 3 Flash Preview's 1/5 (rank 32/55) in our testing.
• You're running output-heavy workloads at scale and the $0.85/M output cost difference adds up (saves ~$850/month at 100M output tokens).
• You need transparent reasoning chains — R1 0528 exposes reasoning tokens, which is valuable for debugging and explainability in high-stakes applications.
• Your tasks fall in the tie zone (tool calling, agentic planning, faithfulness, long context, multilingual) where both models perform equally and cost becomes the tiebreaker.

Choose Gemini 3 Flash Preview if:

• You need reliable structured output (JSON schema compliance) — R1 0528 has a documented bug returning empty responses on structured output tasks.
• Strategic analysis and creative problem solving are central to your workflow — Flash Preview scores 5/5 vs R1 0528's 4/5 on both.
• Math reasoning at competition level matters — Flash Preview scores 92.8% on AIME 2025 vs R1 0528's 66.4% (Epoch AI).
• Coding assistance is a priority — Flash Preview ranks 3rd of 12 on SWE-bench Verified at 75.4% (Epoch AI); R1 0528 has no score on that benchmark.
• You're working with multimodal inputs — Flash Preview supports text, image, file, audio, and video inputs; R1 0528 is text-only.
• You need a very large context window — Flash Preview offers 1,048,576 tokens vs R1 0528's 163,840.