On April 16, 2026, Anthropic quietly dropped Claude Opus 4.7 — and the benchmarks are wild. After months of GPT-5.4 dominating the leaderboards, Anthropic has narrowly retaken the crown for the most powerful generally available large language model. Here's everything that changed, how it compares to its predecessor Opus 4.6, and how it stacks up against GPT-5.4 and Gemini 3.1 Pro.
What Is Claude Opus 4.7?
Claude Opus 4.7 is Anthropic's new flagship model, released on April 16, 2026. It's the direct successor to Opus 4.6 and represents a major step forward in coding performance, scientific reasoning, agentic tool use, and vision capabilities. It's available today across all major cloud platforms — Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry — and notably, pricing stayed flat at $5 per million input tokens and $25 per million output tokens.
What makes this release interesting isn't just that it's an incremental bump — it's that Anthropic managed to push several benchmark numbers past competitors that were previously considered unbeatable.
Opus 4.7 vs Opus 4.6 — What's New?
The jump from 4.6 to 4.7 isn't a minor patch. There are substantial improvements across almost every metric.
Coding Performance
- SWE-bench Verified: 80.8% → 87.6%
- SWE-bench Pro: 53.4% → 64.3%
- CursorBench: 58% → 70%
That's a massive jump — particularly on SWE-bench Pro, where Opus 4.7 gained over 10 percentage points. For developers using Claude Code or building coding agents, this is the single biggest practical upgrade.
Vision — Bigger Images, More Detail
Opus 4.7 accepts images up to 2,576 pixels on the long edge — roughly 3.75 megapixels. Opus 4.6 topped out at about 1.15 megapixels. That's more than triple the resolution, which means better OCR, better handling of dense UI screenshots, and more accurate analysis of high-resolution diagrams and documents.
New "Effort Level" Setting
Anthropic added a new intermediate effort tier that sits between high and max, giving you finer control over the reasoning-latency tradeoff. The company recommends starting with high or xhigh for coding and agentic use cases.
Task Budgets
This is one of the more underrated additions. Task budgets let you give Claude a rough token estimate for an entire agentic loop — including thinking, tool calls, tool results, and final output. The model sees a running countdown and uses it to prioritize work, gracefully wrapping up as the budget is consumed. For long-running agents, this means fewer runaway token bills and more predictable behavior.
Claude Code /ultrareview
Inside the Claude Code environment, 4.7 introduces a new /ultrareview command. Unlike standard code review that looks for syntax errors and style issues, /ultrareview simulates a senior human reviewer — flagging subtle design flaws, logic gaps, and architectural concerns that most linters would miss.
Cybersecurity Safeguards
Opus 4.7 is the first Claude model shipping with automated detection and blocking for prohibited cybersecurity use cases. This is notable because — even despite Anthropic's recent security incidents, including the Claude Code source code leak earlier this year — the company is clearly doubling down on responsible AI deployment from the model side.
Opus 4.7 vs GPT-5.4
GPT-5.4 launched in early March 2026 and held the top spot on most leaderboards for about six weeks. Opus 4.7 has now narrowly overtaken it on directly comparable benchmarks — but the race is closer than the headlines suggest.
"On directly comparable benchmarks, Opus 4.7 only leads GPT-5.4 by 7-4, indicating a tight competitive race."
Where Opus 4.7 Wins
- Coding (SWE-bench Pro): 64.3% vs GPT-5.4's 57.7%
- SWE-bench Verified: 87.6% vs GPT-5.4's lower score
- Scientific reasoning (GPQA Diamond): 94.2%
- Agentic tool invocation and long-horizon coding agents
- Financial analysis and complex instruction following
Where GPT-5.4 Still Leads
- Agentic search: 89.3% vs Opus 4.7's 79.3% — a significant gap
- Multilingual Q&A
- Raw terminal-based coding (without full agent framework)
- Pricing — GPT-5.4 input pricing is half that of Opus 4.7
Pricing Reality Check
Opus 4.7 is priced at $5/M input and $25/M output. That's 2x the input price and 67% higher on output compared to GPT-5.4. If your use case can tolerate GPT-5.4's slightly lower benchmarks, you're looking at meaningful cost savings at scale.
