Key Takeaways
- Claude Opus 5 is Anthropic’s flagship production AI model, released July 24, 2026, at the same API pricing as Claude Opus 4.8: $5 per million input tokens and $25 per million output tokens. It delivers near-Fable-5 performance at approximately half the cost of Fable 5 ($10/$50 per million tokens), making it the strongest value option among frontier-tier models for coding agents, enterprise workflows, and long-context analysis.
- On SWE-bench Verified, the standard benchmark for real-world software engineering task performance, Claude Opus 5 scores 96.0%, up from Opus 4.8’s 88.6%. On SWE-bench Pro, it scores 79.2%, up from Opus 4.8’s 69.2%, putting it within 1.1 percentage points of Claude Fable 5’s 80.3% score on the same benchmark. On Frontier-Bench v0.1 (agentic terminal coding), Opus 5 scores 43.3%, taking the top position among all models at the time of release.
- ARC-AGI-3 is the benchmark most cited in the Claude Opus 5 launch as evidence of a generational capability jump. Opus 5 scores 30.2%, which is approximately three times the score of the next-best model on that benchmark. ARC-AGI-3 tests novel reasoning on tasks requiring genuine abstraction rather than pattern recall, making the result significant for evaluating whether Opus 5 can generalize to genuinely new problems.
- The context window is 1 million tokens with a 128,000-token maximum output length. Opus 5 scores 100% on the Needle in a Haystack benchmark, meaning it can retrieve a specific fact placed anywhere within a 1-million-token document without degradation. The 1M context was available on Opus 4.8 as an opt-in setting; on Opus 5 it is the default configuration.
- Adaptive thinking is enabled by default on Claude Opus 5. The model allocates additional compute to difficult problems automatically, applying extended reasoning where needed without requiring the user to invoke a separate reasoning mode.
- Claude Opus 5 is available on all Anthropic platforms as of July 24, 2026: Claude Max subscribers receive it as the default model, Claude Pro users can select it as the most capable option, and it is available via API using the model string `claude-opus-5`. It is also available on AWS Bedrock.
- The primary limitation of Claude Opus 5 relative to Claude Fable 5 is performance on the hardest coding and reasoning tasks: Fable 5 leads on Frontier Code (Diamond difficulty) with 29.3% against Opus 5’s lower score, and on the Epoch Capabilities Index composite ranking, Opus 5 sits below both Fable 5 and GPT-5.6. For teams where maximum capability on the hardest tasks matters more than cost, Fable 5 remains the stronger choice.
- For production AI agent deployments, Claude Opus 5 is the default recommendation over Fable 5 for most teams: the performance gap is narrow (1.1 percentage points on SWE-bench Pro), the cost difference is substantial (half the per-token price), and Opus 5 is purpose-built for long-running autonomous workflows with self-verification and adaptive judgment built into the default configuration.
Every major AI laboratory releases multiple model tiers, and frontier models routinely top each other on narrower benchmarks. What makes Claude Opus 5’s release notable is not that it performs well, that was expected, but the specific configuration of its performance: near-maximum capability on real-world coding and agentic tasks at the same price point as the previous-generation Opus 4.8. The value calculation for production deployments changed when Opus 5 launched.
This review covers what Opus 5 actually scores, what those scores mean practically, where it leads, where Fable 5 still wins, and which workflows it is the right tool for.
Claude Opus 5: Release, Availability, and Pricing
Claude Opus 5 was released by Anthropic on July 24, 2026, across all platforms simultaneously: Claude.ai, the Claude API, and AWS Bedrock. For Claude Max subscribers, it became the new default model. For Claude Pro users: it is available as a selectable option. API access uses the model string `claude-opus-5`.
Pricing is $5 per million input tokens and $25 per million output tokens. This matches Claude Opus 4.8’s price exactly, meaning production teams running on Opus 4.8 can upgrade to Opus 5 at zero additional cost. Compared to Claude Fable 5 at $10/$50 per million tokens, Opus 5 costs approximately half as much per token on both sides. At scale, this cost difference compounds significantly: a deployment consuming 100 million tokens monthly saves $500,000 annually on input alone by choosing Opus 5 over Fable 5.
The context window is 1 million tokens with a 128,000-token maximum output. The 1M context was available on Opus 4.8 as an extended context opt-in; on Opus 5 it is the default configuration applied to every request.
