On February 5, 2026, Anthropic released Claude Opus 4.6, its first major model of the year and an upgrade aimed at long, complex work: planning more carefully, sustaining agentic tasks for longer, working reliably in large codebases and reviewing code more effectively. Opus 4.6 offered a one-million-token context window in beta, and the release introduced adaptive thinking, in which the model decides for itself when deeper reasoning is worth the cost. Pricing stayed at the $5 and $25 per million token level that Anthropic set with Claude Opus 4.5, making frontier-level capability considerably cheaper than the Opus models of 2025.

Key Facts#

  • Release date: February 5, 2026. API model ID claude-opus-4-6.
  • Pricing per million tokens: $5 input and $25 output. Prompts above 200,000 tokens cost $10 and $37.50, and US-only inference costs 1.1 times the standard rate.
  • Context and output: a 1-million-token context window in beta on the Claude Platform, and up to 128,000 output tokens.
  • Vendor-reported results: 53.0% on Humanity's Last Exam with tools; about 78% on SWE-bench Verified on average; 76% on the eight-needle, one-million-token variant of the MRCR v2 long-context test, versus 18.5% for Claude Sonnet 4.5; and a lead of about 144 Elo points over GPT-5.2 on the GDPval-AA knowledge-work evaluation.
  • API features: adaptive thinking, effort levels from low to max with high as the default, and context compaction in beta for long-running tasks.
  • Product features: agent teams in Claude Code as a research preview, Claude in PowerPoint as a research preview, and upgrades to Claude in Excel.
  • Availability: Claude apps, the Claude API, Amazon Bedrock, Google Cloud's Vertex AI and Microsoft Foundry.

What Happened#

Anthropic's announcement focused on endurance and judgment rather than raw speed. It said Opus 4.6 plans more carefully, sustains agentic tasks for longer and operates more reliably in larger codebases. The benchmark selection reflected that emphasis: Terminal-Bench 2.0 for command-line agents, BrowseComp for finding hard-to-locate information, and GDPval-AA, an evaluation of economically valuable knowledge work, where Anthropic reported a win rate of roughly 70% against OpenAI's GPT-5.2.

The long-context result stood out. On the eight-needle version of MRCR v2, a long-context retrieval test, Opus 4.6 scored 76% at one million tokens, compared with 18.5% for Sonnet 4.5. Large context windows had existed for more than a year, but models often struggled to use the far end of them. A large jump in multi-needle retrieval suggested the window was usable rather than nominal.

Adaptive thinking changed how reasoning is configured. Instead of developers deciding whether to enable extended thinking, the model decides when it helps, while the effort parameter sets how hard it should try overall. Context compaction, in beta, automatically summarizes older parts of a conversation so that long-running agents do not run out of room.

On the product side, Claude Code gained agent teams, a research preview in which several agents coordinate on a task, and Anthropic extended Claude into PowerPoint while improving its Excel integration. On safety, Anthropic said Opus 4.6 is as well aligned as Opus 4.5 and has the lowest over-refusal rate, meaning fewer refusals of harmless requests, of any recent Claude model. Partner quotes came from GitHub, Harvey, Rakuten and Notion, among others.

Background#

The groundwork was laid in November 2025, when Anthropic released Claude Opus 4.5 at $5 and $25 per million tokens, down from the $15 and $75 charged for Claude Opus 4, and introduced the effort parameter. That price change made Opus a realistic option for far more coding and agent workloads.

Competition in the preceding months was fierce. Google launched Gemini 3 in November 2025, and OpenAI released GPT-5.2 in December. CNBC framed the Opus 4.6 launch as part of a move toward a "vibe working" era, with AI agents taking on broader office work, and coverage of the launch noted that enterprise customers make up roughly 80% of Anthropic's business.

Why It Matters for Developers#

The pricing structure is the first thing to model. A one-million-token window is valuable for large codebases and document sets, but prompts above 200,000 tokens are billed at a premium, so the window is best treated as a capability for specific jobs rather than a default. Track input sizes per feature and alert on growth, using the techniques in our guide to observability and cost control for LLM apps.

Adaptive thinking and effort levels also shift responsibility. Rather than tuning reasoning per request, you set an effort budget per workload and let the model decide the rest. That makes evaluation more important, because behavior can vary with task difficulty. Build regression tests around the tasks you care about before switching models.

For .NET teams, the official Anthropic C# SDK implements IChatClient, so Opus 4.6 plugs into Microsoft.Extensions.AI pipelines alongside other providers:

C#
using Anthropic;
using Microsoft.Extensions.AI;

// AnthropicClient reads ANTHROPIC_API_KEY from the environment.
AnthropicClient anthropic = new();

IChatClient client = anthropic.AsIChatClient("claude-opus-4-6")
    .AsBuilder()
    .UseFunctionInvocation()
    .Build();

var response = await client.GetResponseAsync(
    "Review this diff and list changes that could break backward compatibility: ...");
Console.WriteLine(response.Text);

Availability in Microsoft Foundry matters for organizations standardized on Azure. The SDK ships a separate Anthropic.Foundry package for calling Claude through Microsoft Foundry; see our guide to Azure OpenAI and Azure AI Foundry for the surrounding setup. Finally, agent teams and context compaction point to where agent design is heading: multiple coordinated workers and managed memory. Our guide to AI agent architecture patterns covers how to apply those ideas in your own code.

What's Next#

Anthropic kept a fast cadence after Opus 4.6. It enabled a fast mode for the model two days later, released Claude Sonnet 4.6 on February 17, 2026, and followed with Claude Opus 4.7 in April and Claude Opus 4.8 in May, both at the same $5 and $25 pricing. Opus 4.6 remains an active model, and Anthropic's deprecation page lists its tentative retirement as not sooner than February 5, 2027.

For teams choosing an Opus version today, the practical question is whether a newer model's gains justify re-validation. Anthropic's deprecation page notes that Opus 4.7 and later reject non-default temperature, top_p and top_k values, so check your request settings before you upgrade.

Sources#