Claude Opus 5: What's New, Pricing and What Changes for Developers (2026)

On July 24, 2026, Anthropic released Claude Opus 5 (claude-opus-5), its most capable model for agentic coding and enterprise work: a 1 million-token context window, thinking on by default, a new effort ladder that reaches all the way to max and —crucially— the same price as Opus 4.8: $5 per million input tokens and $25 per million output. Anthropic calls it a step-change over Opus 4.8, not an incremental update, and positions it as frontier intelligence at half the cost of Claude Fable 5.

It’s Anthropic’s fourth model in under two months —after Mythos 5, Fable 5 and Sonnet 5 in June— and it lands two weeks after OpenAI overhauled ChatGPT with the GPT-5.6 family. This guide sorts out what actually changes, from the perspective of someone who ships these models in real products: what’s new, what it costs and how to migrate your code without surprises.

Claude Opus 5 at a glance

• Model ID (API): claude-opus-5
• Context window: 1M tokens (both the default and the maximum; there’s no smaller variant)
• Max output: 128k tokens (up to 300k with the batch beta)
• Pricing: $5 / million input · $25 / million output (unchanged from Opus 4.8)
• Reliable knowledge cutoff: May 2026
• Thinking: on by default
• Effort: low, medium, high, xhigh and max
• Available on: Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and the Pro, Max, Team and Enterprise apps

What is Claude Opus 5?

Claude Opus 5 is the flagship of Anthropic’s Opus family, built for complex agentic coding, deep reasoning and long-horizon tasks. “Opus” is the high-capability line for heavy work; only Claude Fable 5 sits above it —pricier, and aimed at very long-running agents.

The difference from earlier versions isn’t a single benchmark number but consistency: Opus 5 stays on task across multistep reasoning chains, completes multi-file features without leaving TODOs or stubs, and verifies its own work without being asked. On Anthropic’s internal test over the Box platform, Opus 5 beats Opus 4.8 by 8% on average, with a 17% gain in due diligence and 11% in data analysis.

The changes that actually matter

1. Thinking is on by default

On Opus 4.8, requests ran without “thinking” unless you set thinking: {"type": "adaptive"}. On Opus 5, thinking is on by default: the model decides when and how much to reason each turn, and the new control is the effort parameter.

One practical gotcha: since max_tokens is a hard limit covering reasoning plus the response, if you migrate a workload that used to run without thinking, revisit that value or you’ll run out of output room.

2. A full effort ladder: from low to max

Opus 5 ships the full effort ladder: low, medium, high, xhigh and max, with max as the new top tier for the deepest possible reasoning. In practice it’s the “toggle between cost and capability” the press highlighted on launch day: dial effort up when the task warrants it, and down for routine work.

Here’s the part that matters for your bill: at low and medium, Opus 5 delivers very high quality at a fraction of the tokens and latency of the higher settings. On Anthropic’s trading benchmark, the model solved tasks using roughly a seventh of the reasoning tokens Opus 4.8 spent.

Breaking change: disabling thinking (thinking: {"type": "disabled"}) is only accepted at effort high or below. Combine it with xhigh or max and the API returns a 400 error.

3. Performance: agentic coding and deep reasoning

The biggest gains over Opus 4.8 are, in order:

• Deep reasoning: sustains multistep analysis across long problem chains.
• Agentic coding and long-horizon tasks: doesn’t lose the thread in extended tool-use loops and finishes large refactors and end-to-end features.
• Test-time compute scaling: turns more effort (up to max) into better results, predictably.
• Code review and bug-finding: surfaces real bugs at a high rate per pass, with few false positives.
• Vision: reads charts, documents and diagrams, and replicates UI and frontend visuals.

On financial modeling, Anthropic reports 9 percentage points higher accuracy than Opus 4.8 and a third fewer turns to reach the result.

4. Frontier intelligence at half the price

This is the economic headline: Opus 5 delivers, per Anthropic, frontier intelligence at half the cost of Claude Fable 5 (which runs $10 / $50 per million). It approaches Fable 5’s capability while you pay $5 / $25. For most real products, that makes Opus 5 the sweet spot between power and budget —Fable 5 stays reserved for the most demanding autonomous agents.

5. API changes you’ll appreciate

Three quiet but useful improvements:

• Mid-conversation tool changes (beta): add or remove tools between turns while preserving the prompt cache, instead of resending the full tool list all session.
• Default fallback mode: when the model refuses a request for safety reasons, the API can automatically fall back to another model instead of returning an error —a direct answer to refusal frustration.
• Lower cache minimum: the minimum cacheable prompt drops to 512 tokens (it was 1,024 on Opus 4.8), so prompts that were previously too short to cache now save money, with no code changes.

