Read more...

🔬 SPECIAL COVERAGE: ARTIFICIAL 🔬 INTELLIGENCE

ANTHROPIC LANZA

CLAUDE OPUS 4.7

📅 April 16, 2026  •  ⏱️ Reading time: 9–11 minutes  •  🏢 Anthropic / San Francisco, CA

🌐 Coverage: Global AI / programming / tech industry  •  📰 Technology & Future

Opus 4.7 is the generally available version of Anthropic's flagship model. According to the company itself, this is "a marked improvement over Opus 4.6 in advanced software engineering, with particular gains in the most difficult tasks." Users report being able to delegate their most complex programming jobs—those that previously required constant supervision—to Opus 4.7 with confidence.

Two main axes define the new version: the ability to verify your own answers before reporting them ("self-checking") and an exponential leap in visual resolution. The model can now process images up to 2,576 pixels on the long edge, which is more than three times the capacity of its predecessor (1.15 megapixels vs. 3.75 megapixels). This opens doors to the interpretation of technical diagrams, dense documents and chemical structures with a precision previously unthinkable.

The numbers speak loudly in the territory that matters most to Anthropic: autonomous programming. In SWE-bench Pro – the benchmark that measures the ability to resolve real issues in open source repositories, in multiple languages – Opus 4.7 jumps from 53.4% to 64.3%, beating GPT-5.4 (57.7%) and Gemini 3.1 Pro (54.2%) by a significant margin. That's a jump of more than 10 percentage points in two months.

↑ Better — ↓ Backward from previous version. Sources: Anthropic System Card, VentureBeat, The Next Web, Vellum AI — April 2026.

🛑 The exception: BrowseComp drops 4 points

Not everything is the green light. BrowseComp, which measures web search and multipage synthesis capability, falls from 83.7% to 79.3%, falling 10 points behind GPT-5.4 (89.3%) and almost 6 behind Gemini 3.1 Pro (85.9%). Teams using web-intensive research agents should take note: Opus 4.7 isn't the first choice for that specific workflow.

1️⃣ Automatic verification of results (self-checking)

Opus 4.7 introduces the ability to review your own responses before reporting them. In complex coding workflows, this means that the model catches logical errors during the planning phase, before committing to an incorrect implementation. Early access users confirm this: "Claude Opus 4.7 catches its own logical flaws during the planning phase and accelerates execution, well above Claude's previous models," a fintech platform recounted in Anthropic's launch post.

2️⃣ "xhigh" effort level and task budgets

Opus 4.7 introduces a new level of reasoning called xhigh, located between high and max. This gives developers granular control over the relationship between depth of reasoning and latency. Claude Code now uses xhigh by default on all subscription plans. In parallel, task budgets in public beta allow developers to set a token consumption ceiling for agent loops, avoiding bill surprises during long debugging sessions.

Anthropic — Official Release Post, 4/16/2026

3️⃣ 3 vision× more resolution and scientific imaging

The visual leap is the most notable in this version: the maximum image resolution goes from 1.15 to 3.75 megapixels. In CharXiv, the benchmark for the interpretation of scientific figures, Opus 4.7 improves by 13 points. On XBOW visual acuity rises from 54.5% to 98.5%. Desktop environment coordinates now map 1:1 to pixels, eliminating the scaling error that limited accuracy in OSWorld.

4️⃣ Persistent memory in long sessions

The model substantially improves the use of file system-based memory. You can recall important notes through long multi-session work sessions and use them to move forward on new tasks with less initial context. For development workflows that span hours or days, this represents a concrete operational advantage.

5️⃣ Command /ultrareview and auto mode for Max plan

Claude Code adds /ultrareview, a command that runs a dedicated review session that reads all changes and flags what a careful human reviewer would detect. In turn, the "auto mode" – previously exclusive to Teams/Enterprise/API – is now available to Max plan subscribers, reducing interruptions in longer tasks.

The release of Opus 4.7 closes a fierce race. Between February 5 and April 16, 2026, the four flagship models of the three great AI powers hit the market in less than ten weeks. The convergence is so pronounced that in university reasoning (GPQA Diamond) the three public models are within 0.2 percentage points of each other. The benchmark has saturated on the border.

According to VentureBeat, Opus 4.7 outperforms GPT-5.4 in 7 of 11 directly comparable benchmarks, but the advantage is not overwhelming. Anthropic itself admits that the model it really aspires to match is its own Claude Mythos Preview, an even more powerful version kept out of public access for security reasons.

💰 The Price Map: Gemini's Pressure

The front where Anthropic doesn't win is price. Gemini 3.1 Pro charges $2 per million tokens input and $12 tokens output, compared to $5 and $25 for Opus 4.7. For high-volume workloads or low code accuracy requirements, the 2.5× difference in price can be decisive. However, analysts and developers conclude that for agentic engineering—where Opus 4.7 saves hours of human labor—the extra cost is justified.

There's one elephant that Anthropic can't hide: Claude Mythos Preview. The company has introduced Opus 4.7 as its most powerful model for general availability, but openly admits that it's not the most advanced model it's ever built. Mythos Preview, described as a system of extraordinarily powerful cyber capabilities, was released the previous week to just a select group of technology and cybersecurity companies under Project Glasswing.

"We released Opus 4.7 with safeguards that automatically detect and block requests that indicate prohibited or high-risk cybersecurity uses," Anthropic wrote. "What we learn from the real-world deployment of these safeguards will help us work toward our ultimate goal of a broad launch of the Mythos-class models."

Anthropic — Post oficial, 16/4/2026

For security professionals who wish to use Opus 4.7 for legitimate purposes—vulnerability research, pen testing, red-teaming—Anthropic launches the Cyber Verification Program, a formal accreditation process that gives access to expanded capabilities under supervision.

