Claude Opus 4.6 Deep Dive: Every New Feature Explained
Comprehensive review of Claude Opus 4.6 covering adaptive thinking, 1M context, Agent Teams, 500+ zero-days, benchmarks, pricing, and real-world performance. Honest assessment.
Quick Rating
| Category | Score | Notes |
|---|---|---|
| Functionality | 9.5/10 | Agent Teams, 1M context, Adaptive Thinking -- packed release |
| Output Quality | 8.5/10 | Coding is best-in-class; writing has regressed |
| Ease of Use | 8.0/10 | Breaking changes and token cost surprises hurt onboarding |
| Price / Value | 7.0/10 | Powerful but expensive; Fast mode is eye-watering |
| Overall | 8.5/10 | The best coding model available, with real caveats |
One-sentence verdict: Claude Opus 4.6 is the strongest model for agentic coding and long-context reasoning we have tested, but it costs meaningfully more to run, writes worse than its predecessor, and ships with breaking changes that will bite you if you are not paying attention.
Introduction
February 5, 2026. Roughly 6:40 PM ET. Anthropic drops Claude Opus 4.6. The developer community barely has time to open the changelog before -- twenty minutes later -- OpenAI fires back with GPT-5.3 Codex. Two flagship releases in a single evening. X loses its mind. The benchmarks start flying. The hot takes start flying faster.
We have spent the last three days putting Opus 4.6 through its paces. Not just running benchmarks, but using it for the actual work we do every day: writing articles, debugging codebases, analyzing long documents, and building agent workflows. This is our honest assessment of everything that changed, what actually matters, and what you can safely ignore.
If you want to see how Opus 4.6 stacks up against GPT-5.2 and Gemini in a full head-to-head, we already published our comparison of ChatGPT, Claude, and Gemini. This article goes deep on Claude specifically -- every feature, every tradeoff, every hidden cost.
What Is Claude Opus 4.6?
Opus 4.6 is the latest flagship model in Anthropic's Claude lineup, released February 5, 2026. It sits at the top above Sonnet 4.5 and Haiku 4.5, and it replaces Opus 4.5 as the "throw the hardest problems at this one" model.
Specs at a Glance
| Spec | Claude Opus 4.6 | Claude Opus 4.5 (previous) |
|---|---|---|
| Model ID | claude-opus-4-6 | claude-opus-4-5-20250620 |
| Context Window | 200K standard, 1M beta | 200K |
| Max Output | 128K tokens | 128K tokens |
| Thinking | Adaptive (new default) | Extended (manual toggle) |
| API Input | $5 / MTok | $5 / MTok |
| API Output | $25 / MTok | $25 / MTok |
| Fast Mode | $30 / $150 per MTok | N/A |
| Agent Teams | Yes (Claude Code) | No |
| Context Compaction | Yes (beta) | No |
Who is it for? Developers building agentic systems. Teams working with large codebases. Researchers who need to reason across very long documents. Anyone whose problems require Claude to think hard and work autonomously for extended stretches.
Who is it NOT for? Writers who loved Opus 4.5's prose quality. People on tight API budgets. Anyone who relied on assistant message prefilling in their prompts (more on that later -- it is gone).
Feature Deep Dive
Adaptive Thinking
This is the change that will affect the most people in practice, and it is worth understanding properly.
Previously, extended thinking was a binary toggle -- on or off. You turned it on, set a budget_tokens value, and Claude would dutifully burn through that many thinking tokens before answering. The problem was obvious: Claude would happily spend 8,000 thinking tokens on "What is the capital of France?" if you told it to.
Adaptive Thinking (thinking: {"type": "adaptive"}) fixes that. Claude now decides for itself how much to think based on the complexity of the problem. A factual lookup gets minimal reasoning. A multi-step debugging session gets deep chains. You can nudge this with an effort parameter: low, medium, high, or max.
Here is what the API call looks like:
response = client.messages.create(
model="claude-opus-4-6",
max_tokens=16000,
thinking={"type": "adaptive"},
output_config={"effort": "medium"},
messages=[{"role": "user", "content": "What is the capital of France?"}]
)The practical impact: It works well. Claude genuinely does think less on easy questions and more on hard ones. The max effort level unlocks the model's best performance on difficult reasoning and is what powers many of the headline benchmark numbers.
