BIG TECH

Anthropic buys time with Opus 4.7 before Mythos

Anthropic may be holding back Mythos, but it doesn’t want its users to go hungry. 

On Thursday, the company released Opus 4.7, the latest addition to the highest tier of its Claude model family. Anthropic said the model offers “notable improvement” in software engineering tasks compared to Opus 4.6, allowing users to “hand off their hardest coding work” and long-running tasks with greater confidence. 

Anthropic also says Opus 4.7 can see images with substantially better resolution, produce more “tasteful and creative” outputs for professional tasks and produce higher-quality documents, slides and designs. 

• The company conceded that Opus 4.7 is less “broadly capable” than Mythos, but still surpasses its predecessors and rivals on several benchmarks, including agentic coding, multidisciplinary reasoning, tool and computer use, and visual reasoning. 

• Anthropic also noted that Opus 4.7 shows considerable gains in instruction following, memory, and multimodal support.

• It also scores roughly the same on safety and alignment benchmarks for behaviors like deception, sycophancy, and cooperation with misuse. 

• As for price, the company is offering Opus 4.7 at the same cost per token as 4.6, at $5 per million input tokens and $25 per million output tokens. The model is currently available in all Claude products and API, as well as Amazon Bedrock, Google Cloud’s Vertex AI, and Microsoft Foundry. 

The company tested the model in preview with several companies, with several reporting gains in efficiency, reliability and speed. In a statement, Box’s Head of AI, Yashodha Bhavnani, told The Deep View that the model was able to do significantly more with less, performing fewer tool calls and offering lower latency for “enhancements that will help enterprises move faster and scale more affordably.”

Along with quelling the appetite of customers chomping at the bit to try Mythos, Anthropic’s Opus 4.7 serves as a testbed for the cyber safeguards it is building as part of Project Glasswing, the company said. This model includes safeguards that automatically detect and block requests that signal prohibited or risky cybersecurity issues. 

“What we learn from the real-world deployment of these safeguards will help us work towards our eventual goal of a broad release of Mythos-class models,” Anthropic said in the announcement.

Anthropic has gotten the industry riled up with all of the chatter around Mythos. As the company figures out what to do with this extremely powerful (and compute-hungry) model, its archrival OpenAI has released GPT-5.4-Cyber and made it available to a wider audience than Mythos. While Anthropic says it intends to use Opus 4.7 to work up to Mythos, the company may be feeling the pressure to ship, especially as some users report switching to OpenAI’s Codex as they run up against Claude’s rapidly depleting quotas and uptime issues. But as users await Mythos, Opus 4.7 might feel like a consolation prize. 

TOGETHER WITH VIKTOR

The one hire every great company needs

With just one hire, marketing gets a board-ready Google Ads + Stripe revenue report, engineering gets a PR review cross-referenced with Linear tickets, and ops gets three vendor contracts summarized before the founder's morning coffee.

So, what is it? Viktor, the AI coworker that 7,000+ teams already rely on. It's not a tool, it's a coworker -- and you can try it for yourself right here.

Get Viktor free with $100 in credits →

PRODUCTS

Codex adds computer-use skills to chase Claude

OpenAI said its Codex coding assistant now has more than 3 million weekly developer users, and the company is introducing a series of upgrades intended to grow that figure and narrow Anthropic's lead in AI coding tools.

A major emphasis of Thursday's updates was expanding what Codex can do by providing access to more tools. For instance, the biggest Codex upgrade is its ability to run desktop apps in the background on your computer. This enables multiple agents to work in parallel without affecting other apps, a major win for developers.

“The first thing we went out to do is to teach Codex and the app to work across a much larger surface area instead of just call and response with some plugins,” said Andrew Ambrosino, Codex App Lead, during a media briefing. 

The Codex app now also has an in-app browser that works natively with the web, allowing developers to comment on the page and provide instructions to agents, according to the release. Codex can also now use gpt-image-1.5 to generate and edit images. 

OpenAI also launched 111 new Plugins for Codex, which combine skills, app integrations, and MCP servers to enable it to access more information and take more actions. The upgrades, as described in the blog post, also include: 

• Software cycle: Codex can address GitHub review comments, run multiple

• terminal tabs, and connect to remote devboxes over SSH in alpha 

• Sidebar: Files are accessible directly in the sidebar for PDF, spreadsheet, slides, and doc preview 

• Summary pane: Located in the sidebar, it allows users to track agent plans, sources, and artifacts

• Personalization: The app can reuse existing conversation threads for improved context, and a new memory feature, in preview, allows it to remember context from previous experiences, such as preferences and information accumulated over time

• Future work: It can schedule future work for itself and wake up automatically, potentially across days or weeks, targeting longer tasks 

• Suggestions: Using context, Codex can propose where to pick up the last project or how to start your workday

The updates are rolling out to all signed-in Codex desktop users. However, there are exceptions. Computer use is limited to macOS users, for now, and Personalization features will roll out to Enterprise, Edu, and EU users soon.

