Claude Opus 4.6 Review: The Best Coding AI Has a Price Problem

I started testing Opus 4.6 within hours of its February 5th launch. Three days later, after running it through a full-stack refactor, a 40-file test suite, and a multi-agent experiment that cost me more than my weekly grocery bill, I can say this: it is the best coding model I have ever used. It is also the most expensive to run at scale, and knowing when NOT to use it might be the most valuable skill a developer learns this year.

What is Claude Opus 4.6?

Claude Opus 4.6 is Anthropic's flagship AI model, released on February 5, 2026. It sits at the top of a three-tier model family (Opus, Sonnet, Haiku) built by a company now valued at $350 billion with over $9 billion in annual revenue. The model is available through the Claude API, claude.ai, and every major cloud platform from day one: AWS Bedrock, Google Cloud Vertex AI, Microsoft Azure Foundry, Snowflake Cortex AI, and GitHub Agent HQ.

What makes this release different from previous Claude updates is scope. This is not an incremental improvement. Opus 4.6 combines a 1-million-token context window (now generally available as of March 13, 2026), 128K token output limits, a four-level Adaptive Thinking system, and an experimental multi-agent capability called Agent Teams. Anthropic is not just building a smarter chatbot. They are building an autonomous software engineering platform. The benchmarks suggest they are pulling it off.

Key Features

1-Million-Token Context Window

Previous Opus models topped out at 200K tokens. Opus 4.6 jumps to 1 million, and as of March 13, 2026, the full 1M window is generally available at standard pricing with no long-context premium. A 900K-token request is billed at the same per-token rate as a 9K one. The model actually uses that context well: on the MRCR v2 long-context retrieval benchmark, Opus 4.6 now scores 78.3%, the highest among frontier models at full context length. Media limits also expanded 6x, supporting up to 600 images or PDF pages per request (up from 100). A Context Compaction feature automatically summarizes older conversation history, preventing the "context rot" that degrades long sessions.

Adaptive Thinking (4 Levels)

Instead of a binary thinking mode (on or off), Opus 4.6 offers four intensity levels: low, medium, high, and max. This is a practical improvement that saves real money. Need a quick variable rename? Use low. Debugging a race condition across three services? Crank it to max. The granularity lets you match intelligence to task complexity instead of paying for deep reasoning on every request.

128K Token Output

The previous output ceiling was roughly 8K tokens. Opus 4.6 pushes that to 128K, which means the model can generate entire files, comprehensive test suites, and multi-page reports in a single response without fragmentation. For developers working on large codebases, this eliminates the frustrating "continue generating" loop that breaks context between responses.

Benchmark Dominance

The numbers tell a clear story. Opus 4.6 scores 65.4% on Terminal-Bench 2.0, which was the highest score recorded at launch (since surpassed by Gemini 3.1 Pro at 68.5%). Its GDPval-AA score of 1606 Elo sits 144 points above GPT-5.2 (1462 Elo), a gap that represents a meaningful difference in knowledge-work reasoning. It ranks first on Humanity's Last Exam, scores 90.2% on BigLaw Bench with 40% perfect scores, and leads BrowseComp for information retrieval. On OSWorld, a benchmark for autonomous desktop GUI operation, Opus 4.6 scores 72.7%, the highest among frontier models and a category where neither Gemini nor GPT has published comparable results. Anthropic also reports the model found 500 zero-day vulnerabilities in open-source codebases, up from roughly 250 with its predecessor.

Using Opus 4.6 in Practice

I ran Opus 4.6 through Claude Code for three days of real development work. The first test was a full-stack refactoring task: restructuring a Next.js application with 40+ components, migrating API routes, and generating corresponding Vitest test files. On Sonnet 4.5, this kind of task requires constant hand-holding. Files get missed, imports break, and the model loses track of what it already changed. Opus 4.6 loaded the entire project into its 1M context window and worked through the changes systematically. It caught dependency chains I had missed and proactively updated import paths across files I had not mentioned.

The Agent Teams experiment was more revealing. I enabled the feature (it requires setting a specific environment variable, since it is still experimental and disabled by default) and spun up 4 agents to parallelize a test-writing task. The agents coordinated through Git, claimed files independently, and produced working test coverage for 12 modules in roughly 20 minutes. The code quality was solid but not expert-level. I spent another 30 minutes cleaning up redundant assertions and improving edge-case coverage. The total API cost for that 20-minute session was approximately $14.