The hidden cost, and it is not small: When Claude decides a problem deserves deep thinking, it generates a LOT of thinking tokens. Developers on the Cursor forums have reported daily API bills jumping 3-5x after switching from Opus 4.5. One user, FinDevAI, reported hitting a $200 daily bill. On the consumer side, HN user smcleod reported that Max subscribers are "exhausting 5-hour limits in 30-40 minutes for light work." This is not a bug -- it is the model doing what it was designed to do. But if you are not monitoring usage carefully, adaptive thinking will surprise you.
The old thinking: {"type": "enabled"} with budget_tokens still works on Opus 4.6 but is officially deprecated. Anthropic will remove it in a future release. If you are building new integrations, use adaptive thinking from the start.
Our take: Adaptive thinking is a genuine improvement to the reasoning system. It makes Claude smarter on hard problems and cheaper on easy ones -- in theory. In practice, the "cheaper on easy ones" part only works if your workload includes a lot of easy problems. If you are primarily using Opus for hard coding tasks, expect higher bills. Set effort to medium as a default and only escalate to high or max when you need it. For comparison, GPT-5.3 takes a different approach: its reasoning pipeline is always-on with no user-facing effort dial, which means less control but also fewer surprises on your bill.
1M Token Context Window
Opus 4.6 can process up to 1 million tokens in a single conversation. That is roughly 750,000 words -- about the combined length of the first four Harry Potter books.
And yes, someone ran exactly that test. Opus 4.6 was given the first four Harry Potter novels as a single prompt (~733K tokens) and asked to find every officially documented spell. It found 49 out of 50. Missed one. That is impressive retrieval across nearly three-quarters of a million tokens of dense narrative text.
On Anthropic's formal benchmark -- MRCR v2 at 1M tokens with 8 needles hidden in the context -- Opus 4.6 scored 76.0%. For comparison, Sonnet 4.5 scored 18.5% on the same test. That is not a typo. It is a 4x improvement.
The restrictions you need to know:
- The 1M context is in beta, not generally available -- you need to be on API tier 4+
- Prompts over 200K tokens trigger premium pricing: $10 / MTok input, $37.50 / MTok output
- That means processing a full 1M token prompt costs about $10 just for input, before any response
The pricing catch in plain terms: A full 1M token prompt with a detailed thinking response could easily run $30-50 for a single API call. This is a tool for specific, high-value workflows -- enterprise document analysis, legal discovery, codebase-wide refactoring. It is not something you turn on for everyday chat.
Our take: The 1M context is impressive and the retrieval quality is genuinely strong. But between the beta restriction and the doubled pricing, this is a feature for teams with specific, well-funded use cases. For most developers, the standard 200K window remains more than enough and far more economical.
128K Output Tokens
Opus 4.6 can generate up to 128K tokens in a single response. That is novel-length output in one shot.
In practice, we have found this most useful for generating large code files, detailed analysis documents, and comprehensive refactors. But there is a subtlety here that matters: adaptive thinking tokens eat into this output budget. A complex reasoning task at max effort might consume 80K tokens on thinking alone, leaving 48K for the actual answer. The doubled output ceiling prevents frustrating truncation in these scenarios.
It is one of those features you rarely need -- until you desperately need it, and then you are grateful it exists.
Context Compaction API
This is a beta feature that lets Claude summarize its own conversation history when things get long. The idea: instead of hitting the context limit and stopping, Claude compacts older messages into a summary and keeps going. Infinite conversations, at least in theory.
response = client.beta.messages.create(
betas=["compact-2026-01-12"],
model="claude-opus-4-6",
max_tokens=4096,
messages=messages,
context_management={
"edits": [{"type": "compact_20260112", "trigger": {"type": "input_tokens", "value": 100000}}]
}
)When input_tokens hits the trigger value (100,000 in this example), Claude automatically summarizes the older context before processing the new message. You can adjust that threshold to balance retention against cost.
