Claude Haiku 4.5 vs Gemini 3 Flash Preview
Both models target the same niche: fast, cheap models for high-volume agentic and coding work, not frontier reasoning. Claude Haiku 4.5 (Anthropic, released 2025-10-15) trades intelligence for speed at a fixed 200K context. Gemini 3 Flash Preview (Google, released 2025-12-17) is newer, cheaper per token, and ships with a 1,048,576-token context window plus audio and video input. The gap in price and context is large enough to matter for real workloads.
Spec vs spec
| Spec | Claude Haiku 4.5 | Gemini 3 Flash Preview |
|---|---|---|
| Context window | 200K | 1.0M |
| Max output | 64K | 66K |
| Input modalities | text, image, file | text, image, file, audio, video |
| Output modalities | text | text |
| Released | Oct 15, 2025 | Dec 17, 2025 |
| Reasoning | optional | optional |
Pricing
Per 1M tokens. Provider price plus the flat 3% fee — the sum is what you pay.
anthropic/claude-haiku-4.5
Input · 1M tokens
$1.00 + 3%$1.03
Output · 1M tokens
$5.00 + 3%$5.15
Cache read · 1M tokens
$0.100 + 3%$0.103
Cache write · 1M tokens
$1.25 + 3%$1.29
FEE — FLAT, EVERY MODEL3%
google/gemini-3-flash-preview
Input · 1M tokens
$0.500 + 3%$0.515
Output · 1M tokens
$3.00 + 3%$3.09
Cache read · 1M tokens
$0.050 + 3%$0.052
Cache write · 1M tokens
$0.083 + 3%$0.086
FEE — FLAT, EVERY MODEL3%
One workload, priced on both
10M input + 2M output tokens at each model's price, flat 3% fee included.
anthropic/claude-haiku-4.5
$20.60
$20.00 provider + 3%
google/gemini-3-flash-previewCheaper
$11.33
$11.00 provider + 3%
Benchmarks
Design Arena categories where both models have results. Higher Elo and lower rank win.
| Claude Haiku 4.5 | Gemini 3 Flash Preview | |||
|---|---|---|---|---|
| Category | Elo | Rank | Elo | Rank |
| 3D | 1149 | #63 | 1261 | #29 |
| Code | 1164 | #63 | 1239 | #33 |
| Game dev | 1162 | #60 | 1232 | #38 |
| Websites | 1164 | #64 | 1239 | #33 |
Head-to-head preference voting. How we filter and rank
Pricing math on a real workload
A 10M input / 2M output token run costs $20.00 on Claude Haiku 4.5 and $11.00 on Gemini 3 Flash Preview — Gemini is the cheaper model here. On OpenKey (provider price x 1.03), Claude Haiku 4.5 lists at $1.03/M input and $5.15/M output (provider: $1.00/$5.00). Gemini 3 Flash Preview lists at $0.515/M input and $3.09/M output (provider: $0.50/$3.00). The input price ratio is 2.0x — Claude Haiku costs twice as much per input token. Both run through OpenKey with one key and the same flat 3% fee on top of provider list price.
Coding and agentic benchmarks
Gemini 3 Flash Preview has design_arena results across both the `agents` and `models` arenas — including fullstack (rank 17, 47.1% win rate), webapps (rank 17, 49.5%), and godotgamedev (rank 8, 52.3%) — plus strong `models`-arena scores in codecategories (rank 33, 57.6% win rate) and gamedev (rank 38, 58.3% win rate). Claude Haiku 4.5 only has `models`-arena data: codecategories rank 63 (44.8% win rate) and gamedev rank 60 (44.6% win rate) — both worse ranks and lower win rates than Gemini's equivalents. Anthropic's own artificial_analysis numbers put Claude Haiku 4.5 at a 43.9 coding index and 16.4 agentic index, but there's no matching artificial_analysis data for Gemini 3 Flash Preview to compare directly.
Context and long-document work
Gemini 3 Flash Preview's 1,048,576-token context dwarfs Claude Haiku 4.5's 200,000 — a context ratio of 0.19, meaning Haiku's window is under a fifth the size. If your workload involves ingesting large codebases, long transcripts, or multi-document RAG context in a single call, Haiku's 200K ceiling will force chunking sooner. Max output is close either way: 64,000 tokens for Haiku vs 65,535 for Gemini.
Modality differences
Claude Haiku 4.5 accepts text, image, and file input, output text only. Gemini 3 Flash Preview accepts text, image, file, audio, and video input, also outputting text only. If your product needs audio transcription-adjacent tasks or video-frame reasoning in the same call, Gemini is the only one of the two that takes that input at all.
When to pick each
Pick Gemini 3 Flash Preview by default for coding assistants, agentic tool-calling loops, and anything touching large context or non-text input — it's cheaper, wider-context, and scores better across the overlapping design_arena categories. Pick Claude Haiku 4.5 if you specifically need the `top_k` sampling parameter (Gemini's supported_parameters list doesn't include it) or you're locked into Anthropic's tokenizer and reasoning-block format for existing tooling.
Which model for which job
| Use case | Pick | Why |
|---|---|---|
| High-volume agentic tool use | Gemini 3 Flash Preview | Ranked in the `agents` arena with rank 8-24 across categories; Haiku has no agents-arena data at all |
| Large codebase / long-document context | Gemini 3 Flash Preview | 1,048,576 token context vs Haiku's 200,000 — over 5x larger |
| Cost-sensitive high-volume text pipeline | Gemini 3 Flash Preview | $11.00 vs $20.00 for a 10M-in/2M-out run |
| Audio or video input processing | Gemini 3 Flash Preview | Only Gemini accepts audio and video input modalities |
| Existing Anthropic tool-calling stack needing top_k | Claude Haiku 4.5 | top_k is in Haiku's supported_parameters list, not in Gemini's |
Questions
- Which model is cheaper for a typical 10M input / 2M output workload?
- Gemini 3 Flash Preview, at $11.00 total on OpenKey pricing versus $20.00 for Claude Haiku 4.5 — roughly 45% cheaper for that mixed workload.
- How much bigger is Gemini's context window?
- Gemini 3 Flash Preview supports 1,048,576 tokens of context versus Claude Haiku 4.5's 200,000 — a context ratio of 0.19, so Haiku's window is under a fifth the size of Gemini's.
- Does Claude Haiku 4.5 win on any benchmark?
- In the `models` arena, Haiku's asciiart score (rank 30, 49.3% win rate) is its best result, but Gemini isn't benchmarked on asciiart for direct comparison. Across the categories both models share — codecategories and gamedev — Gemini ranks higher in both (33 and 38 vs Haiku's 63 and 60).
- Can both models process images?
- Yes — both Claude Haiku 4.5 and Gemini 3 Flash Preview accept text, image, and file input. Gemini additionally accepts audio and video, which Haiku does not support.