Claude Opus 5: The Complete Guide — Near-Fable 5 Intelligence at Half the Price

Update (September 27, 2026): Claude Opus 5.5 was released on September 22, 2026 and is now the current Opus model (API claude-opus-5-5, $4 per million input tokens and $20 per million output tokens, down from Opus 5's $5/$25). Anthropic says it performs at the level of Fable 5.1 on most work and costs 40% less to run than Opus 5 on typical workloads. Opus 5 remains available but is marked legacy in the official model docs. Opus 5 prices and comparisons below reflect the pre-launch state and will be updated. Details: Claude Opus 5.5 complete guide.

Anthropic officially released Claude Opus 5 on July 24, 2026 (local time). It is the fourth model in the Claude 5 family, and the company's headline claim is simple: intelligence that comes close to the frontier-tier Fable 5, at half the price. Pricing is unchanged from its predecessor Opus 4.8, positioning it as the "everyday workhorse" model.

Key specs at a glance

Based on the official announcement and model documentation (verified 2026-07-25):

How much better is it?

Summarizing the benchmarks Anthropic published (note these are the company's own figures, including internal runs):

• Coding: On Frontier-Bench v0.1 it surpasses all other models and more than doubles Opus 4.8's score at a lower cost per task. On CursorBench 3.2 at max effort it lands within 0.5% of Fable 5's peak score at half the cost per task.
• Problem solving: On ARC-AGI 3, which tests novel problems, its score is three times that of the next-best model.
• Work automation: On Zapier AutomationBench its pass rate is around 1.5× the next-best model at the same cost per task; even at its lowest effort setting it passes more tasks than any other model.
• Computer use: On OSWorld 2.0 it outperforms every other model at any given cost, surpassing Fable 5's best result at just over a third of the cost.
• Scientific research: It beats Opus 4.8 on every life-sciences evaluation, most notably +10.2 points on inferring molecular structures from spectroscopy data and +7.7 points on predicting how protein sequence variations affect function (internal benchmarks).

The effort setting — a dial for how deeply the model thinks — lets you trade intelligence against token cost; it defaults to high on the Claude API and Claude Code.

What changes for subscribers

Two practical changes: Claude Max now defaults to Opus 5, and on Claude Pro it becomes the strongest model you can select. It runs within normal plan limits, with no surcharge. For the separate subscription terms of the top-tier Fable 5 and 5.1, see the notice banner above. For Fable 5.1 (released September 1, 2026), see the Fable 5.1 guide.

Safeguards: far fewer interventions than Fable 5

Opus 5 is designed not to advance the frontier in risky dual-use capabilities. Its cyber classifiers are proportionally less restrictive than Fable 5's: finding vulnerabilities in source code is allowed, while binary-based vulnerability scanning, penetration testing, and exploit generation are blocked. Anthropic expects the classifiers to intervene around 85% less often than on Fable 5 (an estimate). In Claude.ai, Claude Code, and Claude Cowork, flagged requests fall back to Opus 4.8 by default, and fallbacks can also be enabled on the API. For biology, requests blocked on Fable 5 now route to Opus 5 rather than Opus 4.8. Separately, Opus 5 itself does not fall back on biology, chemistry, or life-sciences questions; it applies safeguards similar to Opus 4.8’s. What does switch from Opus 5 to Opus 4.8 is a narrow set of offensive cyber requests (exploit generation, binary-based vulnerability scanning, penetration testing), and automatic switching can be turned off under Settings > Capabilities with “Switch models when a message is flagged” (the conversation pauses instead). Source: official help article (checked October 31, 2026; it covers Opus 5 only).

Two changes developers should not miss

Porting Opus 4.8 code as-is will change behaviour, or fail outright, in two places (verified against the official model documentation on 2026-07-26).

1. Thinking is on by default. On Opus 4.8 a request only reasoned if you set thinking: {"type": "adaptive"}; on Opus 5 that same request runs with thinking enabled. The model decides how much to think each turn, and effort controls the depth. The old wire value stays valid and behaves identically to the default. Note that max_tokens caps thinking plus response text combined, so revisit it for workloads that ran without thinking on 4.8.
2. Disabling thinking now has a condition — this is a breaking change. thinking: {"type": "disabled"} is accepted only when effort is high or below. Pair it with xhigh or max and the API returns a 400 error. Either keep thinking off and drop effort to high or below, or keep the effort level and remove the thinking field.