Our take: This is one of the most practically useful features in the entire release, and it has gotten surprisingly little attention. Without compaction, long-running agents bump into context limits and either crash or lose crucial context from earlier in the session. With compaction, they can run indefinitely. The summarization is lossy -- you will lose nuance from early messages -- but for iterative coding sessions and agent workflows, the tradeoff is well worth it. If you are building anything agentic, put this on your integration list now.
Agent Teams
This is the feature that generated the most headlines. And honestly, the most controversy.
Agent Teams let you spin up multiple Claude Code instances that work in parallel on a shared codebase. One agent handles the database layer, another works on the API, a third writes tests. They coordinate, share context, and merge their work.
Anthropic demonstrated this by building a C compiler from scratch. The numbers are worth reading twice:
- 16 parallel Claude instances working simultaneously
- ~2,000 Claude Code sessions total
- $20,000 in API costs
- 100,000 lines of Rust code produced
- 2 billion input tokens consumed
- 99% pass rate on GCC torture tests
- Successfully boots Linux 6.9 on x86, ARM, and RISC-V
- Compiled Doom, QEMU, FFmpeg, SQLite, PostgreSQL, and Redis
A team of AI agents, working in parallel, autonomously produced a working C compiler that boots an operating system and compiles Doom. The "can it compile Doom?" test is the new "can it run Crysis?" and the answer is yes. GPT-5.3 Codex, released the same evening, offers its own multi-agent approach, but it is more tightly scoped to code generation within OpenAI's ecosystem rather than Claude's open-ended terminal-based orchestration.
But we need to talk about the $20,000. The Hacker News thread was predictably split. User PostOnce made a fair point: "$20,000 could produce gibberish... unlike human contractors, you don't get to course-correct." User qarl fired back: "Do you have ANY idea what it costs to have humans write 100,000 lines of code?" Both perspectives have merit.
When Agent Teams make sense:
- Large, well-defined projects with clear module boundaries
- Tasks where genuine parallelism provides speedup (you cannot parallelize everything)
- Situations where the cost of AI is still cheaper than the human alternative
- Exploratory projects where speed matters more than cost
When they do not:
- Small or medium projects where a single Claude session would suffice
- Tasks requiring constant human judgment and course correction
- Budget-constrained work where a $20K surprise bill would be a problem
- Projects that do not decompose cleanly into parallel subtasks
Our take: Agent Teams are impressive technology with a narrow sweet spot right now. For most individual developers, a single Claude Code session remains the right tool -- and if you are evaluating which coding assistant fits your workflow best, we break that down in Cursor vs Copilot vs Claude Code. Agent Teams will matter most for enterprises and larger teams building systems where real parallelism helps. The demo is stunning. The practical applicability for a solo developer? Limited, at least for now.
Fast Mode
Opus 4.6 has a "research preview" fast mode that produces output roughly 2.5x faster. The catch: $30 / MTok input and $150 / MTok output. That is 6x the standard Opus rate.
For context, Fast Mode Opus costs about 10x more per output token than GPT-5.2 ($14 / MTok). That is a real number.
We have used Fast Mode for time-sensitive debugging sessions where waiting 30-60 seconds for a standard response broke our flow. It is genuinely faster and the output quality is identical. But $150 per million output tokens is painful even for well-funded teams.
There is a 50% introductory discount running until February 16, 2026 -- so if you want to try it, now is the time.
Our take: Fast Mode is a feature for moments when time is literally money -- production outages, tight deadlines, latency-sensitive applications. For everything else, standard Opus or Sonnet 4.5 is the smarter choice. This is not the mode you leave on by default.
500+ Zero-Day Vulnerabilities
This might be the most impressive and least-discussed part of the entire release.
Anthropic pointed Opus 4.6 at well-known open-source codebases -- GhostScript, OpenSC, CGIF, and others -- and it found over 500 previously unknown high-severity vulnerabilities. Some of these bugs had been hiding in heavily-tested, widely-audited code for decades.