Opus 4.7 vs Gemini 3.1 Pro
Google released Gemini 3.1 Pro back in February 2026, and it's been a strong contender — especially in vision and long-context workloads.
Where Opus 4.7 Wins
- SWE-bench Verified: 87.6% vs Gemini 3.1 Pro's 80.6%
- SWE-bench Pro: 64.3% vs Gemini's 54.2%
- Agentic workflows requiring strict instruction following
- Scientific reasoning benchmarks
Where Gemini 3.1 Pro Still Wins
- Context length — Gemini's native context window is still larger
- Vision on very high-resolution documents
- Pricing — $1.25/M input and $5/M output, roughly 25% the cost of Opus 4.7
For teams processing huge document batches or doing heavy vision work on architectural drawings, medical scans, or other ultra-high-res assets, Gemini 3.1 Pro remains the price-to-performance sweet spot.
Head-to-Head Benchmark Summary
Here's a quick side-by-side of where each model ranks across the core benchmarks:
| Benchmark |
Claude Opus 4.7
|
GPT-5.4
|
Gemini 3.1 Pro
|
|---|---|---|---|
| SWE-bench Verified | 87.6% | — | 80.6% |
| SWE-bench Pro | 64.3% | 57.7% | 54.2% |
| Agentic Search | 79.3% | 89.3% | — |
| GPQA Diamond (Science) | 94.2% | — | — |
| Vision Resolution | 3.75 MP | — | — |
| Input Pricing (per 1M tokens) | $5.00 | $2.50 | $1.25 |
| Output Pricing (per 1M tokens) | $25.00 | $15.00 | $5.00 |
Sources: VentureBeat, Vellum AI, Eden AI, Apiyi Benchmark Analysis. Em dashes indicate published scores not available for direct comparison.
Which Model Should You Actually Use?
Benchmarks are useful, but they're not the full story. Here's the practical breakdown:
Use Claude Opus 4.7 if…
- You're building reliable coding agents or multi-step agentic workflows
- Strict instruction following matters more than raw speed or cost
- You need the absolute best SWE-bench performance for complex codebases
- You're already using Claude Code and want /ultrareview
- Budget isn't the primary constraint
Use GPT-5.4 if…
- You want one model that handles coding, docs, reasoning, and business workflows without heavy optimization
- Agentic search is a core part of your application
- Multilingual support is critical
- You need the best balance of capability and cost
Use Gemini 3.1 Pro if…
- You're processing huge context windows (millions of tokens)
- Vision work on ultra-high-resolution documents is a core workload
- Cost matters — Gemini 3.1 Pro is significantly cheaper than either competitor
- You're already in Google's ecosystem
The Bigger Picture
What's most interesting about the Opus 4.7 release isn't any single benchmark number — it's that the race between Anthropic, OpenAI, and Google is now down to margins. A few months ago, benchmark gaps were double digits. Today, we're arguing over 7% differences on a handful of evaluations.
For developers and businesses, this means the choice of frontier model increasingly comes down to use case, pricing, and ecosystem fit — not raw capability. All three models are capable enough to handle most real-world tasks. The question is where you're willing to pay the premium, where you can afford to save a dollar, and which API aligns best with your existing tooling.
Anthropic's recent operational missteps are worth noting too. Even despite the Claude Code source code leak back in March, the company has continued shipping strong models and expanding enterprise availability. For teams evaluating AI vendors, the real question is whether Anthropic's safety-first branding still holds up — but on the model capability side, Opus 4.7 is a serious piece of work.
Bottom Line
Claude Opus 4.7 is a legitimate leap forward from 4.6 — especially for coding. It has retaken the top spot across most benchmarks, though not by dominant margins. If you're building coding agents or complex workflows, it's worth evaluating. If you're optimizing for cost or need massive context windows, GPT-5.4 or Gemini 3.1 Pro may still be the better call. The good news for everyone: the competition is pushing these models forward at a pace that makes even six-month-old releases feel outdated.
Sources: VentureBeat, Vellum AI, Claude API Docs, GitHub Changelog, Inc., Eden AI
Ready to Build Something?
Let's talk about your project and see how we can help.
Schedule a Consultation