Less than a month ago, OpenAI rolled out sweeping organizational changes that signaled a decisive push into enterprise products. The pace and intentionality behind the latest Codex developments reinforce that commitment. The Claude Code desktop redesign, which launched this week, also focused on enabling developers to run parallel agents and tasks, reflecting the scale at which developers are using these AI tools. Ultimately, several of the Codex upgrades further narrow the gap with Claude Claude, underscoring that the race to capture the developer market keeps heating up.

TOGETHER WITH GRANOLA

The Deep View Team is Obsessed with Granola

We’re not talking about the food (although a few of us do have an unhealthy interest in that, too). We’re talking about the AI notepad that turns your chicken scratch into something elegant, organized, and downright beautiful.

Since day one, Granola has checked all the boxes we look for in a notetaking app, like… 

• Working seamlessly across our various  teams and workflows ✅

• Documenting, summarizing, and prioritizing all our discussions, so we can focus on what’s most useful and relevant ✅

• Having full context across all meetings for a holistic view of trends and patterns ✅

We estimate Granola saves us around 10 hours per week, per person – that’s no small feat. Try it for yourself and get your first month free with code THEDEEPVIEW.

RESEARCH

OpenAI initiates rise of science-first AI models

While all the AI model buzz is centered on coding right now, OpenAI just released its first frontier reasoning model built purely for scientists.

Called GPT-Rosalind, this life sciences model is purpose-built for lab workflows. It has an in-depth understanding of genomics, chemistry, and protein engineering, and it's especially designed to support research in drug discovery, translational medicine, and biology more generally.  

OpenAI believes this model will address key bottlenecks in the life sciences:

1. Data volume exceeds human capacity: Biology is an extremely data‑rich field where modern tools have outpaced humans’ ability to keep up.

2. Specialized knowledge across biology subfields: Deeper specialization has created silos that slow translation of insights across domains.

"On average, it takes roughly 10 to 15 years to go from target discovery to regulatory approval for new drugs in the United States, and only one in 10 drugs entering clinical trials are ultimately approved… We think there's a real opportunity to help researchers move faster through some of the most complex and time‑intensive parts of the scientific process,” the OpenAI team said in a briefing with journalists.  

• The new model is named after Rosalind Franklin, the scientist who helped discover the shape of DNA.

• It's better than other AI models at science tasks. On tests that measure understanding of proteins, DNA, and chemical reactions, GPT-Rosalind scored higher than OpenAI's other models. In one test with real scientists, the AI actually performed better than 95% of human experts at predicting how certain RNA sequences work.

• It comes with special tools for scientists. OpenAI is also releasing a "Life Sciences research plugin" that connects the AI to more than 50 scientific databases and tools, so researchers can do things like look up protein structures, search DNA sequences, and review scientific papers all in one place.

• Access is limited for safety reasons. Because this technology could be misused to create something harmful, OpenAI is granting access only to approved organizations conducting legitimate research, such as Amgen, Moderna, the Allen Institute, and Thermo Fisher Scientific.

“The opportunity we set out to achieve was for every scientist to become their own research team, to tackle more by having the complementary expert-level knowledge and skills at their fingertips,” the OpenAI team shared during its press briefing.

OpenAI's GPT-Rosalind represents further evidence that today's LLMs are moving away from a one-model-to-rule-them-all approach (the AGI path) toward specialized, domain-specific, and even task-specific models. Along with Claude Mythos and GPT-5.4-Cyber, GPT-Rosalind also represents a new phase of models in which the most powerful models aren't being made broadly available to everyone, because they could also be adapted by bad actors for nefarious purposes that could cause harm. This trend, along with personal AI agents, is emerging as one of AI's key turning points of 2026. 

LINKS

• Anthropic to expand London presence with new office for 800 staff 

• Voice actors rally against Hollywood’s push for AI dubbing

• UK launches $675 million sovereign AI fund for AI independence

• White House moves to give US agencies access to Anthropic Mythos

• AI fintech Slash Financial raises $100 million at $1.4 billion valuation

• Big Tech firms push for data center roll outs in EU after success in Spain

• Perplexity Personal Computer: Perplexity’s AI assistant is now available for subscribers on Mac.

• DeepL: The voice AI company unveiled its voice-to-voice translation suite and API for developers.

• Roblox Assistant: Roblox has upgraded its natural language AI tool for game development. 

• Nano Banana 2: Google’s image model now includes Gemini’s Personal Intelligence.

• Canva 2.0: Canva unveiled an AI-powered redesign, both of its architecture layers and intelligent workflows.

2026 Salary Report: U.S. vs Global hiring.

• Explore salary benchmarks for AI engineers, data scientists, product managers, and more.

• Discover top-tier talent from Latin America, Africa, and South East Asia—ready to hit the ground running.

• See how companies are saving up to 67% on salaries by hiring globally.

Download the report

(sponsored)

GAMES

POLL RESULTS

Do you use the desktop app for ChatGPT, Claude, and/or Gemini?

Yes (53%)
No (40%)
Other (7%)