The launch comes at the time of the greatest commercial momentum in Anthropic's history. Claude's traffic grew approximately 5 times in the last year. The company reached a valuation of USD 380,000 million in its Series G in February 2026. And at the time of this news, its annualized revenue rate (ARR) had climbed to $30 billion by April 2026, driven primarily by enterprise adoption and the success of Claude Code. Eight of the world's ten largest companies by capitalization are now Claude's clients.

Investment reports suggest that venture capital firms are offering valuations of up to $800 billion — more than double February's Series G valuation — though the company has not confirmed any new rounds. The context is relevant: in a market that converges technically, the battle is fought in distribution, ecosystem of tools and business trust.

✅ Clear strengths of Opus 4.7

►    World's best public model for autonomous programming (SWE-bench Pro: 64.3%).

►    Self-checking: reduces errors without human intervention in long workflows.

►    More powerful 3× vision: opens up new applications in science, patents, and technical documents.

►    Same price as Opus 4.6: more performance at the same cost for those who already use it.

►    Multi-cloud availability: AWS Bedrock, Google Vertex AI, Microsoft Foundry — frictionless integration.

⚠️ Points of attention and open-ended questions

►    BrowseComp drops 4 points: not the optimal choice for intensive web research agents.

►    New tokenizer: the same input can generate up to 1.35× more tokens — price neutrality can be misleading for long prompts.

►    Legacy prompts: The increased literality of the model can break down backwards-optimized instructions.

►    Mythos as a ceiling: the market knows that there is a more powerful model. The tension between security and access will be the debate next quarter.

►    Pricing competition: Gemini 3.1 Pro at $2/$12 pushes on high-volume loads. MiniMax M2.5 (open-weight) at $0.30/$1.20 threatens in the cost segment.

Recommended Metadata

Table of strategic keywords

🛠️ Advanced SEO/AEO Strategies 2026

►TechArticle Schema + SoftwareApplication Schema: Mark the release as a technology article and the model as a software application (applicationCategory: "ArtificialIntelligence").

►AEO — Answer Engine Optimization: Include direct response paragraph in the first 100 words to capture featured snippets in Google SGE, Perplexity, and ChatGPT Search.

►E-E-A-T: credit the author with AI expertise; cite primary sources (Anthropic, VentureBeat, The Next Web, Vellum AI) with outbound links of authority.

►Freshness signal: update the article with each Anthropic release; keep the modification date updated in the Schema and in the tag <meta http-equiv="last-modified">.

►Semantic SEO: cover the entire topic cluster — "LLMs", "agentic AI", "AI programming", "LLM benchmarks", "Anthropic AI security" — as interconnected satellite pages.

►Core Web Vitals: LCP < 2.5 s, CLS < 0.1, INP < 200 ms — a must-have for high-traffic articles in competitive tech niche.

►Featured Snippets: definiciones breves (40–60 palabras) para términos como “SWE-bench”, “xhigh effort”, “task budgets” y “Cyber Verification Program”.

►Internal linking: link from GPT-5.4, Gemini 3.1 Pro, Claude Code, and Anthropic coverages to consolidate thematic authority in the domain.

📈 FAQ Schema — Frequently Asked Questions

❓ When was Claude Opus 4.7 released?

Claude Opus 4.7 was released on April 16, 2026 by Anthropic, exactly ten weeks after Claude Opus 4.6 (released on February 5, 2026). It is available from the day of its announcement on all Claude plans and in the API.

❓ How much does Claude Opus 4.7 cost?

The API price remains the same as Opus 4.6: $5 per million input tokens and $25 per million output tokens. The Batch API offers a 50% discount. The prompt cache can reduce the cost of entry by up to 90%.

❓ How does Claude Opus 4.7 beat GPT-5.4?

Opus 4.7 outperforms GPT-5.4 primarily in autonomous programming (SWE-bench Pro: 64.3% vs 57.7%), professional knowledge work (GDPVal-AA Elo 1753 vs 1674), and autonomous desktop use (OSWorld: 78% vs 73.1%). GPT-5.4 maintains the advantage in agentive web search (BrowseComp: 89.3% vs 79.3%).

❓ What is Claude Mythos and why is he not available to everyone?

Claude Mythos Preview is Anthropic's most advanced model, with cybersecurity capabilities far superior to Opus 4.7. It is restricted to a select group of technology and security companies because of the dual-use risks it poses, under Project Glasswing.

►    Anthropic — “Introducing Claude Opus 4.7” (post oficial, anthropic.com/news) — 16/4/2026

►    Axios — “AnthropicReleases Claude Opus 4.7, concedes it trails unreleased Mythos” — 16/4/2026

►    CNBC — “Anthropicrolls out Claude Opus 4.7, an AI model that is less risky than Mythos” — 16/4/2026

►    VentureBeat — “AnthropicReleases Claude Opus 4.7, narrowly retaking lead” — 16/4/2026

►    The Next Web — “Claude Opus 4.7 leads on SWE-bench and agentic reasoning” — 16/4/2026

►    Vellum AI — “Claude Opus 4.7 Benchmarks Explained” — 17/4/2026

►    9to5Mac — “AnthropicReveals New Opus 4.7 Model” — 16/4/2026

►    9to5Mac — “AnthropicLaunches Claude Design following Opus 4.7” — 17/4/2026

►    Nerd Level Tech — “Claude Opus 4.7: Benchmarks, Features & Pricing” — 17/4/2026

►    Build Fast With AI — “Claude Opus 4.7: Full Review, Benchmarks & Features” — 17/4/2026

――――――――――

🤖 Professional journalistic coverage with advanced 🤖 SEO optimization

Bell Ville, Córdoba, Argentina — April 18, 2026

Close