Anthropic
Claude Opus 5 (Adaptive Reasoning, Max Effort)
Released 
July 2026

Anthropic
Claude Opus 5 (Adaptive Reasoning, Max Effort)

Anthropic's near-frontier reasoning model with a per-request effort dial for agentic coding at controllable cost.

Modality:
Text
Image
PDF
model ID
anthropic/claude-opus-5

Output Speed *

55.40
tok/s

Intelligence Index *

50.7
/ 100

Context Window *

1000000
tokens

Input price

30
Anytoken

Output price

150
Anytoken
Claude Opus 5: Near-Frontier Reasoning With a Cost-Control Effort Dial Claude Opus 5 is Anthropic's flagship Opus-tier model, released July 24, 2026. It sits below Claude Fable 5 in maximum capability but delivers much of that intelligence at roughly half the per-task cost, keeping the same base price as Opus 4.8. Its defining feature is a per-request effort dial (low/medium/high/xhigh/max) that lets teams trade reasoning depth for token savings. With adaptive thinking on by default, a 1M-token context window, and strong agentic-coding scores, Opus 5 fits demanding autonomous coding, debugging, and enterprise knowledge work where accuracy matters and budget must be managed. Start building with Claude Opus 5 via the AnyAPI.ai API

Performance

Where Opus 5 Wins: Agentic Coding Per Dollar

Opus 5 is built for autonomous software engineering and knowledge work. On Frontier-Bench v0.1 it more than doubles Opus 4.8's score and surpasses all other models at a lower cost per task, and on CursorBench 3.2 at max effort it lands within 0.5% of Fable 5's peak at roughly half the cost. The reason this matters is the effort dial: teams route hard problems to higher effort and routine work to lower effort, preserving accuracy while trimming token spend. In production, that makes Opus 5 a practical default for coding agents that must balance quality against per-task cost.

Benchmarks

Independent Benchmarks: Terminal-Bench, ARC-AGI 3 and Intelligence Index

On the Artificial Analysis Intelligence Index (v4.1), Claude Opus 5 (max effort) scores 51 and ranked #1 at launch across a field of 586 evaluated models. On Terminal-Bench 2.1, Artificial Analysis reports 89.1% at max effort—a 14.5-point jump over Opus 4.8 and just 0.4 points behind GPT-5.6 Sol. On ARC-AGI 3, a test of novel problem-solving, Opus 5 scored roughly 30%, about three times the next-best model, indicating genuine ability to work out unfamiliar rule systems rather than recall. Note that several coding figures are effort-sensitive and vendor-configured.

Output Speed

*
55.40
tok/s

Intelligence Index

*
50.7
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
93
%

HLE *

Adherence to multi-step structured instructions.
55
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

Claude Opus 5 accepts text and images (including PDFs/files) and returns text only—there is no image, audio, or video output. It carries a 1M-token context window as both default and maximum, with 128K max output tokens (up to 300K via the Batch API beta). Adaptive thinking is on by default with high as the default effort. The two most production-relevant details: thinking tokens count toward the same output limit, so review max_tokens when migrating; and effort is not strictly monotonic in quality—max can underperform xhigh on some tasks. Model ID: claude-opus-5.
Verified Specifications — 
Claude Opus 5 (Adaptive Reasoning, Max Effort)
*
Input modalities
Text
Image
PDF
output modalities
Text
Context window
1000000
 tokens
Maximum output tokens
128000
Reasoning
Yes
Knowledge cutoff
July 2026
Pricing (standard)
30
 AnyTokens in
 / 
150
 AnyTokens out

Comparison

Claude Opus 5 vs Claude Fable 5: Capability Ceiling or Cost Control?

Opus 5 and Fable 5 are natural alternatives: both are Anthropic frontier-tier models with a 1M-token context window, adaptive thinking, and image/file input, aimed at autonomous coding and knowledge work. Anthropic positions Opus 5 as coming close to Fable 5's intelligence at roughly half the per-task cost, and Fable 5's safety classifiers actually fall back to Opus 5 when a request is blocked. The practical decision is whether your workload needs Fable 5's absolute capability ceiling for long-running autonomous tasks, or whether Opus 5's near-frontier quality at lower cost and burn rate is enough.

Dimension
Claude Opus 5 (Adaptive Reasoning, Max Effort)
Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback)
Context window *
1000000
tokens
1000000
tokens
Output speed *
55.40
tok/s
68.64
tok/s
Intelligence Index *
50.7
49.7
Input pricing
30
AnyToken
60
AnyToken
Output pricing
150
AnyToken
300
AnyToken
Knowledge cutoff *
July 2026
June 2026

Choose Claude Opus 5 when you want near-frontier reasoning and agentic-coding quality with materially lower per-task cost, a later May 2026 knowledge cutoff, fewer spurious refusals, and fine-grained control over spend via the effort dial—ideal for regular advanced coding and enterprise workflows. Choose Claude Fable 5 when you need the absolute capability ceiling for long-running, ambiguous, multi-day autonomous tasks and can absorb roughly double the cost and its higher token burn rate. For most teams, Opus 5 is the more economical default.