Benchmark Performance
SWE-bench Verified and SWE-bench Pro
SWE-bench is the most widely used benchmark for evaluating real-world software engineering performance. Tasks involve resolving actual GitHub issues from production codebases, requiring the model to understand a repository, identify the correct file and function, produce a working patch, and avoid breaking existing tests. The Verified variant uses a curated set of resolvable issues; the Pro variant uses harder tasks that filter out issues solvable by pattern-matching alone.
Claude Opus 5 scores 96.0% on SWE-bench Verified and 79.2% on SWE-bench Pro. The Verified score is the higher of the two numbers and reflects performance across the broader task set. The Pro score is the more meaningful signal: at 79.2%, Opus 5 is 10 percentage points ahead of GPT-5.5 (58.6%) and within 1.1 percentage points of Claude Fable 5 (80.3%).
The improvement from Opus 4.8 is substantial. Opus 4.8 scored 88.6% on SWE-bench Verified and 69.2% on SWE-bench Pro. Opus 5 improves by 7.4 and 10 percentage points respectively, representing a meaningful capability jump rather than an incremental update.
Frontier-Bench v0.1
Frontier-Bench v0.1 evaluates agentic terminal coding: tasks where a model must operate autonomously in a command-line environment, navigate a codebase, use tools, run tests, interpret results, and complete complex engineering objectives without human assistance. Claude Opus 5 scores 43.3% on this benchmark, placing it at the top of the published leaderboard at the time of its release and more than doubling Opus 4.8’s score on the same evaluation.
For teams building AI coding agents that operate in CI/CD pipelines, development terminals, or autonomous engineering workflows, Frontier-Bench is the most relevant benchmark. The 43.3% score reflects Opus 5’s design orientation: it is purpose-built for extended agentic tasks that require multi-step reasoning, error recovery, and judgment about when to proceed versus verify.
ARC-AGI-3
ARC-AGI-3 (Abstraction and Reasoning Corpus, third generation) tests novel reasoning on tasks that cannot be solved through pattern recall or memorized solutions. Each task presents a new visual or structural puzzle that requires genuine abstraction, and the benchmark is designed to be resistant to models that have high performance on standard reasoning evaluations without demonstrating true generalization.
Claude Opus 5 scores 30.2% on ARC-AGI-3, approximately three times the score of the next-best model on the benchmark at the time of release. Anthropic highlighted this result in the launch announcement as evidence that Opus 5 represents a qualitative capability advancement rather than a performance optimization. The result suggests the model can generalize to genuinely novel problem types rather than only improving on tasks within the distribution of its training data.
Needle in a Haystack (1M Context)
Needle in a Haystack is a retrieval evaluation where a specific fact is placed at varying positions within a very long document, and the model is asked to locate and return it. A model that scores 100% retrieves the fact correctly regardless of where it is placed in the document, no degradation at the beginning, middle, or end of the 1M-token context.
Claude Opus 5 scores 100% on this evaluation, confirming that the 1M-token context is not merely nominal. For practical applications involving long document analysis, legal contracts, research corpora, large codebases, extensive transcripts, this result means Opus 5 can be trusted to retrieve any specific piece of information from anywhere within a million-token context window without performance degradation.
Key Capabilities
Agentic Workflows
Claude Opus 5 is built for long-running autonomous operation. Anthropic describes it as capable of running agents for hours or overnight, handling multi-step workflows, adapting strategies when encountering obstacles, and recovering from errors without requiring human intervention at each decision point. This positions it as the infrastructure model for AI agent deployments where the agent must complete a complex objective autonomously rather than responding to single-turn requests.
The practical implication for engineering teams is that Opus 5 is the appropriate base model for autonomous coding agents, research agents, and enterprise workflow automation where the task involves many sequential decisions rather than a single response. The self-verification capability, the model’s tendency to check its own outputs before completing a task, is a meaningful design feature for agent deployments where incorrect intermediate steps compound into larger failures downstream.
Adaptive Thinking
Adaptive thinking is enabled by default on Claude Opus 5. The model applies extended reasoning automatically on problems it identifies as requiring more compute, without the user needing to invoke a separate mode or flag. On simpler tasks, it responds directly; on complex reasoning, coding, or analytical tasks, it allocates additional thought before returning an answer. This makes the effective reasoning quality higher than the baseline performance on fixed-compute evaluations might suggest for real-world usage patterns.
