Claude Haiku 4.5 vs Qwen3 235B A22B Instruct 2507
Claude Haiku 4.5 is Anthropic's fast, near-frontier model released October 2025. Qwen3 235B A22B Instruct 2507 is Qwen's mixture-of-experts model with 22B active parameters, released July 2025. Both are text-capable; only Haiku 4.5 handles images and files. The gap that matters most: Qwen3 is priced for volume, Haiku 4.5 is priced for capability. Both run on OpenKey with one key and a flat 3% fee on top of provider list price.
Spec vs spec
| Spec | Claude Haiku 4.5 | Qwen3 235B A22B Instruct 2507 |
|---|---|---|
| Context window | 200K | 262K |
| Max output | 64K | 16K |
| Input modalities | text, image, file | text |
| Output modalities | text | text |
| Knowledge cutoff | — | Jun 30, 2025 |
| Released | Oct 15, 2025 | Jul 21, 2025 |
| Reasoning | 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%
qwen/qwen3-235b-a22b-2507
Input · 1M tokens
$0.090 + 3%$0.093
Output · 1M tokens
$0.100 + 3%$0.103
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%
qwen/qwen3-235b-a22b-2507Cheaper
$1.13
$1.10 provider + 3%
Benchmarks
Design Arena categories where both models have results. Higher Elo and lower rank win.
| Claude Haiku 4.5 | Qwen3 235B A22B Instruct 2507 | |||
|---|---|---|---|---|
| Category | Elo | Rank | Elo | Rank |
| 3D | 1149 | #63 | 1071 | #80 |
| Code | 1164 | #63 | 1088 | #81 |
| Data viz | 1168 | #59 | 1102 | #77 |
| Game dev | 1162 | #60 | 1018 | #94 |
| UI components | 1155 | #60 | 1022 | #86 |
| Websites | 1164 | #64 | 1101 | #83 |
Head-to-head preference voting. How we filter and rank
Pricing math
Provider list price: Claude Haiku 4.5 is $1.00/M input, $5.00/M output. Qwen3 235B is $0.09/M input, $0.10/M output. On OpenKey that's Haiku 4.5 at $1.03/M input and $5.15/M output (1.03x for the 3% fee), Qwen3 at $0.0927/M input and $0.103/M output.
For a 10M-input / 2M-output workload: Haiku 4.5 costs $20.00, Qwen3 costs $1.10. That's the `input_price_ratio` of 11.11x on input tokens playing out at scale — Qwen3 is roughly 18x cheaper on this specific workload. If you're processing millions of tokens a day and don't need vision or top-tier code quality, that delta compounds fast.
Coding and UI generation
On Design Arena's model arena, Haiku 4.5 beats Qwen3 235B on every category both models were scored in: codecategories (1164 elo, rank 63 vs 1088 elo, rank 81), dataviz (1168 elo, rank 59 vs 1102 elo, rank 77), gamedev (1162 elo, rank 60 vs 1018 elo, rank 94), uicomponent (1155 elo, rank 60 vs 1022 elo, rank 86), website (1164 elo, rank 64 vs 1101 elo, rank 83), and 3d (1149 elo, rank 63 vs 1071 elo, rank 80). Haiku 4.5 also has Artificial Analysis coding_index 43.9 and agentic_index 16.4, metrics not reported for Qwen3. If code generation or UI output quality is the bottleneck, Haiku 4.5 is the safer default.
Context and output length
Qwen3 235B has the larger context window at 262,144 tokens versus Haiku 4.5's 200,000 — a context_ratio of 0.76 (Haiku's window is 76% of Qwen3's). But Haiku 4.5 allows longer completions: 64,000 max output tokens versus Qwen3's 16,384. If you're summarizing very long documents and need long generated responses (long reports, extended code files), Haiku 4.5's output ceiling matters more than the modest context difference. If you're just doing long-context retrieval with short answers, Qwen3's larger window is the edge.
Modality and parameters
Haiku 4.5 accepts text, image, and file input; Qwen3 235B is text-only. Haiku 4.5 also supports `reasoning` and `include_reasoning` parameters (reasoning is optional, not mandatory), useful for step-by-step tasks without forcing a slower reasoning mode. Qwen3 235B supports a wider set of sampling controls (`min_p`, `top_logprobs`, `logit_bias`, `repetition_penalty`) plus prompt caching is unavailable on either — Haiku 4.5 does offer cache pricing ($0.10/M read, $1.25/M write provider-side) which Qwen3 doesn't expose. If your pipeline needs image or document input at all, Qwen3 is disqualified outright.
Which model for which job
| Use case | Pick | Why |
|---|---|---|
| High-volume text classification/extraction | Qwen3 235B A22B Instruct 2507 | Costs $1.10 vs $20.00 for a 10M-in/2M-out workload |
| UI component or website code generation | Claude Haiku 4.5 | Higher elo on uicomponent (1155 vs 1022) and website (1164 vs 1101) |
| Document or image analysis | Claude Haiku 4.5 | Only Haiku 4.5 accepts image and file input |
| Long single-document retrieval | Qwen3 235B A22B Instruct 2507 | 262,144 token context vs 200,000 |
| Long-form generation (reports, full files) | Claude Haiku 4.5 | 64,000 max output tokens vs 16,384 |
| Agentic tool-use workflows | Claude Haiku 4.5 | Reports an agentic_index of 16.4; no equivalent score exists for Qwen3 |
Questions
- Which model is cheaper on OpenKey?
- Qwen3 235B A22B Instruct 2507, by a wide margin. On OpenKey it's $0.0927/M input and $0.103/M output versus Haiku 4.5's $1.03/M input and $5.15/M output. A 10M-input/2M-output run costs $1.10 on Qwen3 versus $20.00 on Haiku 4.5.
- Which model wins on coding benchmarks?
- Claude Haiku 4.5. On Design Arena's codecategories test it scores 1164 elo (rank 63) versus Qwen3's 1088 elo (rank 81). Haiku 4.5 also has an Artificial Analysis coding_index of 43.9, a metric not reported for Qwen3.
- Does either model support image input?
- Only Claude Haiku 4.5. Its input modalities are text, image, and file; Qwen3 235B A22B Instruct 2507 accepts text only. If your workload involves screenshots, PDFs, or scanned documents, Qwen3 is not an option.
- Which has the bigger context window?
- Qwen3 235B A22B Instruct 2507, at 262,144 tokens versus Haiku 4.5's 200,000 — Haiku's window is about 76% the size of Qwen3's. Haiku 4.5 compensates with a much larger max output of 64,000 tokens versus Qwen3's 16,384.