The model does not just scan for known patterns. According to Anthropic, it reads code the way a human security researcher would: looking at past commit history to find similar bugs that were never fixed, spotting patterns that tend to cause problems, and understanding program logic well enough to know what input would break it. In the CGIF case, triggering the vulnerability required understanding the LZW compression algorithm and its relationship to the GIF format -- genuine algorithmic reasoning, not pattern matching.
Each finding was validated by either an Anthropic team member or an independent security researcher.
What this means in practice: If you are responsible for code security, Opus 4.6 is now a serious tool in your arsenal. It is not replacing security researchers, but it can cover far more code far faster than any human team. The fact that it found bugs in GhostScript -- a codebase that has been audited by humans for years -- is a strong signal about the model's capability.
Anthropic also introduced six new cybersecurity-specific probes that monitor model behavior in real time, designed to detect and block potential malicious use of these same capabilities. The dual-use concern is legitimate: the same skills that help defenders also help attackers. Anthropic seems to be taking that tension seriously.
Self-Correction Improvements
This is harder to benchmark than the other features, but it is easy to feel in practice.
Opus 4.6 is measurably better at catching and fixing its own mistakes during code generation and multi-step tasks. Where Opus 4.5 would sometimes confidently produce buggy code and barrel forward, 4.6 more frequently pauses, reviews its own output, and corrects errors before presenting the final result.
Ishank Choudhary, writing on Medium, captured it well: "The model doesn't rush, doesn't jump straight to output, but pauses, structures the problem, and then proceeds step by step."
In our own testing, we noticed fewer "wait, that is obviously wrong" moments during complex refactors compared to Opus 4.5. The model does not always catch its errors -- no model does -- but the hit rate is noticeably higher. This plays into the Terminal-Bench improvement (59.8% to 65.4%) -- sustained autonomous performance requires a model that can recover from its own mistakes, and Opus 4.6 does that better.
To put that Terminal-Bench jump in perspective: the benchmark measures multi-step agentic tasks that require sustained correctness over long sessions, not just one-shot answers. A 5.6 percentage point gain there means fewer dead-end paths, fewer wasted token loops, and fewer moments where the model confidently barrels into a broken solution. On SWE-bench Verified, the gap between Opus 4.5 (80.9%) and Opus 4.6 (80.8%) is effectively zero -- but SWE-bench tests isolated patches, not sustained multi-turn work. The self-correction improvements show up most clearly when Claude has to plan, execute, hit a wall, and recover, which is exactly what real-world coding sessions look like.
Benchmark Results
Numbers without context are noise. Here is what the benchmarks actually tell us, and where they do not tell the full story.
Where Opus 4.6 Leads
| Benchmark | Opus 4.6 | GPT-5.2 | Opus 4.5 | What It Measures |
|---|---|---|---|---|
| Terminal-Bench 2.0 | 65.4% | 64.7% | 59.8% | Sustained agentic coding |
| ARC AGI 2 | 68.8% | 54.2% | 37.6% | Novel reasoning and pattern matching |
| GDPval-AA | 1606 Elo | 1462 Elo | 1416 Elo | Real-world knowledge work |
| BrowseComp | 84.0% | 77.9% | 67.8% | Web browsing and research |
| OSWorld | 72.7% | 41.2% | 66.3% | Computer use tasks |
| BigLaw Bench | 90.2% | -- | -- | Legal reasoning |
| SWE-bench Verified | 80.8% | 80.0% | 80.9% | Software engineering |
The ARC AGI 2 number is the one that made us sit up. Going from 37.6% (Opus 4.5) to 68.8% -- nearly doubling -- on a test specifically designed to measure novel reasoning is not an incremental gain. That is a qualitative shift. You can feel it when working with the model: it handles unusual, unfamiliar problems with more confidence and fewer stumbles.
The GDPval-AA result is also worth noting: 144 Elo points over GPT-5.2 on economically valuable knowledge work tasks is a large gap in Elo terms. This benchmark, covering finance and legal work, likely contributed to the Wall Street reaction we discuss below.
Where Opus 4.6 Loses
No model wins everything, and intellectual honesty requires acknowledging the gaps:
- GPQA Diamond: GPT-5.2 Pro scores 93.2% on graduate-level science reasoning. Anthropic has not published an Opus 4.6 number for this benchmark, which usually means it is not competitive.
