A.I.
Frontier models
Claude Opus 5
Claude Opus 5 is Anthropic’s default start for most workloads: complex agentic coding and enterprise work. The API id is claude-opus-5. Pricing is $5 input and $25 output per million tokens.
- Status
- Archived
- Price
- $5 in, $25 outUS dollars per million tokens (word pieces) in and out, unless noted
Training
Anthropic lists a reliable knowledge and training data cutoff of May 2026. Parameter count, training FLOPs, and the full post-training recipe for this SKU are not published on the models overview or choosing guide.
Inference
Adaptive thinking is supported. Default effort is high. Relative to the current Claude lineup, Opus sits in a moderate latency class. Effort steers how much the model thinks.
Context
The documented context window is 1 million tokens. Max output is 128K tokens. Inputs are text and images; output is text. The model supports multilingual use, vision, and tool use.
Methods
Anthropic’s choosing guide names an effort parameter, tools, and computer-use class workloads for Opus 5.
Benchmarking
Numbers below are from Anthropic’s Fable 5.1 announce (Opus 5 column). OpenAI’s Sol and Astra tables use different harnesses, so treat those as separate prints.
| Eval | Opus 5 | Fable 5.1 | Fable 5 | GPT-5.6 Sol (Anthropic column) |
|---|---|---|---|---|
| Terminal-Bench-Science 0.1 | 29.0% | 52.6% | 24.7% | 22.4% |
| Terminal-Bench 4.0 | 52.3% | 55.8% | 42.0% | 37.3% |
| GDPval-AA v2 (score) | 1824 | 1853 | 1723 | 1711 |
| OSWorld 2.0 partial | 75.4% | 77.9% | 72.9% | - |
| OSWorld 2.0 strict | 39.6% | 41.7% | 36.1% | - |
| Humanity's Last Exam (no tools) | 56.6% | 60.9% | 57.8% | - |
| Humanity's Last Exam (with tools) | 63.6% | 65.0% | 63.8% | - |
| AutomationBench | 26.9% | 31.4% | 17.1% | 19.6% |
| CursorBench 3.2.0 | 70.0% | 73.4% | 70.5% | 67.2% |
Anthropic notes a Terminal-Bench-Science standard error of about ±3.5 to 4.5 points. OSWorld 2.0 is the August 2026 task release, not comparable to older OSWorld 2.0 prints. Anthropic notes biology tasks that hit Fable safeguards were completed by Opus 5.
Hardware
Anthropic trains and runs Claude on AWS Trainium, Google TPUs, and NVIDIA GPUs. Amazon is the primary cloud and training partner. Anthropic reported more than 1 million Trainium2 chips in use as of April 2026. Per-SKU serving silicon is not published. Buyers purchase API tokens, not a named Claude GPU SKU.
Access and pricing
Pricing is $5 input and $25 output per million tokens on the Claude API and partner clouds.
Sources
- Models overview
- Choosing a model
- Introducing Claude Fable 5.1 and Mythos 5.1 (includes Opus 5 capability columns)
- Amazon compute partnership (20 Apr 2026)
- Google/Broadcom compute partnership (6 Apr 2026)