Across 13 comparable benchmarks, Haiku 4.5 has better reported results on 10 and worse on 3.
Moderate evidence13 of 13 shared results countedHow we compare
Price basis: Standard rates. The price section below shows the applicable periods.
| Fact | gpt-oss-20b | Haiku 4.5 |
|---|---|---|
| Context window | 128K tokens | 200K tokens |
| Input / 1M | $0.030 | $1.00 |
| Output / 1M | $0.140 | $5.00 |
| Blended (4:1) | $0.052 | $1.15 |
| Status | Available | Available |
| Released | August 6, 2025 | October 15, 2025 |
Blended is 4:1 input to output with 90% cache hits at listed rates — set your own mix.
Mean of 8 shared benchmarks grouped by capability on a 0–100 scale. How averages are built
Haiku 4.5 ahead on 2 of 2 counted
gpt-oss-20b ahead on 1 of 1 counted
Haiku 4.5 ahead on 2 of 2 counted
Haiku 4.5 ahead on 2 of 2 counted
Haiku 4.5 ahead on 1 of 1 counted
Split 1–1 of 2 counted
No shared comparable results for Abstract reasoning, Research & search, Long context, Cybersecurity.
Published scores, with matching modes where available. Gaps show gpt-oss-20b relative to Haiku 4.5. Reading the results
| Benchmark | gpt-oss-20b | Haiku 4.5 | Gap |
|---|---|---|---|
| SWE-bench VerifiedRepository Engineering · Accuracy | 34.0Thinking · no tools | 73.3≠Thinking · tools · 128K | −39.3 pts#103 vs #47 of 108 |
| Terminal Bench HardAgentic Development · Accuracy | 10.6Thinking High · tools | 27.3≠Normal · tools | −16.7 pts#115 vs #74 of 157 |
| AIME2025Mathematics · Accuracy | 98.7Thinking · tools | 96.3≠Thinking · tools · 128K | +2.4 pts#13 vs #18 of 86 |
| GPQA DiamondScientific Reasoning · Accuracy | 71.5Thinking · no tools | 73.3≠Extended · no tools | −1.8 pts#130 vs #126 of 200 |
| SciCodeScientific Computing · Score | 38.9Thinking High · no tools | 42.2≠Thinking · no tools | 0.92×#79 vs #72 of 98 |
| Pinch BenchTool Orchestration · Accuracy | 66.0Thinking · tools | 82.0Thinking · tools | −16.0 pts#35 vs #22 of 38 |
| PinchBench v2Tool Orchestration · Average score (%) | 36.3Reported best | 67.7Reported best | −31.4 pts#48 vs #33 of 51 |
| GDP.pdfDocuments & Charts · Task pass rate (%) | 2.0Thinking High · no tools | 3.8≠Thinking · no tools | −1.8 pts#84 vs #78 of 96 |
| τ²-BenchService Workflows · Accuracy | 47.7Thinking · tools | 33.0≠Normal · tools | +14.7 pts#34 vs #39 of 40 |
| τ³-BankingService Workflows · Score | 7.0Thinking High · tools | 9.3≠Thinking · tools | 0.75×#99 vs #95 of 120 |
| HLEKnowledge Exams · Accuracy | 17.3Thinking · tools | 9.7≠Extended · no tools | +7.6 pts#114 vs #129 of 151 |
| MMLU-ProKnowledge Exams · Accuracy | 74.0Thinking · no tools | 80.0≠Normal · no tools | −6.0 pts#84 vs #62 of 153 |
| ECICapability Indices · ECI score (capability index, higher is better) | 138Reported best | 142Reported best | 0.97×#106 vs #90 of 168 |
Repository Engineering · Accuracy
Agentic Development · Accuracy
Mathematics · Accuracy
Scientific Reasoning · Accuracy
Scientific Computing · Score
Tool Orchestration · Accuracy
Tool Orchestration · Average score (%)
Documents & Charts · Task pass rate (%)
Service Workflows · Accuracy
Service Workflows · Score
A dash means no result is available.
