Skip to content
CLAUDEXIA
enru
Claudexia TeamMODELS

Claude Opus 5: the flagship guide

What Opus 5 is, what changed from the previous revision, when it earns its price, and how to call it through Claudexia on either API format.

Anthropic shipped Claude Opus 5 on July 24, 2026. We added it to the Claudexia lineup the same day, and since then it's the model to reach for first when the task is genuinely hard.

What Opus 5 is

Opus 5 is the flagship model in the Claude lineup. It took over from Opus 4.8 as the default pick for heavy work: architecture, multi-file refactors, long agentic runs. Through Claudexia you call it by id claude-opus-5, the same way you'd call any other model on our endpoint.

What changed

Three things show up in practice right away.

  1. Speed. On long agentic runs the difference from the previous revision is visible without any benchmarking, especially on scenarios with many steps.
  2. Mid-conversation tool changes. The set of available tools used to be locked in at the start of a conversation. Now you can change it mid-dialogue without starting a fresh session.
  3. Automatic fallback to Opus 4.8. If a safety classifier flags a request, the call doesn't error out, it quietly routes to Opus 4.8 instead. For your code this means: budget and logic should assume some slice of traffic gets physically served by the neighboring model.

When it earns its price

The flagship is for tasks that require holding many relationships in mind at once:

  • reading an unfamiliar repository to understand the whole architecture
  • refactoring that touches a dozen files coherently
  • debugging races and intermittent bugs where the cause isn't obvious on the first pass
  • agent planning where the model chooses its own next step across several iterations
  • long code or contract review where losing context halfway through matters

If the task can be described as "take this input and turn it into that output by pattern", Opus 5 is overkill for it.

How to call it

One base URL: https://api.claudexia.tech. Both formats work, Anthropic Messages and OpenAI Chat Completions, and the same sk_cdx_... key fits either one.

Via the Messages API:

curl https://api.claudexia.tech/v1/messages \
  -H "x-api-key: sk_cdx_..." \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{
    "model": "claude-opus-5",
    "max_tokens": 1024,
    "messages": [{"role": "user", "content": "Review this repository structure and propose a refactor plan"}]
  }'

Via the OpenAI SDK in Python:

from openai import OpenAI

client = OpenAI(api_key="sk_cdx_...", base_url="https://api.claudexia.tech/v1")

resp = client.chat.completions.create(
    model="claude-opus-5",
    messages=[{"role": "user", "content": "Review this repository structure"}],
)
print(resp.choices[0].message.content)

Via the Anthropic SDK in TypeScript:

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic({
  apiKey: "sk_cdx_...",
  baseURL: "https://api.claudexia.tech",
});

const msg = await client.messages.create({
  model: "claude-opus-5",
  max_tokens: 1024,
  messages: [{ role: "user", content: "Plan a refactor of the auth module" }],
});

Same key across all three examples, only the client and the URL path change.

What it costs

Anthropic's own list price for Opus 5 is $5 per million input tokens and $25 per million output tokens. Through Claudexia the same model runs $0.4 per million input and $0.4 per million output, several times cheaper, which is why we call this our actual price rather than an introductory discount.

Estimating your own bill is easier with the calculator on the site: pick the model, enter your volumes, get a number without doing the arithmetic yourself. The dashboard also breaks usage down per key, so if a team splits projects across several keys, spend stays visible separately.

Honest note: when a cheaper tier is enough

A flagship is still a flagship, even at our price. For classifying inbound messages, format conversion, parsing logs, or generating boilerplate from a pattern, reach for claude-sonnet-5 or gpt-5.6-luna instead — the quality gap barely shows on that kind of work, and the bill drops noticeably. A pattern that works well: let Opus 5 read the task once and produce a plan, then have a cheaper model execute the mechanics against that plan.

In short

Opus 5, the flagship since July 24, 2026: faster than the previous revision, can swap tools mid-conversation, and quietly falls back to Opus 4.8 when the safety classifier flags a request. Reach for it when you need to hold the whole system in mind, and hand routine work to cheaper models. Call it as claude-opus-5 through either API format on one key.

← Back to all posts