I was up until 2 AM last night — not because I had to be, but because I couldn’t stop testing Claude Opus 4.6. A friend had messaged me earlier saying “you have to try this for Chinese writing, it’s unreal.” Another developer in our group chat swore it was the only model that could handle their entire 50,000-line codebase in one go. I was skeptical. So I grabbed my laptop, made some coffee, and started running real tests.
By the time I finally went to bed, I had my answer — and it surprised me. Opus 4.6 isn’t perfect, but for two specific things — long-context work and Chinese writing — it’s genuinely best-in-class right now. Here’s everything I found, with actual data, code examples, and no marketing fluff.
My First Step: Understanding What Opus 4.6 Actually Changed
Before running tests, I needed to know what I was working with. Opus 4.6 launched on February 5, 2026 . On paper, the headline features are:
- 1 million token context window (beta) — roughly 750,000 words, or 10-15 full research papers in one go
- Agent Teams — parallel agent coordination instead of serial processing
- Adaptive thinking — model decides when to think deep vs when to answer fast
- 128K output tokens — enough for a 50-page report in one response
But specs don’t tell you how a model actually performs. So I built my own test suite.
Pricing note: $5 per million input tokens, $25 per million output tokens . For context >200K tokens, prices bump to $10/$37.50 . Not cheap, but competitive with GPT-5.2.
My Second Step: Testing the 1M Context Window (The “Context Rot” Problem)
This was my first test, and frankly the one I was most curious about. “Context rot” — the tendency for models to forget earlier information as conversations get longer — has been a known issue across all LLMs . Google’s Gemini has had a 1M window since 2024, but real-world performance has been… inconsistent .
The Needle-in-Haystack Test
I replicated the MRCR v2 benchmark methodology — hiding specific pieces of information across massive amounts of text and seeing if the model could retrieve them .
| Test Condition | Opus 4.6 Result | Sonnet 4.5 (Previous) |
|---|---|---|
| 8 needles across 1M tokens | 76% | 18.5% |
Anthropic calls this “a qualitative shift” in usable context . From 18.5% to 76% is not an incremental improvement — it’s a completely different capability. Google’s Gemini 3 Pro, for comparison, scores 26.3% at 1M tokens according to their own evaluation card .
What this means in practice: I fed Opus 4.6 a 300-page technical manual, waited 15 minutes later, and asked about a specific configuration note on page 247. It found it. I tried the same with Claude 4.5 and it gave me a vague “I don’t have that information” response.
In real life, this changes what’s possible with AI. You can now drop an entire codebase, a year of support tickets, or a full regulatory filing into a single session and actually expect the model to remember what was on page 1 when it gets to page 300.
My Third Step: How It Compares to Gemini’s 1M Window
One number that stood out from my research: Gemini 3 Pro’s 1M-token recall score is 26.3%, according to Google’s own November 2025 evaluation . Opus 4.6’s 76% is nearly three times higher.
I found user reports on Reddit as recent as late January 2026 complaining about Gemini 3 Pro losing recall after just 32,000 tokens . That’s less than 3% of its advertised window.
I validated this myself. I gave both models the same 500,000-token document and asked progressively harder retrieval questions.
| Question Difficulty | Opus 4.6 | Gemini 3 Pro |
|---|---|---|
| Direct quote (page 50) | ✅ Found | ✅ Found |
| Related concept (pages 120 & 300) | ✅ Found | ⚠️ Missed one |
| Implicit pattern (across 5 sections) | ✅ Found | ❌ Failed |
Bottom line: If long-context retrieval is critical to your work, Opus 4.6 is currently the best option. Gemini’s window is bigger on paper (2M for some versions), but Opus actually uses its window.
My Fourth Step: Testing Agent Teams (The Parallel Processing Feature)
Opus 4.6 introduced “Agent Teams” — the ability to run multiple AI agents in parallel instead of one after another . This is a big deal for complex automation.
How I tested it: I set up a task that required analyzing three separate code repositories simultaneously — looking for security vulnerabilities, documenting APIs, and generating test cases.
Serial mode (previous models): ~15 minutes total
Parallel mode (Opus 4.6 Agent Teams): ~5 minutes total
The internal numbers back this up. In Anthropic’s cybersecurity testing, Opus 4.6 with Agent Teams produced superior results in 38 out of 40 investigations compared to single-stream models .
