Compare
Four models, one page
Pick the models you are actually choosing between. Every row is scaled within itself, so a bar means something next to its neighbours and nothing across rows.
| Specification | gpt-oss-20bOpenAI | Claude Opus 5Anthropic |
|---|---|---|
| Released | Aug 5, 2025 | Jul 24, 2026 |
| Context window | 131K | 1M |
| Max output | 131K | 128K |
| Input / 1M | $0.03 | $5 |
| Output / 1M | $0.13 | $25 |
| Cost per task | $0.01 | $2.25 |
| Arena Elo | 961 | 1,393 |
| Serving providers | 11 | 5 |
| Parameters | 21.5B | Undisclosed |
| Licence | apache-2.0 | Proprietary |
| Capabilities |
|
|
Metrics side by side
Each row is scaled to the largest value in that row — bars compare within a row, never across rows
- gpt-oss-20b
- Claude Opus 5
Intelligence Index
gpt-oss-20b15.2Claude Opus 563.1Coding Index
gpt-oss-20b20.7Claude Opus 578.0Agentic Index
gpt-oss-20b3.1Claude Opus 559.2Output speed
gpt-oss-20b252 t/sClaude Opus 587 t/sContext window
gpt-oss-20b131KClaude Opus 51MLatency · lower is better
gpt-oss-20b259msClaude Opus 51.12sBlended price / 1M · lower is better
gpt-oss-20b$0.055Claude Opus 5$10Cost per task · lower is better
gpt-oss-20b$0.01Claude Opus 5$2.25
Rows marked “lower is better” still draw a longer bar for a larger number — read the value, not just the length. Arena Elo is in the specification table above instead: it has no meaningful zero, so a bar would flatten the gaps.
View as table
| Metric | gpt-oss-20b | Claude Opus 5 |
|---|---|---|
| Intelligence Index | 15.2 | 63.1 |
| Coding Index | 20.7 | 78.0 |
| Agentic Index | 3.1 | 59.2 |
| Output speed | 252 t/s | 87 t/s |
| Context window | 131K | 1M |
| Latency · lower is better | 259ms | 1.12s |
| Blended price / 1M · lower is better | $0.055 | $10 |
| Cost per task · lower is better | $0.01 | $2.25 |
Evaluation scores
Percentage correct on a common 0–100% scale
- gpt-oss-20b
- Claude Opus 5
GPQA Diamond
gpt-oss-20b68.8%Claude Opus 593.2%Humanity's Last Exam
gpt-oss-20b11.0%Claude Opus 554.9%SciCode
gpt-oss-20b34.4%Claude Opus 555.7%τ²-bench
gpt-oss-20b60.2%Claude Opus 542.1%Terminal-Bench Hard
gpt-oss-20b10.6%Claude Opus 5not measuredLiveCodeBench
gpt-oss-20b77.7%Claude Opus 5not measuredAA-LCR long context
gpt-oss-20b33.3%Claude Opus 575.7%AIME 2025
gpt-oss-20b89.3%Claude Opus 5not measured
A missing bar means that evaluation was not run for that model — it is not a zero.
View as table
| Evaluation | gpt-oss-20b | Claude Opus 5 |
|---|---|---|
| GPQA Diamond | 68.8% | 93.2% |
| Humanity's Last Exam | 11.0% | 54.9% |
| SciCode | 34.4% | 55.7% |
| τ²-bench | 60.2% | 42.1% |
| Terminal-Bench Hard | 10.6% | — |
| LiveCodeBench | 77.7% | — |
| AA-LCR long context | 33.3% | 75.7% |
| AIME 2025 | 89.3% | — |
Estimated from list pricing: 50K input tokens plus 80K output tokens for reasoning models (25K for non-reasoning).