One caveat: with thinking disabled, Opus 5 occasionally writes a tool call into its text output instead of emitting a tool_use block, or leaks internal XML tags into the visible response. Where you can, leave thinking on and manage cost with a lower effort level instead.

The five effort levels

Effort sets how deeply the model reasons. Opus 5 supports the full ladder — low, medium, high, xhigh and max — with high as the default and max as the top tier for the deepest reasoning.

Anthropic states that Opus 5 converts additional effort into better results more reliably than any earlier Opus model, which makes the level you pick carry more weight. The recommended approach is to start at the default and adjust in both directions based on your own evals: step down where quality holds to save tokens and latency, step up only for the most demanding work. When running at xhigh or max, set a generous max_tokens so the model has room to think and act across subagents and tool calls.

Efficiency at the lower levels is listed as an improvement in its own right: low and medium produce strong quality at a fraction of the tokens and latency of the higher settings.

Three features launched alongside

• Mid-conversation tool changes (beta): add or remove tools between turns while preserving the prompt cache, instead of resending a fixed tool list for the life of a session. Include the mid-conversation-tool-changes-2026-07-01 beta header.
• Default fallbacks mode (beta): the fallbacks parameter gained a "default" mode that applies Anthropic's recommended fallback models by refusal category, rather than a list you maintain yourself. Use the server-side-fallback-2026-07-01 beta header, which supports both the default mode and explicit lists (the earlier server-side-fallback-2026-06-01 header accepts lists only).
• Lower prompt cache minimum: the minimum cacheable prompt length drops to 512 tokens from 1,024 on Opus 4.8, so prompts that were previously too short to cache now create cache entries with no code changes.

If you are on Opus 4.8

Since pricing is identical, API users can upgrade at zero switching cost by changing the model ID to claude-opus-5; Anthropic provides an official migration guide. Note that the legacy Opus 4.1 (claude-opus-4-1-20250805) was retired on August 5, 2026, so if you were still using it, migrate immediately. Like prior Opus models, Opus 5 has no data retention requirements for general access (a difference from Fable 5).

The prompts matter more than the code. Opus 5 verifies its own work without being asked, so verification instructions carried over from earlier models — "include a final verification step", "use a subagent to verify" — are worth removing; leaving them in causes over-verification. Beyond that, default responses and written deliverables run longer, the model narrates progress more often in agentic sessions, and it delegates to subagents more readily in multi-agent frameworks. Vision also improved, so any prompt-side workarounds you tuned for earlier models are worth re-validating.

Where you can use it

• Claude API — available to all customers as claude-opus-5
• AWS — through Claude in Amazon Bedrock as anthropic.claude-opus-5
• Google Cloud — as claude-opus-5
• Microsoft Foundry — through Claude in Microsoft Foundry

One exception: fast mode is currently on the Claude API only (research preview) and is not available on Amazon Bedrock, Google Cloud or Microsoft Foundry. It is priced at $10 per million input tokens and $50 per million output tokens — twice the standard rate. Opus 4.8 remains available on all of these platforms.

Bottom line

Opus 5 is aimed at "more, for less." It matters most to users who need frontier-class performance but found Fable 5's cost and restrictions burdensome — developers running long coding-agent sessions, and teams automating document and analysis work. For the latest terms and full specs, see Anthropic's official announcement and model documentation.

Related articles

• Claude Opus 4.8: The Complete Guide to What's New and How to Use It
• Claude Fable 5 & Mythos 5: The Complete Guide
• Claude Model Selection Guide: Opus 5.5 vs Sonnet 5.5 vs Haiku 5.5
• Claude Sonnet 5 Complete Guide: What's New, Pricing, and Migration Gotchas
• Claude vs ChatGPT vs Gemini: Why the Right Pick Depends on the Task