Limitations & Trade-offs

Where Claude Opus 5 (Adaptive Reasoning, Max Effort) falls short

1
Below-average output speed. Independent Artificial Analysis measurements put Opus 5 around 50–54 tokens per second across effort levels—below the median for reasoning models in its price tier (roughly 63–68 t/s). Combined with adaptive thinking and time-to-first-answer-token latency that can run tens of seconds at max effort, this makes Opus 5 a poor fit for strict low-latency, real-time chat or interactive UX. For latency-sensitive, high-volume workloads, a lighter model such as Sonnet 5 is preferable.
2
Premium reasoning tier, not a high-volume default. Opus 5 occupies the higher-cost end of the Claude lineup, and outputs are comparatively expensive—reasoning tokens count toward the same output limit, so extended thinking directly inflates cost. Filling the 1M-token context on every call in a loop is unwise. For cost-sensitive, simple, high-throughput tasks, a cheaper sibling delivers better economics; reserve Opus 5 for work where the extra accuracy justifies the spend.
3
Effort is not monotonic in quality. Raising the effort dial does not always improve results: on Frontier-Bench, Opus 5's xhigh setting (44.4) outperforms its more expensive max setting (43.3). A longer reasoning trace can talk the model out of a correct early answer. This means teams cannot assume max effort is optimal and must benchmark effort levels per workload, adding evaluation overhead before routing production traffic.
4
Text-only output and vendor-configured benchmarks. Opus 5 accepts text, images, and PDFs but returns text only—no image, audio, or video generation. Additionally, several headline coding scores (Frontier-Bench, CursorBench) are Anthropic-run and include a safety-fallback policy that routed a fraction of trials to Opus 4.8, so they partly evaluate a configured agent system rather than an untouched endpoint. Validate on your own workload rather than relying solely on vendor figures.

Best-Fit Workloads

Where this model earns its place

01

Autonomous agentic coding

Opus 5 is Anthropic's recommended starting model for complex agentic coding. It more than doubles Opus 4.8 on Frontier-Bench and reaches 89.1% on Terminal-Bench 2.1 (Artificial Analysis, max effort), with strong tool calling and beta mid-conversation tool changes that preserve the prompt cache. This suits multi-file repository agents in tools like Cursor or Devin. Tune effort per task to control token spend, and benchmark effort levels since max is not always optimal.

02

Difficult debugging and root-cause analysis

Anthropic and independent write-ups highlight Opus 5's strength on hard debugging and root-cause tasks, where it self-verifies and recovers from errors without intervention. Its adaptive thinking and later May 2026 knowledge cutoff help with recent libraries and APIs. This fits production incident triage and fixing subtle regressions across large codebases. For simple, mechanical fixes, a cheaper model is more economical.

03

Whole-codebase and long-document analysis


The 1M-token context window—default and maximum, with no input surcharge for long contexts—lets Opus 5 reason over an entire mid-sized codebase or large document corpus in a single request. Prompt caching reduces the cost of repeated stable context. This suits architecture review, cross-file refactor planning, and large-corpus knowledge work. A 1M window is capacity, not an instruction to fill it; irrelevant context still degrades accuracy, so structure prompts and consider RAG for very large retrieval needs.

04

Enterprise knowledge work

Opus 5 leads on knowledge-work evaluations like GDPval-AA and is positioned as a daily driver for serious business tasks, with the most aligned Opus behavior and roughly 85% fewer spurious refusals than prior models. This fits research synthesis, analysis, and document-heavy enterprise workflows where reliability and fewer refusals matter. The effort dial lets teams keep routine queries cheap while reserving deeper reasoning for high-stakes tasks.

Pricing in anytokens via AnyAPI
Input
30
Output
150
Cache write
37.5
Cache read
3

Integration

Access Claude Opus 5 (Adaptive Reasoning, Max Effort) via AnyAPI.ai

Access Claude Opus 5 (Adaptive Reasoning, Max Effort) through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate Claude Opus 5 (Adaptive Reasoning, Max Effort) without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access Claude Opus 5 (Adaptive Reasoning, Max Effort) and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test Claude Opus 5 (Adaptive Reasoning, Max Effort) against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use Claude Opus 5 (Adaptive Reasoning, Max Effort) from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use Claude Opus 5 (Adaptive Reasoning, Max Effort) for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

Claude Opus 5 has a 1M-token context window, which is both the default and the maximum—there is no smaller standard variant or beta header needed to unlock it. Maximum output is 128K tokens, extendable to 300K via the Message Batches API beta. There is no input surcharge for long contexts.

Opus 5 exposes a per-request effort control (low, medium, high, xhigh, max, default high) that decides how much reasoning compute the model spends. Lower effort preserves much of the quality while using fewer tokens and costing less. Effort is not strictly monotonic—on some benchmarks xhigh beats the more expensive max—so benchmark per workload.

Not universally. Anthropic positions Opus 5 as approaching Fable 5's intelligence at roughly half the per-task cost, and on some coding benchmarks Opus 5 even leads. Fable 5 remains the choice for the absolute capability ceiling on long-running autonomous tasks. For most advanced work, Opus 5 offers better cost-efficiency.

The Claude Opus 5 API accepts text, images, and documents such as PDFs as input, and returns text only. It does not generate images, audio, or video. It supports function/tool calling, JSON output via response_format (without strict schema enforcement), streaming, prompt caching, and batch processing.

Independent Artificial Analysis measurements report roughly 50–54 output tokens per second across effort levels on Anthropic's API—below the median for reasoning models in its price tier. Time to first answer token can reach tens of seconds at max effort. Opus 5 prioritizes reasoning quality over raw speed, so it is not ideal for strict low-latency use.

* Benchmark data source: Artificial Analysis artificialanalysis.ai