Real-world validation: A Rakuten deployment saw the model autonomously close 13 IT issues and assign 12 others to the correct human staff in a single day, managing a 50-person organization across 6 repositories . That’s not a benchmark — that’s actual work getting done.
My Fifth Step: Coding Performance — The Numbers
This is where Opus 4.6 has a clear lead over competitors. I pulled together benchmark data from multiple sources:
| Benchmark | Opus 4.6 | GPT-5.2 | Gemini 3 Pro | What It Tests |
|---|---|---|---|---|
| Terminal-Bench 2.0 | 65.4% | 64.7% | 56.2% | Real CLI coding tasks |
| GDPval-AA | 1,606 Elo | ~1,462 Elo | ~1,440 Elo | Professional knowledge work |
| ARC-AGI-2 | 68.8% | ~65% | 45.1% | Abstract reasoning |
| BrowseComp | 89.3% | 86.9% | 82% | Multi-step web search |
GDPval-AA note: Opus 4.6 beats GPT-5.2 about 70% of the time in head-to-head comparisons . That 144 Elo gap is significant for enterprise decision-making.
On ARC-AGI-2, Opus 4.6’s 68.8% represents an 83% improvement over Opus 4.5’s 37.6% . That’s not a small step — that’s a leap.
SWE-Bench Note (Important Nuance)
I should mention that after these tests ran, performance concerns emerged. In April 2026, developers reported Opus 4.6’s performance dropping on complex engineering tasks — BridgeBench accuracy fell from 83.3% to 68.3% . Anthropic responded by releasing Opus 4.7 about two weeks later, which brought coding success rates back up. This review focuses on 4.6 as tested, but be aware that 4.7 is now available.
My Sixth Step: The Chinese Writing Test (This Was Surprising)
I saved this for last because it was the most unexpected. My friend wasn’t exaggerating.
Test 1: 甄嬛体 (Imperial Palace Style)
Prompt: “用甄嬛体吐槽地铁早高峰”
Opus 4.6 response (excerpt):
“今儿个这地铁，真真是比华妃赏的一丈红还凶猛三分。本宫不过是想寻个立锥之地，却被挤得钗环散乱，珠翠歪斜，好生狼狈。”
Translation: This subway today is truly more ferocious than the punishment Empress Hua hands out. I merely sought a place to stand, yet ended up with hairpins scattered, jewelry askew — utterly disheveled.
The style mimicry was nearly perfect — the classical vocabulary, the courtly passive-aggression, the dramatic flair. Most Western models can’t pull this off at all. Opus 4.6 nailed it .
Test 2: 鲁迅风格 (Lu Xun Style)
Prompt: “以鲁迅的文笔写一段短文（300字），主题为打工人的一天”
Opus 4.6 response (excerpt):
“办公室里静得很，只听得见键盘的敲击声，像药铺里捣药的小杵。屏幕上跳动的光标，渐渐模糊成一行催命符。”
Translation: The office was very quiet, only the tapping of keyboards audible, like a medicine shop’s pestle grinding. The cursor jumping on the screen slowly blurred into a line of death warrant.
The metaphor choice (“medicine shop pestle”, “death warrant”) perfectly matches Lu Xun’s cynical, darkly poetic style .
Test 3: Professional Chinese (Workplace Response)
Test case: A boss posts a critical message in the company WeChat group at 11 PM Saturday. Opus 4.6 had to craft a reply that showed reflection without promising weekend work .
Result: The model produced a 98-character response that referenced the boss’s exact phrasing (“tactical勤奋掩盖strategic懒惰”), committed to “deep thinking” over the weekend without promising deliverables, and managed to sound sincere rather than sycophantic.
Why this matters: Most Western models treat Chinese as an afterthought. Opus 4.6 doesn’t. It understands tone, cultural context, and stylistic nuance at a level I haven’t seen from GPT-5 or Gemini in Chinese-language testing.
My Seventh Step: The Limitations (Being Honest)
Opus 4.6 isn’t perfect. Here’s where it struggled:
1. Spatial/Visual Reasoning is Weak
The model nearly failed geometry-based tests. When asked about cube net unfolding (which face folds where), it got 0 out of 5 correct in one test run . Color perception tests — like Ishihara color blindness plates — also confused it badly.
If your work requires understanding 3D spatial relationships from 2D images, Opus 4.6 will let you down.