Opus 5 vs Opus 4.8, Sonnet 5 and Fable 5

Which Anthropic model should you pick in 2026? Here’s the current family at a glance:

In short: Sonnet 5 if you prioritize speed and cost, Opus 5 as the workhorse for complex work, and Fable 5 only when you need the ceiling of capability and can pay for it.

Claude Opus 5 pricing

Opus 5 keeps Opus 4.8’s rate, with no increase:

Since thinking counts as output tokens, your real bill depends heavily on the effort level you choose. Before moving a workload to Opus 5, estimate the cost with an LLM token counter and compare it with what you already pay —I put together a full breakdown in how much the OpenAI, Claude and Gemini APIs cost in 2026.

What changes for developers using the API

Migrating from Opus 4.8 is, at its core, changing the model ID:

import anthropic

client = anthropic.Anthropic()

with client.messages.stream(
    model="claude-opus-5",          # before: "claude-opus-4-8"
    max_tokens=64000,               # leave headroom: thinking also consumes max_tokens
    output_config={"effort": "high"},  # low | medium | high | xhigh | max
    messages=[
        {"role": "user", "content": "Refactor this module and explain the changes."}
    ],
) as stream:
    response = stream.get_final_message()

print(response)

Beyond the ID, three things to review:

1. Thinking on by default. You don’t need the thinking field. If your integration must run without thinking, remember it’s only valid at effort: high or below.
2. Remove inherited verification instructions. Opus 5 self-verifies; phrases like “add a final verification step” or “use a subagent to verify” cause over-verification. Delete them from your old prompts.
3. Longer, more narrated responses. By default Opus 5 writes longer deliverables, narrates its progress more often and delegates to subagents more readily. If you need concise answers, ask for it explicitly in the prompt.

A minor but important detail: with thinking disabled, Opus 5 can occasionally write a tool call into its text instead of emitting a tool_use block. Anthropic’s recommendation is to keep thinking on and control spend by lowering effort.

Where to use Claude Opus 5 today

It’s available from launch day on:

• Claude API as claude-opus-5 (to all customers).
• Amazon Bedrock (anthropic.claude-opus-5), Google Cloud and Microsoft Foundry.
• The Claude apps for Pro, Max, Team and Enterprise plans.

Opus 4.8 stays available on all of those platforms, so you can migrate at your own pace. If you build agents that orchestrate multiple models, Opus 5 fits as the reasoning “brain” —I cover it in detail in the guide to agentic AI and AI agents in 2026.

Early reactions

It wasn’t all applause. In the first hours, part of the community complained about usage limits in the apps —many users expected a quota reset that didn’t arrive with the launch— and that Fable 5, the bigger sibling, had gained a reputation for “burning” tokens and blowing budgets. Anthropic seems to have taken note: Opus 5 is noticeably more efficient at lower effort levels and ships automatic fallback to soften refusals. As always, the proof is in your own use case: measure before you commit to a model.

What it means for your business

If you build AI into your product or your processes, the practical read on Opus 5 is simple: more capability at the same price, plus a lever (effort) to tune cost and quality task by task. That makes viable cases that used to fail on budget —code review, document analysis, support agents, back-office automation— without jumping to the family’s most expensive tier.

At OrtaMarco I help SMBs and teams integrate these models into real workflows: picking the right model, controlling token spend and getting the agent to production. If you’re evaluating Opus 5 for your business, get in touch and we’ll ground it in your case.

Frequently asked questions

When was Claude Opus 5 released? On July 24, 2026, available from day one on the Claude API, Amazon Bedrock, Google Cloud, Microsoft Foundry and the Claude apps (Pro, Max, Team and Enterprise).

How much does Claude Opus 5 cost? $5 per million input tokens and $25 output —the same as Opus 4.8— with up to 90% savings from prompt caching and 50% from batching.

How is it different from Opus 4.8? Thinking on by default, an effort ladder up to max, self-verification, greater consistency on long tasks and far lower reasoning-token spend, all at the same price.

What is the effort ladder? A five-level parameter (low to max) that tunes how much the model reasons: the toggle between cost and capability.

How do I migrate from Opus 4.8? Change the ID to claude-opus-5, revisit max_tokens, remove manual verification instructions and note that disabling thinking requires effort: high or below.

Opus 5 or Fable 5? Opus 5 for complex day-to-day work; Fable 5 only when you need maximum capability and can pay double.

---

Sources: Claude Opus (Anthropic), Models overview (Anthropic Docs) and the effort documentation. Facts verified on July 24, 2026.