"Agent Teams is genuinely useful for parallelizable tasks. It is not magic, and it is not cheap, but for test generation, documentation, and boilerplate refactoring, the time savings are real."

Addy Osmani, Chrome engineering lead, on Agent Teams

# Enable Agent Teams (experimental, disabled by default)
export CLAUDE_AGENT_TEAMS=1

# Start a multi-agent session with 4 workers
claude-code --agents 4 --task "Write Vitest tests for all components in src/components/"

# Monitor agent coordination
claude-code --agents status

The Adaptive Thinking feature proved its value quickly. I defaulted to "medium" for general development and switched to "high" only for complex debugging. The cost difference is noticeable: a 30-minute session on "high" consumed roughly 3x the tokens of the same duration on "low." After three days, I settled on using "low" for file generation and formatting, "medium" for feature development, and "high" for architecture decisions and debugging.

Pricing

Anthropic held pricing steady from Opus 4.5 despite the capability jump, and in March 2026 eliminated the long-context premium entirely. The full 1M window now costs $5/$25 per MTok regardless of context length, a significant improvement over the previous $10/$37.50 for requests above 200K tokens. But $25 per million output tokens is still the most expensive option in the frontier model market. GPT-5.2 charges roughly $15 for output; Gemini 3.1 Pro charges $12. For agent workflows that consume billions of tokens, the gap compounds fast. The C compiler demonstration that Anthropic used to showcase Agent Teams consumed 2 billion input tokens and 140 million output tokens. Rough math: that is about $13,500 in API costs for a single project (down from $20,000 before the pricing update). Consumer access is available through claude.ai on Pro ($20/month), Max, Team, and Enterprise plans. Pricing verified March 2026.

Pros and Cons

How It Compares

Opus 4.6 wins on expert knowledge work and output capacity by clear margins. GPT-5.2 is the better value for general-purpose work at lower output pricing. Gemini 3.1 Pro leads on abstract reasoning benchmarks (77.1% ARC-AGI-2) at the lowest price point ($2/$12 per MTok) but trades latency for intelligence. The right choice depends on your use case: expert knowledge tasks and agent workflows favor Opus 4.6; async pipelines with budget constraints favor Gemini 3.1 Pro.

Who Should Use Claude Opus 4.6?

This is a specialist tool. It is the best at what it does, but what it does is not cheap. Match your usage to where the model genuinely outperforms the alternatives.

• Best for: Professional developers and engineering teams building autonomous AI agents, working with large codebases (100K+ lines), or needing top-tier code generation and debugging. The benchmark leads are real and translate to noticeably better results on complex tasks.
• Great for: Enterprise teams that need multi-cloud deployment flexibility and can leverage the Batch API and prompt caching to manage costs. Also strong for legal analysis (90.2% BigLaw Bench) and security research (500 zero-day findings).
• Skip if: You primarily need a general-purpose AI assistant, work on smaller projects, or have tight API budgets. Use Sonnet 4.5 for 80% of daily development tasks. It is 40% cheaper on input and 40% cheaper on output, and handles most coding work well. Consider Gemini 3 Pro if you need the largest context window at the lowest price.

Safety and Alignment

Anthropic's System Card for Opus 4.6 reports that the model maintains "exceptionally low rates of deception, sycophancy, or misuse cooperation," on par with Opus 4.5, which was previously considered the most highly aligned frontier model. The model's safety evaluations show it refuses harmful requests at rates consistent with its predecessor while being significantly more capable.

There is a paradox worth noting. The same enhanced planning capabilities that make Opus 4.6 a superior coding agent also make it theoretically more capable of subtle task completion if misaligned. The System Card acknowledges "increased competence in subtly completing suspicious side tasks" as a capability that exists but is not observed in practice. Anthropic frames this as a known trade-off of building more capable models and addresses it through Constitutional AI training and ongoing red-teaming. For enterprise deployments, this transparency matters more than the conclusion: you can read exactly what the model can and cannot be trusted with.

Enterprise Deployments

Beyond benchmarks, Opus 4.6 is seeing real enterprise adoption. Rakuten deployed the model for autonomous issue management across a 50-person engineering organization spanning 6 repositories, using Agent Teams to triage, assign, and in some cases resolve issues without human intervention. On the financial services side, Opus 4.6 achieved 60.7% accuracy on SEC filings research tasks, a 5.47 percentage point improvement over Opus 4.5 and the highest score among frontier models for this specific use case. Legal teams report the model maps structural vulnerabilities and clause inconsistencies across multi-source documents with roughly 10 percentage points better accuracy than previous generations.

