What It Is
Claude Opus 4.7 is Anthropic's latest flagship model, released April 16, 2026. It replaces Opus 4.6 with notable improvements in coding, vision, and long-task handling—but comes with a hidden price increase that's sparking major community backlash.
Bottom line: Opus 4.7 clears 70% on CursorBench (vs 58% for Opus 4.6), sees images at 3x resolution, and handles multi-hour agentic workflows. But the new tokenizer silently increases your bill by 35-50%.
Technical Specs
| Parameter | Opus 4.7 | Opus 4.6 | Change |
|---|---|---|---|
| Context Window | 1M tokens | 1M tokens | Same |
| Max Output | 128K tokens | 128K tokens | Same |
| Max Image Resolution | 2576px / 3.75MP | 1568px / 1.15MP | +3.26x |
| Official Pricing | $15/$75 per M | $15/$75 per M | Same |
| Effort Levels | low/med/high/xhigh/max | low/med/high/max | NEW: xhigh |
| Thinking Mode | Adaptive ONLY | Extended + Adaptive | BREAKING |
| Tokenizer | NEW (1.0-1.35x more tokens) | Previous | HIDDEN COST |
Vision Improvements
The 3x resolution claim is real: Opus 4.7 processes images up to 3.75 megapixels (2576px long side), versus 1.15MP for previous models. This unlocks:
- Computer use agents reading full screen captures
- Dense document understanding without downsampling
- Pixel-perfect coordinate mapping (no scale-factor math)
- Professional diagram and chart interpretation
Benchmarks vs Competitors
Coding Benchmarks
| Benchmark | Opus 4.7 | Opus 4.6 | GPT-5.4 | Mythos |
|---|---|---|---|---|
| SWE-bench Pro | 64.3% | 53.4% | 57.7% | ~80% |
| SWE-bench Verified | 87.6% | 80.8% | - | 93.9% |
| CursorBench | 70% | 58% | - | - |
| Terminal-Bench 2.0 | 69.4% | 65.4% | - | - |
| GDPVal-AA (Elo) | 1753 | 1619 | 1674 | - |
Knowledge Benchmarks
| Benchmark | Opus 4.7 | Opus 4.6 |
|---|---|---|
| GPQA Diamond | 82.1% | 69.1% |
| ChartQA Reasoning | 91.0% | 84.7% |
Effort Level Scaling
| Effort | Opus 4.7 | Opus 4.6 |
|---|---|---|
| low | ~52% | ~45% |
| medium | ~60% | ~52% |
| high | ~65% | ~58% |
| xhigh (NEW) | ~71% | N/A |
| max | ~73% | ~62% |
Critical finding: Opus 4.7 at xhigh (~71%) beats Opus 4.6 at max (~62%) with fewer tokens.
The Tokenizer Tax Controversy
Here's what Anthropic's announcement doesn't mention:
The new tokenizer maps the same content to 1.35x more tokens on average. Same $/token pricing means 35-50% higher actual cost per request.
Reddit thread "Opus 4.7 is 50% more expensive with context regression?!" has 176 upvotes and 75 angry comments. Key complaints:
- Hidden price increase: Workloads that were cost-sensitive on 4.6 become unsustainable on 4.7.
- Context regression: MRCR v2 benchmark shows dramatic drops—91.9% → 59.2% at 256K context, 78.3% → 32.2% at 1M context.
- API breaking changes: Extended thinking budgets removed entirely, confusing developers.
Real Cost Impact
| Factor | Effect |
|---|---|
| New tokenizer | 1.0x-1.35x more tokens per prompt |
| xhigh default for coding | More output tokens |
| Combined | 35-50% bill increase |
API Breaking Changes
Previous system (Opus 4.6):
thinking={"type": "enabled", "budget_tokens": 10000} # This worked
New system (Opus 4.7):
thinking={"type": "enabled"} # No budget_tokens! Returns 400 error if you try
Other changes:
- Reasoning summary NOT displayed by default—must add
"display": "summarized" - Sampling parameters (temperature, top_p) no longer supported
- Adaptive Thinking is the only mode—no fixed budgets
Community Sentiment
Positive
- Cursor CTO: "State-of-the-art model on the market" for coding
- Hex: "Strongest model Hex has evaluated—correctly reports missing data instead of hallucinating"
- Devin: "Works coherently for hours, pushes through hard problems"
Negative
- simonw (notable developer): "The adaptive thinking thing is very confusing, especially having written code against previous thinking budget modes"
- Reddit user: "First they reduced Opus 4.6 tokens, we could all feel it. Now 4.7 has a tokenizer that costs 50% more. My subscription ended yesterday—I'm not even trying this."
- Gizmodo: "Anthropic Releases Claude Opus 4.7 to Remind Everyone How Great Mythos Is"
Mythos Comparison
Anthropic's own statement: "Opus 4.7 is less broadly capable than our most powerful model, Claude Mythos Preview."
| Metric | Opus 4.7 | Mythos Preview |
|---|---|---|
| SWE-bench Verified | 87.6% | 93.9% |
| Cybench | ~67% | 100% pass@1 |
| Availability | Generally Available | Restricted (Glasswing) |
| Pricing | $15/$75 | $25/$125 |
Migration Decision Matrix
| Strategy | When to Use |
|---|---|
| Migrate to 4.7 | Best coding/vision/agentic capability; can absorb 35-50% cost increase |
| Route between 4.6/4.7/Sonnet | Cost-sensitive workloads; need per-task modeling |
| Stay on Opus 4.6 | Cost-critical apps; proven reliability needed |
Availability
- Claude.ai
- Claude API (
claude-opus-4-7) - Amazon Bedrock
- Google Cloud Vertex AI
- Microsoft Foundry
The Takeaway
Opus 4.7 is a real capability upgrade—70% on CursorBench, 3x vision resolution, sustained multi-hour reasoning. But the tokenizer change is a hidden price hike that's generating significant backlash. If your workload was cost-optimized for 4.6, you'll need to re-audit your prompts and model routing strategy.