- MMMU: Gemini 3 Pro leads on multimodal understanding. Claude has never been the multimodal champion.
- Math (AIME 2025): OpenAI's o4-mini scores 99.5% -- essentially perfect. Claude is not in the same league for pure mathematical reasoning.
- Speed: GPT-5.2 responds roughly 4x faster than standard Opus 4.6. For interactive work, this speed gap is a real usability issue.
The SWE-bench Oddity
Notice that on SWE-bench Verified, Opus 4.6 (80.8%) actually scores marginally lower than Opus 4.5 (80.9%). The difference is within noise, but it is interesting that the most-watched coding benchmark did not improve. Our read: Anthropic optimized for broader agentic capability rather than narrowly maximizing one benchmark. That is probably the right call -- SWE-bench is one test, and Opus 4.6 is clearly stronger across the full suite of evaluations.
Pricing (February 2026)
API Pricing
| Tier | Input | Output |
|---|---|---|
| Standard | $5 / MTok | $25 / MTok |
| Fast Mode (research preview) | $30 / MTok | $150 / MTok |
| Long Context (>200K tokens) | $10 / MTok | $37.50 / MTok |
| Batch | $2.50 / MTok | $12.50 / MTok |
For comparison: GPT-5.2 costs $1.75 / $14 per MTok. Gemini 2.5 Pro costs $1.25 / $10 per MTok. Opus 4.6 is the most expensive frontier model by a wide margin -- roughly 2-3x more expensive than the competition on both input and output.
Consumer Plans
| Plan | Price | What You Get |
|---|---|---|
| Pro | $20/mo | Access to Opus 4.6, usage limits |
| Max | From $100/mo | Higher limits, priority access |
Is It Worth It?
If you are on the consumer side ($20/mo Pro plan), the upgrade to Opus 4.6 is free and the answer is straightforward: yes, try it. You can always switch back to Opus 4.5 for writing tasks in the model selector. The main annoyance is that your usage limits will feel tighter because adaptive thinking burns through tokens faster.
On the API side, the math is harder. The base pricing is unchanged from Opus 4.5 -- $5 / $25 per MTok -- so on paper, nothing changed. In practice, a query that cost $0.03 with Opus 4.5 can easily cost $0.10-0.15 with Opus 4.6 at default effort settings because of all those extra thinking tokens. The effective cost per task has gone up even though the rate card has not.
For many workloads -- especially ones that do not need frontier reasoning -- Sonnet 4.5 at $3 / $15 per MTok or GPT-5.2 at $1.75 / $14 per MTok will get you 85-90% of the quality at a fraction of the cost. Save Opus for the problems that actually need it.
What Developers Are Actually Saying
Three days in, the community is split in a way that tells you a lot about who this model was built for.
The coding crowd is thrilled. Boris Cherny from Anthropic called it the "best model yet... more agentic, more intelligent." Dan Shipper from Every.to said it is the "best AI coder we've tested." If you have read Cursor vs Copilot vs Claude Code, you know that bar was already high. Opus 4.6 raises it.
The writing crowd is frustrated. A Reddit post titled "Opus 4.6 lobotomized" hit 167 upvotes quickly, with users complaining that the model's prose took a noticeable step backward from Opus 4.5. The writing feels more mechanical, less nuanced, less willing to take creative risks. We have tested this ourselves -- running the same writing prompts through Opus 4.5 and 4.6 side by side -- and the difference is real. Opus 4.5 produces more natural, varied prose. It is not subtle.
The emerging consensus: Use 4.6 for coding and reasoning. Keep 4.5 around for writing. That is not ideal -- switching between models adds friction -- but it reflects reality right now.
The cost shock is real. HN user smcleod reported Max subscribers are "exhausting 5-hour limits in 30-40 minutes for light work." On the API side, a Cursor Forum user (FinDevAI) posted about a $200 daily bill with Opus 4.6 as the default model. Adaptive thinking at default settings generates substantially more tokens than people expect. The message from the community is consistent: monitor your usage for the first week. Set alerts. You will probably be surprised.