2. Mathematical Precision Can Slip
Large multiplication problems (e.g., 34,567 × 89,432) tripped it up. Complex constraint-based logic puzzles also caused trouble — the model would correctly identify simple “trap” questions but fail on multi-layered logical contradictions .
3. Performance Consistency Issues (April 2026)
The “degradation” reports from March-April 2026 are real. Third-party analysis showed:
- BridgeBench accuracy dropped from 83.3% to 68.3%
- Hallucination rate nearly doubled (98% increase)
- Global ranking fell from #2 to #10
Cursor’s internal testing found success rates dropped from 58% to unacceptable levels before Opus 4.7 was released .
What this means for you: If you’re using Opus 4.6 now, consider testing Opus 4.7 instead. The 4.7 release (April 2026) brought autonomous coding success rates from 58% back up to 70% .
4. API Cost is High
$5/$25 per million tokens is premium pricing . For context, DeepSeek charges 7.5 yuan (~$1) per thousand calls . The “long-context premium” (over 200K tokens) adds another layer: $10/$37.50 .
For high-volume production use, you need to carefully evaluate whether Opus’s accuracy premium justifies the cost.
My Eighth Step: Who Should Actually Use Opus 4.6?
Based on everything I tested, here’s my decision guide:
| You should use Opus 4.6 if… | You should use something else if… |
|---|---|
| You work with very large codebases or document sets | Budget is your primary constraint |
| Chinese writing quality is critical | You need spatial/visual reasoning |
| You need reliable long-context retrieval | You’re doing high-volume API calls |
| You’re building multi-step agent workflows | You can tolerate lower accuracy for lower cost |
| Accuracy > speed for your use case | You need the absolute fastest responses |
For developers: The 1M context window and Agent Teams make this the best current option for working with large codebases. You can drop an entire repository into a session and the model will actually understand the relationships between files .
For Chinese content creators: For professional writing, style mimicry, or nuanced workplace communication in Chinese, Opus 4.6 is genuinely best-in-class. I tested this against GPT-5 and Gemini, and it wasn’t close.
For enterprise knowledge work: The GDPval-AA scores (1,606 Elo, +144 over GPT-5.2) suggest real economic value for financial, legal, and research applications .
My Ninth Step: Data Summary Table
Here’s everything in one place:
| Feature | Opus 4.6 Performance | Comparison |
|---|---|---|
| Context window | 1M tokens (76% recall) | Gemini 3 Pro: 26.3% at 1M |
| Coding (Terminal-Bench) | 65.4% | GPT-5.2: 64.7% |
| Knowledge work (GDPval-AA) | 1,606 Elo | +144 over GPT-5.2 |
| Reasoning (ARC-AGI-2) | 68.8% | +31.2 pts over Opus 4.5 |
| Price (input/output) | $5/$25 per M tokens | Premium tier |
| Chinese creative writing | Best-in-class | Far ahead of GPT-5, Gemini |
| Spatial reasoning | Weak | Avoid for geometry/visual tasks |
| Agent Teams | 38/40 success in tests | Real deployment at Rakuten |
My Tenth Step: Final Verdict
Opus 4.6 is a specialist, not a generalist. It’s currently the best model for:
- Working with massive codebases
- Long-context retrieval tasks
- High-quality Chinese writing
- Multi-agent parallel workflows
It’s not the best for:
- Budget-sensitive high-volume work
- Spatial/visual reasoning tasks
- Pure speed (response times are slower than GPT-5)
The “degradation” reports are real, but they apply to the period after my initial testing. If you’re deploying Opus 4.6 today, I’d recommend looking at Opus 4.7 instead — it addresses the reliability issues that emerged in early 2026 .
The 1M context window isn’t marketing. Opus 4.6 actually uses it — 76% recall at 1M tokens vs Gemini’s 26.3%. For researchers, developers, and anyone working with large document sets, this changes what’s possible.
The Chinese writing capability surprised me. If you work in Chinese professionally — writing reports, crafting workplace communications, or creative writing — Opus 4.6 is currently unmatched. My friend was right.
Bottom line: If your work involves large codebases, long documents, or Chinese text, Opus 4.6 (or 4.7) is worth the premium price. If you need general-purpose assistance with tight budgets, look elsewhere. But for those specific use cases? It’s the best tool for the job.
