1.0M contextVisionTool callingReasoningPrompt cachingPartner-served

Pricing

USD per 1M tokens
Input$5.00
Output$25.00
Cached input$0.5090% below fresh input
Cache write (5m)$6.25
Cache write (1h)$10.00

Billed per token with no minimum and no monthly fee. Price your workload

Capabilities and limits

Context window
1.0M tokens
Max output
63K tokens
Modality
multimodal
Tool calling
Supported
Vision
Supported
Reasoning
Optional
Reasoning effort
low, medium, high, xhigh, max (default: high)
Prompt caching
Supported (platform_managed)
Cache minimum
512 tokens

Measured performance

7d
Time to first token (p50)
3.0 s
Time to first token (p95)
6.0 s
Throughput (avg)
95.3 tok/s
Throughput (p50)
150.0 tok/s

Measured on real Run BiOS traffic over the trailing 7d. Your figures will vary with prompt shape and region.

Call it now — OpenAI-compatible
curl https://api.runbios.ai/v1/chat/completions \
  -H "Authorization: Bearer $BIOS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-opus-5",
    "messages": [{ "role": "user", "content": "Hello" }]
  }'

Existing OpenAI SDK code works by changing the base URL and the key. API overview

Model card

Curated by Run BiOS · benchmark figures, where present, are the card’s own, not our measurements

Claude Opus 5 is Anthropic's strongest Opus-tier model — a hybrid reasoning model built for serious coding and AI agents, featuring a 1M-token context window. Anthropic describes it as a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price, and as a step-change improvement for the Opus tier: stronger coding, more capable agents, and sharper professional work. It is an upgrade to Claude Opus 4.8, with gains in agentic coding, computer use, and long-horizon knowledge work, as well as improvements in mathematical and scientific reasoning.

On coding and knowledge-work evaluations such as Frontier-Bench and GDPval-AA, Opus 5 is the state of the art among the models Anthropic compared it against, though it remains behind Claude Mythos 5 on cybersecurity tasks.

Capabilities

  • Effort settings — The model's effort level controls how hard it works: it reaches its best performance at higher effort on the highest-value tasks, and runs at much lower cost on tasks that do not need the ceiling of its capabilities. On Run BiOS, Opus 5 exposes the low, high, and max reasoning-effort levels.
  • Long context — a 1M-token context window.
  • Multimodal input — text and image input, with text output.
  • Tool use and agents — built for production agentic workflows: orchestrating complex multi-tool tasks, planning deliberately, verifying its own work, and sustaining long-running work autonomously.
  • Prompt caching and batch processing — Anthropic advertises up to 90% cost savings with prompt caching and 50% with batch processing on its own API.

Benchmark results

Results reported in the Claude Opus 5 system card and Anthropic's launch materials. Unless noted, coding results are averages over five trials in Anthropic's standard system-card configuration.

Benchmark Claude Opus 5 Reference points
SWE-bench Verified 96.0% Claude Fable 5: 95.0% · Claude Opus 4.8: 88.6%
SWE-bench Pro 79.2% Claude Fable 5: 80.0% · Claude Opus 4.8: 69.2%
SWE-bench Multilingual 89.5%
SWE-bench Multimodal 59.4%
FrontierCode 1.1 Main 53.4% (medium effort) Claude Fable 5: 53.5%
FrontierCode 1.1 Extended 63.6% best verified result
Frontier-Bench v0.1 43.3% Claude Fable 5: 33.7% · Claude Opus 4.8: 18.7%
ARC-AGI-3 30.2% Claude Opus 4.8: 1.5%
OSWorld 2.0 (computer use) 70.6%
CursorBench 3.2 70.0% (max effort) Claude Fable 5: 70.5%
GDPval-AA v2 (knowledge work) 1861 Elo over 100 points ahead of Claude Fable 5 in Artificial Analysis' run
AA-Briefcase (agentic knowledge work) 1720 Elo +146 over Claude Fable 5 in Artificial Analysis' run

Anthropic's launch reporting adds:

  • On Frontier-Bench v0.1, Opus 5 more than doubles Opus 4.8's performance at a lower cost per task.
  • On ARC-AGI 3, an evaluation of novel problem solving, Opus 5's score is three times as high as the next-best model Anthropic compared against.
  • On Zapier AutomationBench, which measures whether models complete business tasks end to end, Opus 5's pass rate is around 1.5× the next-best model for the same cost per task; even at its lowest effort setting it passes more tasks than any other model in the comparison.
  • On OSWorld 2.0, Opus 5 outperforms every other model at any given cost, surpassing Fable 5's best result at just over a third of the cost.
  • On Anthropic's internal life-sciences evaluations, Opus 5 improves on Opus 4.8 on every one — most notably organic-chemistry tasks such as inferring molecular structures from spectroscopy data (+10.2 percentage points) and protein-variant function prediction (+7.7 percentage points).

Working with the model

In Anthropic's evaluations and early-access testing, Opus 5 stood out for verifying its own work and iterating until it succeeds:

  • On one Frontier-Bench task, given a drawing of a machine part and no way to view it directly, Opus 5 wrote its own computer-vision pipeline to pull the geometry from the raw pixels, then reconstructed the full part as a 3D model — repeatedly; no competing model with the same setup solved it in five attempts.
  • Given a real bug in a popular open-source package manager, Opus 5 found the root cause and fixed an edge case the community's patch had missed.
  • Anthropic's early-access customers reported the largest gains on harder, longer-horizon work — debugging and root-cause analysis, multi-day enterprise projects, financial modeling, legal agent work, and deck/document production — with several reporting equal-or-better quality than Opus 4.8 at meaningfully fewer tokens, turns, and wall-clock time.

Safety and evaluations

Anthropic's pre-deployment evaluation of Opus 5 is documented in the Claude Opus 5 System Card. Anthropic reports testing across software engineering, mathematical and scientific reasoning, long context, agentic search and multi-agent orchestration, multimodal and computer-use tasks, real-world professional work, and multilingual, healthcare, and life-sciences domains, with Opus 5 substantially stronger than Opus 4.8 across the board.

On cybersecurity specifically, Anthropic describes Opus 5 as a general-purpose model not specifically trained for cyber tasks, and reports five capability evaluations — ExploitBench, OSS-Fuzz, Firefox 147, CyScenarioBench, and ExploitGym — alongside external cyber-range testing by the UK AI Security Institute. Anthropic states that the model's cyber capabilities exceed those of its predecessor but fall short of Claude Mythos 5.