Claude Opus 5

Overview

Claude Opus 5 is Anthropic's Opus-tier flagship, released on July 24, 2026 and available on every platform the same day. 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 its documentation now names Opus 5 the recommended starting point for complex agentic coding and enterprise work — the role Opus 4.8 held from May.

The pitch is a price-performance move rather than a new capability ceiling. Fable 5 remains Anthropic's most capable widely released model and still sits above the Opus tier for workloads that need the highest available capability. What Opus 5 changes is what you get at $5/$25: Anthropic reports it approaches Fable-level results on several agentic evaluations while using 26% fewer tokens on average than Opus 4.8, which on tool-heavy workloads moves the bill more than the unchanged sticker price suggests.

For developers, migration is close to a model-ID swap — claude-opus-5, a fixed ID with no date suffix — with two exceptions worth auditing before deploying: adaptive thinking is now on by default, and disabling it is capped at high effort.

Capabilities

• Long-horizon agentic execution: Anthropic's headline claim is Frontier-Bench v0.1, where Opus 5 "more than doubles Opus 4.8's performance at a lower cost per task." Customer statements in the launch post consistently describe gains concentrated on longer-horizon work rather than single-turn tasks.
• Self-verification without prompting: Opus 5 checks its own work unprompted. Anthropic's migration guidance is to remove verification instructions carried over from earlier prompts — leaving them in causes over-verification.
• Token efficiency: 26% fewer tokens on average than Opus 4.8 for the same work, per Anthropic. Goldman Sachs, quoted in the launch post, reports its trading benchmark completing with "a seventh of the reasoning tokens" of Opus 4.8.
• Computer use: On OSWorld 2.0, Anthropic says Opus 5 surpasses Fable 5's best result "at just over a third of the cost."
• Abstract reasoning: On ARC-AGI 3, Anthropic reports a score "three times as high as the next-best model" — the largest relative claim in the announcement, and one made without a published absolute figure.
• Scientific research: 10.2 percentage points higher than Opus 4.8 on organic chemistry tasks and 7.7 points higher on protein tasks.
• Willingness to disagree: Several testers describe Opus 5 pushing back on proposed designs rather than complying. Anthropic frames this as judgment; in practice it means prompts that assume compliance may get an argument instead.

Technical Specifications

• API model ID: claude-opus-5 (Amazon Bedrock: anthropic.claude-opus-5; Google Cloud: claude-opus-5)
• Context window: 1M tokens, at standard API pricing (no long-context premium)
• Max output: 128K tokens per request. On the Message Batches API, up to 300K output tokens via the output-300k-2026-03-24 beta header.
• Thinking: Adaptive thinking, on when the thinking parameter is omitted. Extended thinking (thinking.type: "enabled") is not supported.
• Effort levels: low, medium, high (default on the Claude API and Claude Code), xhigh, max
• Training data cutoff: May 2026 (reliable knowledge cutoff: May 2026)
• Minimum cacheable prompt: 512 tokens — half the 1,024 required on Opus 4.8
• Comparative latency: Moderate — slower than Sonnet 5 and Haiku 4.5, faster than Fable 5
• Architecture: Transformer-based, developed under Anthropic's Constitutional AI approach.

Supported features

Effort, adaptive thinking, task budgets (beta), compaction (beta), context editing (beta), the memory tool, code execution, programmatic tool calling, tool search, high-resolution vision, prompt caching, Message Batches, structured outputs, mid-conversation tool changes (beta) and automatic fallbacks (beta).

Not supported: Priority Tier, which Opus 4.8 kept — organizations with a Priority Tier commitment must plan capacity separately. The server-side web fetch tool is also unavailable on Opus 5.

API Behavior

Opus 5 inherits the Opus 4.7/4.8 request surface. Four things still return a 400 error if you are arriving from Opus 4.6 or earlier:

• thinking: {"type": "enabled", "budget_tokens": N} — removed. Use adaptive thinking and control depth with output_config.effort.
• temperature, top_p, top_k — removed. Steer behavior with prompting.
• Last-assistant-turn prefills — removed. Use structured outputs (output_config.format) or a system-prompt instruction.
• thinking.display — defaults to "omitted", so thinking blocks arrive with an empty thinking field. Pass "summarized" if you surface reasoning to users.