The Bigger Picture: Vibe Coding in 2026

Opus 4.6 lands in a market that barely existed two years ago. Today, 92% of US developers use AI coding tools daily. An estimated 41% of all code is now AI-generated. But the picture is not entirely rosy: 63% of developers report spending more time debugging AI-generated code than they would have spent writing it manually, and 44% of engineering leaders observe declining fundamental coding skills among junior developers. Gartner projects 80% of organizations will shrink dedicated developer teams by 2030.

Anthropic is betting that the solution to AI-generated code quality problems is better AI models, not fewer AI models. Agent Teams pushes this thesis to its logical extreme: replace not just the code-writing step, but the entire coordination layer of a development team. Whether that bet pays off depends on whether models like Opus 4.6 can close the quality gap fast enough to justify the cost. Right now, the quality is impressive but not expert-level. The cost is manageable for high-value projects but prohibitive for everyday development. That gap is where the real competition will play out over the next 12 months.

Alternatives to Consider

If Opus 4.6 is not the right fit, these are the strongest alternatives in the current market.

GPT-5.2 (OpenAI): Better if you need a lower output price ($15 vs $25 per MTok) with strong general reasoning. The Codex agent system is more mature than Agent Teams, though limited to single-agent workflows. Best for mixed workloads that span coding, writing, and analysis.

Gemini 3.1 Pro (Google): The cost-effective frontier option at $2/$12 per MTok with the highest ARC-AGI-2 score (77.1%) among broadly available models. Excels at abstract reasoning and tool coordination but has a 35-second time-to-first-token on High mode. Best for async pipelines and budget-conscious enterprise teams. Read our Gemini 3.1 Pro review.

Claude Sonnet 4.5: If you are already in the Claude ecosystem, route most work to Sonnet ($3/$15 per MTok) and reserve Opus for the 10-15% of tasks that genuinely need it. Sonnet handles standard development, code review, and analysis well. This is the strategy Anthropic themselves recommend.

Frequently Asked Questions

Is Claude Opus 4.6 worth the price?

It depends on your workload. For complex multi-file refactors, deep debugging, agent workflows, and expert-knowledge tasks (legal, finance, security), Opus 4.6 delivers measurably better results than any competitor and the time savings easily justify the cost. For routine development, code review, and general-purpose tasks, Sonnet 4.6 at $3/$15 per MTok handles 80% of the work at a fraction of the price. The recommended strategy is to route most work to Sonnet and escalate to Opus only when the task demands it.

How does the 1M context window compare to Gemini?

Both Opus 4.6 and Gemini 3.1 Pro offer 1M token context windows, but the retrieval quality differs dramatically. Opus 4.6 scores 78.3% on MRCR v2 at full context length, meaning it reliably finds and uses information across the entire window. Gemini 3.1 Pro drops to 26.3% at 1M tokens, making its full context unreliable for large-document retrieval. Opus also outputs up to 128K tokens (vs Gemini's 64K) and now supports 600 images or PDF pages per request.

What is Agent Teams and should I use it?

Agent Teams lets up to 16 Claude instances work in parallel on a shared codebase, coordinating through Git without a central orchestrator. It is still experimental and disabled by default. It works well for parallelizable tasks like test generation, documentation, and boilerplate refactoring. It is not recommended for tasks requiring tight coordination or where code quality needs to be expert-level out of the box. Expect to spend time cleaning up outputs.

Can Claude Opus 4.6 handle audio and video?

Opus 4.6 accepts text, images, and PDFs as input but does not natively process audio or video at the API level. Gemini 3.1 Pro is currently the only frontier model with native multimodal input spanning text, images, audio (approximately 8.4 hours), and video (around 45 minutes). If your workflow requires audio or video understanding, Gemini has the edge.

Final Verdict

Claude Opus 4.6 is the most capable AI coding model available in early 2026. Its benchmark leads are not marginal; they are decisive. The 65.4% Terminal-Bench score, the 144-point Elo gap over GPT-5.2, the 1M context window that actually retrieves information reliably, and the 128K output limit combine to create a model that handles complex autonomous coding tasks better than anything else on the market. Agent Teams is rough around the edges and expensive, but it is a genuine glimpse of where AI development is heading.

The catch is price. At $25 per million output tokens, Opus 4.6 is a precision tool, not an everyday driver. Use it strategically. Use Sonnet for the bulk of your work, use Haiku for validation, and bring in Opus when the task demands the best. With that discipline, it is worth every token.