Wall Street noticed. The same week Opus 4.6 launched alongside Anthropic's expanded Claude Cowork platform, software stocks saw a $285 billion selloff. Thomson Reuters fell nearly 16%. LegalZoom dropped almost 20%. Investors read the GDPval-AA benchmarks -- showing Claude outperforming everything on finance and legal tasks -- alongside the Agent Teams demo and decided that a lot of white-collar software was about to be disrupted. Whether that panic is justified is a longer debate. But it tells you how seriously the financial world is taking these capabilities.
Breaking Changes
If you are an API developer, read this section before upgrading. Not after.
Prefill removal -- the big one. Prefilling assistant messages -- the common pattern where you start Claude's response with specific text like {"role": "assistant", "content": "Here is the JSON:"} -- returns a hard 400 error on Opus 4.6. Not a deprecation warning. Not a gentle nudge. A 400. If your application uses prefilling for structured output, it will break immediately on upgrade. Migrate to output_config.format with json_schema for guaranteed JSON, or use system prompt instructions instead.
Deprecated parameters:
thinking: {"type": "enabled"}withbudget_tokens-- still works but deprecated. Switch to adaptive thinking now.- The
betas=["interleaved-thinking-2025-05-14"]header -- remove it. Interleaved thinking is standard on Opus 4.6. output_formatparameter -- moved tooutput_config.format. Old parameter still works but is deprecated.
JSON escaping changes: Opus 4.6 may produce slightly different JSON string escaping in tool call arguments, including different handling of Unicode escapes and forward slash escaping. If you have brittle JSON parsing downstream, test carefully.
These are not theoretical risks. Test in staging before you switch model IDs in production.
Pros and Cons
Pros:
- Best-in-class agentic coding performance (Terminal-Bench 65.4%)
- Near-doubling of novel reasoning capability (ARC AGI 2: 37.6% to 68.8%)
- Adaptive thinking reduces wasted compute on simple tasks
- 1M token context with strong retrieval quality (76% MRCR at 8-needle)
- Agent Teams enable genuinely new parallel workflows
- Context Compaction makes long-running agents viable
- 500+ zero-day discoveries demonstrate serious security value
- Improved self-correction during code generation
- 128K output handles deep reasoning chains without truncation
Cons:
- Writing quality has clearly regressed versus Opus 4.5
- 3-5x higher token consumption in practice from adaptive thinking
- Most expensive frontier model on the market by a wide margin
- Fast Mode pricing is extreme ($150/MTok output)
- 1M context is beta-only with doubled pricing
- Prefill removal breaks existing integrations with no grace period
- Response speed still lags well behind GPT-5.2 and Gemini
- Agent Teams cost is hard to predict and control
- Consumer usage limits feel dramatically tighter due to token consumption
Who Should Upgrade Now
Get it now if you:
- Build agentic coding systems or use Claude Code heavily
- Work with large codebases that benefit from long context
- Need the strongest available reasoning for complex technical problems
- Are on the $20/mo Pro plan (it is free to try, just keep 4.5 accessible)
- Do security auditing and want to scan code for vulnerabilities
Wait or skip if you:
- Primarily use Claude for writing and creative work
- Are on a tight API budget and cannot absorb 3-5x token increases
- Rely on assistant message prefilling in production
- Need the fastest possible response times for interactive work
- Are happy with Opus 4.5's current performance for your use case
Alternatives to Consider
GPT-5.3 Codex -- Dropped the same evening, literally 20 minutes after Opus 4.6. OpenAI's answer to agent-powered coding. Notably faster and cheaper per token. If you are already in the OpenAI ecosystem, it deserves evaluation. Check our comparison of ChatGPT, Claude, and Gemini for the full breakdown.
Gemini 3 Pro -- Leads on multimodal tasks and has the most generous context window pricing. If your workflow is research-heavy with lots of documents and images, Gemini might be the better fit at a fraction of the cost.
Claude Opus 4.5 -- Still available and still excellent. Better writing quality, same base API pricing, proven stability. If Opus 4.5 is working well for you and you do not need Agent Teams or 1M context, there is no rush to switch. Anthropic has not announced a deprecation timeline.