| Benchmark | gpt-oss-20b | Haiku 4.5 |
|---|---|---|
| AIME 2024 | 96.0 | — |
| CritPt | 1.4 | — |
| FrontierMath | — | 4.1 |
| FrontierMath - Tier 4 | — | 2.1 |
| MMLU | 85.3 | — |
| Code Migration | — | 7.5 |
| CodeForces | 2516 | — |
| LiveCodeBench | — | 62.0 |
| SWE-Bench Pro - Public | — | 39.5 |
| SWE-Bench Pro V2 Hard | — | 25.5 |
| Terminal-Bench 1.0 | — | 41.8 |
| Vibe Code Bench v1.1 | — | 11.4 |
| Claw Bench | — | 89.4 |
| MCP-Atlas | — | 40.2 |
| OSWorld 2.1 | — | 15.7 |
| ExploitBench | — | 13.7 |
| AA Intelligence Index v5.0.1 | — | 7.4 |
| LiveBench | — | 61.3 |
| LiveBench 2.0 | — | 46.0 |
| Vals Index | — | 22.9 |
| Creative Writing | 665 | — |
| IF Bench | — | 54.3 |
| Context Arena | — | 36.3 |
| MMMU | — | 73.2 |
| MMMU-Pro | — | 58.6 |
| AA-AnalystAgent | — | 15.0 |
| AutomationBench | — | 0.46 |
| EMB (Excel Modeling) | — | 22.7 |
| Finance Agent v2 | — | 31.0 |
| Harvey Lab-AA | — | 61.0 |
| Harvey's Legal Agent Benchmark | — | 0.83 |
| Legal Research Bench | — | 10.6 |
| Public Benefits Bench v1.1 | — | 54.3 |
| SAGE | — | 31.8 |
| ARC-AGI-1 | — | 47.7 |
| ARC-AGI-2 | — | 4.5 |
| BALROG | — | 31.2 |
| Vending-Bench 2 | — | 459 |
| BrowseComp | 28.3 | — |
| GAIA | — | 56.4 |
| MedCode | — | 32.7 |
| MedScribe | — | 85.2 |
Time-of-day providers use peak-hour rates; fixed-price providers use standard rates. Off-peak discounts remain in the published details.
Drag the handle above the bar for cache hits, the one below it for the input : output mix · 5% steps
Blended price = effective input rate at 90% cache hits × 80% + output rate × 20%
gpt-oss-20b costs95%lessthan Haiku 4.5 per million tokens — 22.2× cheaper.Using standard rates at this token mix.
gpt-oss-20b: Standard pricing. DeepInfra official models API (openai/gpt-oss-20b) verified 2026-11-13
Haiku 4.5: Standard pricing. Anthropic official pricing (platform.claude.com/docs/en/about-claude/pricing) verified 2026-11-13
| Price / 1M tokens | gpt-oss-20b | Haiku 4.5 | Ratio |
|---|---|---|---|
| Input | $0.03 | $1.00 | 0.03× |
| Output | $0.14 | $5.00 | 0.03× |
| Cached input | — | $0.10 | — |
| Cache write | — | — | — |
| Blended | $0.052 | $1.15 | 0.05× |
Estimates use the selected token mix. Cache writes, tool fees and long-context surcharges are excluded.
Reported cost per task, paired with the score from the same run. Settings and sources may differ; costs do not affect the verdict.
| Benchmark | gpt-oss-20b | Haiku 4.5 |
|---|---|---|
| AA Intelligence Index v5.0.1 | Not reported | |
| ARC-AGI-1 | Not reported | |
| AutomationBench | Not reported | $0.07 / task Average · Pass Rate: 0.46 Reported best (effort unspecified) · zapier-private-heldout |
| LiveBench 2.0 | Not reported | $0.934 / task Average · Overall score (equal-weight mean of 7 suite scores, 0-100): 46.0 Max (task-specific tools) |
| Public Benefits Bench v1.1 | Not reported | $0.2245 / task Average · Criterion pass rate (%): 54.3 Unspecified Reasoning (With Tools) |
| Attribute | gpt-oss-20b | Haiku 4.5 |
|---|---|---|
| Availability | Available | Available |
| Provider | OpenAI | Anthropic |
| Model type | Reasoning model | Multimodal model |
| First release | August 6, 2025 | October 15, 2025 |
| Context window | 128K tokens | 200K tokens |
| Max output | 128K tokens | 64K tokens |
| Total parameters | 21B | — |
| Active parameters | 3.6B | — |
| Architecture | Mixture of Experts (MoE) | Undisclosed |
| Documented capabilities | Reasoning, Multilingual | Reasoning, Multilingual |
| Input modalities | Text | Text, Image |
| Output modalities | Text | Text |
| Weights | Open weights | Closed / API only |
| Licence | Apache 2.0 | — |
| Commercial use | Commercial use allowed | Proprietary |
These are reported-score comparisons, not controlled head-to-head tests.
What people found comparing these models, or using either one. Comments also appear on the models' own pages.
Used gpt-oss-20b or Haiku 4.5? Sign in with Google to share what worked, what did not, and which API or setup you tested.
No community comments on either model yet.
Alternatives with the most benchmark overlap for this matchup.