Long-Context Analysis
The combination of 1M-token context, 128K-token output, and perfect Needle in a Haystack performance makes Claude Opus 5 the strongest model in its price tier for tasks involving very large documents. Legal document review, large codebase analysis, research synthesis across many sources, and processing of extensive enterprise data sets all become viable single-context operations. The 128K output capacity means that tasks requiring both ingestion of a large input and generation of a substantial output (a detailed analysis of a large contract, a comprehensive rewrite of a large codebase component) are feasible without output truncation.
Coding and Software Engineering
At 96.0% on SWE-bench Verified and 79.2% on SWE-bench Pro, Claude Opus 5 is the most capable model available for software engineering tasks in its price bracket. In real-world coding evaluations, the model demonstrates an ability to understand large repositories, identify root causes of bugs rather than superficial symptoms, propose minimal correct patches rather than over-engineered solutions, write focused regression tests for fixed issues, and provide verification steps that pass locally. These characteristics matter for production use: an AI coding assistant that over-engineers fixes or misidentifies root causes creates as many problems as it solves.
Claude Opus 5 vs Claude Fable 5
Claude Fable 5 is Anthropic’s maximum-capability model, positioned above Opus 5 in the model hierarchy and priced at $10/$50 per million input/output tokens. On most benchmarks, the gap between Opus 5 and Fable 5 is narrow.
On SWE-bench Pro, the difference is 1.1 percentage points (Fable 5 at 80.3%, Opus 5 at 79.2%). On Frontier-Bench agentic coding, Opus 5 takes the top score and Fable 5’s published score falls below it on that specific evaluation. The area where Fable 5 leads more clearly is Frontier Code Diamond difficulty, the hardest coding tasks in the evaluation distribution, where Fable 5 scores 29.3% against a lower Opus 5 score. For teams whose production workload concentrates on the most difficult possible engineering tasks at scale, Fable 5’s edge on Diamond-difficulty problems may justify the additional cost. For most production deployments, the 1.1 percentage point gap on SWE-bench Pro does not justify a 2x increase in per-token cost.
The practical recommendation: use Opus 5 as the default for production agents and complex workflows. Escalate to Fable 5 for tasks that specifically require Diamond-difficulty reasoning, or for cases where the highest possible performance matters more than cost efficiency.
Claude Opus 5 vs GPT-5.5 and GPT-5.6
On the benchmarks most relevant to software engineering, Claude Opus 5 leads GPT-5.5 clearly. On SWE-bench Pro, Opus 5 scores 79.2% versus GPT-5.5’s 58.6%, a 20.6 percentage point gap. On Frontier Code Diamond difficulty, GPT-5.5 scores 5.7% against Fable 5’s 29.3%, with Opus 5 performing substantially above GPT-5.5 on this task type. On Frontier-Bench agentic terminal coding, Opus 5 holds the top score.
GPT-5.5 and GPT-5.6 (the successor model) lead on certain evaluations including math benchmarks and long-context retrieval. The Epoch Capabilities Index composite ranking places Opus 5 behind both Fable 5 and GPT-5.6 in overall composite capability, this reflects that GPT-5.6 is a stronger general model on averaged capabilities, while Opus 5’s advantage concentrates in coding and agentic task performance. For teams whose primary use case is software engineering, autonomous agents, or enterprise coding workflows, Opus 5’s domain-specific advantage is more relevant than the composite index ranking.
Who Should Use Claude Opus 5
Claude Opus 5 is the right choice for engineering teams building AI coding agents, production AI pipelines, or autonomous workflow systems where the agent must complete multi-step tasks reliably without constant human oversight. The combination of SWE-bench Pro performance, Frontier-Bench agentic coding leadership, 1M-token context, and same-as-Opus-4.8 pricing makes it the natural default upgrade for teams currently running on Opus 4.8.
For enterprise teams with large-document analysis requirements, legal review, research synthesis, large-codebase operations, the 1M-token context with 100% retrieval fidelity and 128K output removes a constraint that previously required document splitting, chunking, or retrieval-augmented approaches. Opus 5 can process an entire contract, codebase, or research corpus in a single context.
Researchers and advanced users working on novel reasoning problems who want access to a model with meaningfully higher generalization capability (as measured by ARC-AGI-3) at a production-accessible price point have a clear option in Opus 5. The 3x ARC-AGI-3 improvement over the next-best model is the result in the launch materials most directly relevant to this use case.
Teams that primarily need a general-purpose chat or writing assistant and are not running high-volume agentic workflows or complex coding tasks will find the cost difference between Opus 5 and less expensive models like Claude Haiku 4.5 or Sonnet 5 harder to justify. Opus 5’s pricing reflects its position as a production reasoning and coding model, not a general-purpose conversational model. For conversational or content use cases, lower-tier models at lower price points offer better value.
