Back to all articles
5 min read

Claude Opus 4.6: 1M context, pricing and limits in 2026

By Dorian Laurenceau

Updated

Fact-checked October 19, 2026. This is a guide to a specific version. The official Opus 4.6 page now labels it legacy and says it remains available. Consult our current model overview when selecting a new model.

LearnIA has not run a paid model comparison. Product facts come from the linked sources; the fictional exercise and acceptance criteria below are our proposed evaluation method.

Correcting outdated launch information

The March 13 announcement made 1M context generally available on Claude Platform without the premium for requests above 200K tokens or a required beta header. A guide still treating that beta arrangement as the current default is outdated.

The model specifications list the API identifier claude-opus-4-6, a 1M context window, standard maximum output of 128K tokens and text/image inputs with text output. Capacity does not guarantee correct use of every document detail.

API pricing and a transparent calculation

The Claude Platform price list lists $5 per million input tokens and $25 per million output tokens. Opus 4.5 has the same standard rates there: our earlier claim of a 67% cut between these two versions was incorrect. App subscriptions, API access, cache, tools and cloud offerings need separate checks.

A fictional request with 20,000 input and 2,000 output tokens costs 0.02 × 5 + 0.002 × 25 = $0.15, excluding cache, tools and taxes. Repeated attempts add cost. With a zero API budget, prepare the task and rubric without making a paid request; only use an already authorised account within available allowances.

Judge adaptive thinking by verifiable output

The model page identifies adaptive thinking. For an evaluation, inspect conclusions, sources, calculations and missing information. A persuasive explanation is not evidence that every source was read correctly. Ask for checkable justifications and source passages rather than treating internal reasoning as a validation method.

Start with a short table of facts, status and source. Verify it before requesting polished narrative. This exposes errors early and avoids spending review time on an attractive but unusable document.

What a published benchmark does and does not show

The March context announcement reports 78.3% on MRCR v2, a long-context retrieval benchmark. This is an Anthropic-published result, not an accuracy rate for your reports. Keep launch figures and later measurements attached to their dates and conditions. Source: 1M context announcement.

Record model version, task set, tools, budget, mode and date before citing a score. Our benchmark reading guide explains why coding tests, preference votes and document retrieval cannot be merged into one universal verdict.

Workshop: preserve an unresolved decision

Download the fictional comparison pack. It contains French-language source documents and a machine-readable answer key: a schedule, revised quotation and pending approval. Write the expected facts before inspecting any model output.

The reference preserves a planned budget of €1,200, a €980 pre-tax quotation and a €1,176 tax-inclusive total. November 19 is proposed; Laura’s approval is still pending. A quotation is not evidence of payment. A correct sum cannot compensate for inventing approval.

Prepare an Opus 4.6 evaluation

Prepare your instruction here. This module sends no data to a model: copy your brief into your chosen tool.

View the starter instruction
Read the three source documents and produce a table of fact, status and exact source. Distinguish budget, quotation and payment. Preserve proposed dates as proposals. State missing approvals. Do not invent decisions or claim to have verified an unreadable file.

Verify the output (0/4)

Accept the output only if those five reference elements are retained without invented approval. Check each citation against the files. Repeat with a case containing no date: the expected answer is “not stated”. Use the comparison workbook to record observations. Its blank cells are not results of completed tests.

Loading diagram…

Keep the version or evaluate a replacement?

Retain the same reference cases, format and rejection criteria when evaluating a replacement. Test the downstream workflow too: valid JSON can still contain a false date. Record retries and review effort alongside API cost. A newer name is a reason to investigate, not proof of fitness for your task.

Start with our source-based document summary method if you do not need an API. Check provider documentation and actual account availability on the day you make an integration decision.

Question 1 of 3.

Test Your Knowledge

0 / 3 answered

Q1.Does 1M context guarantee an accurate summary?

3 questions remaining

Go deeper — free guide

Module 1 — LLM Anatomy & Prompt Structure

Understand how LLMs work and construct clear, reusable prompts.

D

Dorian Laurenceau

Full-Stack Developer & Learning Designer

Full-stack web developer and learning designer. I spent 4 years as a freelance full-stack developer and 4 years teaching React, JavaScript, HTML/CSS and WordPress to adult learners. Today I design learning paths in web development and AI, grounded in learning science. I founded learn-prompting.fr to make AI practical and accessible, and built the Bluff app to gamify political transparency.

Prompt EngineeringLLMsFull-Stack DevelopmentLearning DesignReact
Published: February 6, 2026Updated: October 19, 2026

One practical AI workflow in your inbox

Methods, resources and new exercises. Free, unsubscribe at any time.