Claude Sonnet 4.5 -- At $3 / $15 per MTok, Sonnet gives you roughly 85-90% of Opus quality for most tasks at dramatically lower cost. For latency-sensitive applications or budget-constrained workloads, Sonnet remains the smart default.
Final Verdict
Rating: 8.5 / 10
Claude Opus 4.6 is the strongest model available for agentic coding, complex reasoning, and long-context work. The ARC AGI 2 score is not just a benchmark win -- it represents a real, noticeable improvement in how the model handles novel problems. Agent Teams and Context Compaction open up workflows that were previously impossible. The security capabilities are quietly impressive and underreported.
But this is not a clean, across-the-board upgrade. The writing regression is real and disappointing. The token costs are significantly higher in practice than the pricing page suggests. The breaking changes -- especially prefill removal -- will cause real production pain for teams that do not test first. And the speed gap with GPT-5.2 remains frustrating for interactive work.
Our recommendation: If you code with Claude, upgrade. If you write with Claude, keep Opus 4.5 accessible. If you are building agents, this is the model you have been waiting for -- just watch your bill.
For tips on getting the most out of Opus 4.6's reasoning capabilities, including how to structure prompts for adaptive thinking, check out our Prompt Engineering 101 guide.
FAQ
Is Claude Opus 4.6 free to use? On the Pro plan ($20/mo), you get access to Opus 4.6 at no additional cost, subject to usage limits. On the API, you pay per token at $5 / $25 per MTok for standard usage. There is no free tier for Claude.
Can I still use Claude Opus 4.5?
Yes. Opus 4.5 remains available via its model ID (claude-opus-4-5-20250620). Anthropic has not announced a deprecation date. For writing tasks, we actually recommend keeping it around.
How much does the 1M context window cost? Prompts exceeding 200K tokens are charged at $10 / MTok input and $37.50 / MTok output -- double the standard rate. A full 1M token input costs approximately $10 before any output is generated. Add a detailed thinking response and a single call could run $30-50.
Is Adaptive Thinking worth the extra token cost?
For complex reasoning and coding tasks, yes -- the quality improvement is real. For simple queries, set the effort to low or medium to control costs. Monitor your token usage closely for the first week. The 3-5x increase catches almost everyone off guard.
Should I switch from GPT-5.2 to Opus 4.6? It depends on your priorities. Opus 4.6 is better for agentic coding, long-context reasoning, and complex knowledge work. GPT-5.2 is faster, cheaper, and better at math. If response speed and cost matter most, GPT-5.2 might still be the better fit. If you need the absolute best reasoning and coding and are willing to pay the premium, Opus 4.6 justifies it.
What is the difference between Adaptive Thinking and Extended Thinking?
Extended Thinking was a binary toggle: you turned it on, set a fixed budget_tokens value, and Claude used that many thinking tokens regardless of problem difficulty. Adaptive Thinking, the new default on Opus 4.6, lets the model decide how much to think based on the question's complexity. Simple factual lookups get minimal reasoning. Hard multi-step problems get deep chains. You can nudge the depth with an effort parameter (low, medium, high, or max). Extended Thinking still works on Opus 4.6 but is officially deprecated.
Is my data safe with Claude Opus 4.6? Anthropic does not train on API data by default -- your prompts and responses are not used to improve models unless you explicitly opt in. On the consumer side (Pro and Max plans), Anthropic's privacy policy states that conversations may be used for safety research and model improvement unless you opt out. For enterprise and API users, Anthropic offers a zero-retention option. If you are processing sensitive data, the API with zero-retention is the safest path.
Is Claude Opus 4.6 worth upgrading from Pro to Max? If you are a heavy user hitting Pro rate limits regularly -- especially with adaptive thinking burning through tokens faster -- Max gives you higher limits and priority access. At $100/mo it is a significant jump from $20. Our recommendation: try Opus 4.6 on Pro first for a week. If you find yourself constantly waiting for rate limits to reset during coding sessions, Max pays for itself. If you use Claude casually or mainly for writing, Pro is plenty.