Verdict
Claude Opus 5 is the most capable widely available model for software engineering and agentic tasks at its price point. The 96.0% SWE-bench Verified score, 79.2% SWE-bench Pro score, 43.3% Frontier-Bench agentic coding score, 30.2% ARC-AGI-3 score, and 100% long-context retrieval fidelity represent a meaningful capability upgrade from Opus 4.8 at the same price.
The narrow gap from Fable 5 on most benchmarks (1.1 points on SWE-bench Pro) combined with the 2x cost advantage makes Opus 5 the default recommendation for production agent deployments where cost-per-token matters at scale. It earns the headline: the model that topped the leaderboards most production teams care about in 2026, while pricing itself within reach of the teams that need to run it at volume.
Frequently Asked Questions
When was Claude Opus 5 released?
Claude Opus 5 was released by Anthropic on July 24, 2026. It launched simultaneously across Claude.ai, the Claude API, and AWS Bedrock. Claude Max subscribers received it as the new default model on the same day, and Claude Pro users can select it as the most powerful option available.
How much does Claude Opus 5 cost?
Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens via the API. This is the same price as Claude Opus 4.8, the previous generation model. Compared to Claude Fable 5 at $10/$50 per million tokens, Opus 5 costs approximately half as much per token on both input and output. There is no additional charge for the 1M-token context window or adaptive thinking.
What is Claude Opus 5’s context window?
Claude Opus 5 has a 1-million-token context window, now set as the default configuration rather than an opt-in extended context mode as it was on Opus 4.8. The maximum output length is 128,000 tokens. The model scores 100% on Needle in a Haystack retrieval across the full 1M-token range, meaning performance does not degrade at any position within the context window.
How does Claude Opus 5 perform on SWE-bench?
Claude Opus 5 scores 96.0% on SWE-bench Verified and 79.2% on SWE-bench Pro. These are improvements of 7.4 and 10 percentage points respectively over Claude Opus 4.8’s scores of 88.6% and 69.2%. On SWE-bench Pro, Opus 5 is within 1.1 percentage points of Claude Fable 5’s 80.3% score, and 20.6 percentage points ahead of GPT-5.5’s 58.6% score.
What is ARC-AGI-3 and why does Opus 5’s score matter?
ARC-AGI-3 is the third generation of the Abstraction and Reasoning Corpus benchmark, designed to test novel reasoning on tasks that cannot be solved through memorized patterns or training data recall. Claude Opus 5 scores 30.2% on ARC-AGI-3, approximately three times the score of the next-best model at the time of release. Anthropic cited this result as evidence that Opus 5 represents a genuine capability improvement in generalization rather than only improved performance on tasks within training distribution.
What is the API model string for Claude Opus 5?
The API model string for Claude Opus 5 is `claude-opus-5`. Developers currently using `claude-opus-4-8` or similar model strings can update their API calls to `claude-opus-5` to access the new model without changes to request format, pricing, or token limits.
Is Claude Opus 5 better than Claude Fable 5?
Claude Fable 5 remains Anthropic’s highest-capability model and leads Claude Opus 5 on the hardest coding tasks (Frontier Code Diamond difficulty) and on the Epoch Capabilities Index composite ranking. On SWE-bench Pro, Fable 5 leads by 1.1 percentage points (80.3% vs 79.2%). On Frontier-Bench agentic coding, Opus 5 holds the higher published score. For most production deployments, the capability gap between Opus 5 and Fable 5 is narrow enough that Opus 5’s 2x cost advantage makes it the better choice. Teams running workloads where absolute maximum capability on the hardest possible tasks is more important than cost efficiency should use Fable 5.
What is adaptive thinking in Claude Opus 5?
Adaptive thinking is a default feature of Claude Opus 5 that allows the model to allocate additional compute to problems it identifies as requiring extended reasoning, without the user needing to select a separate reasoning mode. On simple tasks, the model responds directly. On complex coding, analysis, or reasoning tasks, it applies more extended internal thinking before producing its response. This improves output quality on difficult tasks without requiring configuration changes or explicit prompting from the user.
Where can I access Claude Opus 5?
Claude Opus 5 is available via Claude.ai (as the default model on Max plans, selectable on Pro plans), the Claude API using the model string `claude-opus-5`, and AWS Bedrock. It launched on all three platforms simultaneously on July 24, 2026. There is no waitlist or limited access period; it is generally available across all supported platforms.