The two changes from Opus 4.8

Thinking is on by default. A request with no thinking field ran without thinking on Opus 4.8; on Opus 5 it runs with adaptive thinking. max_tokens remains a hard cap on total output — thinking plus visible text — so any workload that relied on the old default needs its max_tokens re-checked or thinking: {"type": "disabled"} set explicitly.

Disabling thinking is capped at high effort. thinking: {"type": "disabled"} with xhigh or max effort is rejected with a 400, enforced per request. Anthropic also warns that with thinking off, the model can occasionally emit tool calls as plain text or include internal XML tags in visible output — preferring lower effort with thinking is the recommended path.

New betas

• Mid-conversation tool changes (mid-conversation-tool-changes-2026-07-01) — add or remove tools between turns without invalidating prompt-cache hits on earlier turns. Previously any change to the tool list invalidated the entire cached prefix.
• Automatic fallbacks (server-side-fallback-2026-07-01) — pass fallbacks: "default" and the API selects a recommended fallback model based on the refusal category, instead of requiring you to name one. Opus 5's cyber-category refusals fall back to Opus 4.8.

Behavioral Notes

Anthropic's migration guidance for Opus 5 is mostly about what to take out of prompts tuned for Opus 4.8:

• Delete self-verification scaffolding. Opus 5 verifies its own work; instructions telling it to do so cause over-verification.
• Ask for length explicitly. Default visible responses and written deliverables run longer than on Opus 4.8, and lowering effort reduces thinking volume without reliably shortening the visible answer. Prompt for conciseness or a target length instead of reaching for the effort dial.
• Constrain scope on narrow tasks. Left unbounded, Opus 5 does more than asked.
• Steer or cap subagents. In multi-agent frameworks, state which scenarios warrant delegation, or set a limit.
• Re-run your effort sweep. Anthropic reports low and medium are stronger on Opus 5 than on earlier Opus models and worth testing as cost and latency controls, while max is worth testing where capability matters more than token spend. If you run at xhigh or max, start max_tokens at 64K and tune from there.

Use Cases

• Autonomous software development: Overnight refactors, large migrations, and multi-file feature work from a single well-specified brief — the workload Anthropic's Frontier-Bench and CursorBench claims target.
• Enterprise knowledge work: Financial modeling, due diligence and research synthesis. Morgan Stanley, Databricks and Goldman Sachs all describe financial workloads as the area where they measured the largest gain over Opus 4.8.
• Agentic automation: Zapier reports Opus 5 topping its AutomationBench leaderboard "without spending more tokens than prior Claude models" — the combination that matters when a workflow runs thousands of times a day.
• Computer use: Browser and desktop automation, where Anthropic's OSWorld 2.0 cost claim is the strongest published argument for choosing Opus 5 over Fable 5.
• Scientific and technical analysis: Chemistry and protein work, where Anthropic publishes concrete percentage-point gains over Opus 4.8.
• Code review and debugging: Note the standing Opus caveat — instructions like "only report high-severity issues" are followed literally and can depress measured recall. Report everything and filter downstream.

Performance / Benchmarks

Anthropic's launch post publishes comparisons, not a score table. Every capability claim in the announcement is stated relative to another model or to cost, with no absolute per-benchmark figures in the text. The Claude Opus 5 System Card, published July 24, 2026, carries the full capability and safety evaluation.

What Anthropic states in the announcement:

Figures that come from customers, not Anthropic

The launch post's most quotable percentages sit inside partner statements describing internal evaluations. They are not Anthropic measurements, the evaluations are not public, and none of them can be reproduced:

• Box: Opus 5 "outperforms Opus 4.8 by 8%."
• Goldman Sachs: strongest Opus model on its trading benchmark, using "a seventh of the reasoning tokens" of Opus 4.8.
• Perplexity: on first-turn redlines, "nearly double Opus 4.8."
• Devin (Cognition): on FrontierCode 1.1, "approaches Fable-level performance at half the cost."

Other figures circulating from the post — an 11% improvement in data analysis workflows, 17% in due diligence, 22% over Opus 4.7 on agentic coding, 60% less time on financial modeling — carry the same caveat.

How to read these numbers

Every figure above is vendor-reported or customer-reported. Anthropic chose the benchmarks, the harnesses and the effort levels; partners chose their own. Relative claims of the form "n× the next-best model" are the weakest kind of published result, because neither the absolute score nor the comparison set is stated. Treat all of it as an upper bound and validate on your own evaluation set — particularly the 26% token reduction, which is the claim most likely to change a budget and the easiest to measure yourself.

Limitations

• Not the capability ceiling: Fable 5 remains Anthropic's most capable widely released model, and Opus 5 stays behind Claude Mythos 5 on biology research and offensive cybersecurity by Anthropic's own statement.
• No Priority Tier: Unsupported on Opus 5 while Opus 4.8 keeps it — a real migration blocker for organizations with a throughput commitment.
• No web fetch: The server-side web fetch tool is unavailable; web search remains.
• Longer default output: Visible responses and deliverables run longer than on Opus 4.8, and effort is not a reliable lever for shortening them — you have to prompt for length.
• Thinking-off is constrained: Disabling thinking is capped at high effort, and Anthropic acknowledges degraded output formatting in that mode.
• Cost: $5/$25 per million tokens sits well above the Sonnet and Haiku tiers; fast mode doubles it.
• Reduced parameter surface: No temperature/top_p/top_k, no budget_tokens, no assistant prefill.
• Cybersecurity refusals: Source-code vulnerability finding is permitted, but binary-based scanning, penetration testing and exploit generation are blocked. Legitimate security work can hit these classifiers.

Pricing & Access

API Pricing

• Input: $5 / MTok
• Output: $25 / MTok
• Fast mode: twice the base price, for up to 2.5× the default output speed — offered on the Claude Platform and through usage credits in Claude Code
• Prompt caching: cache reads at roughly 10% of base input price; minimum cacheable prefix 512 tokens
• Batch API: 50% discount

Anthropic states that Opus 5 has no data retention requirements for general access — a difference from Fable 5, which requires 30-day retention.

Consumer Plans

Claude's consumer tiers are Pro ($20/month) and Max (from $100/month), alongside a Free tier. Anthropic does not publish which model backs each tier, so we do not claim Opus 5 is reachable on any particular plan — check claude.com/pricing for current entitlements.

Safety and access programs

Anthropic describes Opus 5 as its "most aligned model to date," reporting the lowest deceptive-behavior rates it has measured and a score of 2.3 on overall misaligned behavior in its automated behavioral audit. It reports safety classifiers intervening 85% less often than on Fable 5, and biology-related requests blocked on Fable 5 now route to Opus 5 rather than Opus 4.8. Enterprises whose legitimate cybersecurity work falls inside the blocked categories can apply to the Cyber Verification Program.

Ecosystem & Tools

Official SDKs

• Python SDK — full-featured with async support
• TypeScript SDK — type-safe for Node.js and browsers
• REST API — direct HTTP access from any language

SDKs also ship for Java, Go, Ruby, C#, and PHP, along with the ant command-line interface.

Developer Tools

• Anthropic Console — playground and prompt development
• Claude Code — agentic coding in the terminal, IDE, and web; effort defaults to high here as on the Claude API
• Claude.ai — browser-based chat, alongside Claude Cowork

Community & Resources

• Introducing Claude Opus 5 — official announcement, July 24, 2026
• Claude Opus 5 System Card — safety and capability evaluation
• Models Overview — specs for all current Claude models
• Migration Guide — moving from Opus 4.8 and earlier
• Pricing Page