[7min] Anthropic and OpenAI Release New Models as Software Stocks Lose $1T
Plus: Claude Opus 4.6 tops agentic benchmarks with 1M context window, GPT-5.3-Codex rated first 'High' cybersecurity risk, and ChatGPT's market share craters to 45%
Anthropic and OpenAI Release New Models as Software Stocks Lose $1T
February 06, 2026
This digest is AI-curated. LLMs can make mistakes. Always verify critical details.
Anthropic and OpenAI both dropped major releases today. Anthropic launched Claude Opus 4.6 with agent teams that can split tasks and execute in parallel, topping most agentic benchmarks and featuring a 1M token context window. OpenAI countered with GPT-5.3-Codex, its first model rated "High" cybersecurity risk, and Frontier, an enterprise platform for orchestrating AI coworkers across an organization's tech stack. To prove the concept, Anthropic had 16 agents build a working C compiler from scratch in hours. Wall Street reacted swiftly: software stocks shed over $1 trillion in market value this week as investors absorbed the implications. ServiceNow lost 25% in a month. Salesforce fell 40% from its high. JPMorgan called it "being sentenced before trial," but the rout reflects genuine uncertainty about which enterprise software survives when anyone can vibe code custom tools. Meanwhile, new data shows ChatGPT's chatbot market share has eroded from 69% to 45% as Gemini, Claude, and Grok gain ground. Separately, a $100 billion Nvidia-OpenAI deal appears to have evaporated, and Claude Opus 4.6 discovered over 500 zero-day vulnerabilities during pre-release testing — a reminder that more capable models cut both ways.
Anthropic releases Claude Opus 4.6 with agent teams, 1M token context window
Anthropic launched Claude Opus 4.6, its most powerful model, featuring a 1M token context window (up from the Opus tier's previous limit) and a new "agent teams" capability in Claude Code that lets multiple AI agents split and execute tasks in parallel. The model tops most agentic benchmarks, scoring 76% on MRCR v2 (vs. Sonnet 4.5's 18.5%) and 65.4% on Terminal-Bench 2.0.
New integrations put Claude directly inside Excel and PowerPoint as sidebar tools, and the API gains "Adaptive Thinking" and "Compaction" features to manage long contexts. Anthropic's system card notes Opus 4.6 is slightly more vulnerable to indirect prompt injections than its predecessor, a concern for agentic deployments. (source)
OpenAI launches GPT-5.3-Codex, its first model rated 'High' cybersecurity risk
OpenAI released GPT-5.3-Codex, a coding model that merges GPT-5.2-Codex's programming skills with GPT-5.2's reasoning while running 25% faster. It tops agentic benchmarks including Terminal-Bench 2.0 (beating Opus 4.6 by 12 points) and scores 64.7% on OSWorld, nearly double its predecessor's 38.2%.
OpenAI flagged this as its first model to receive a "High" cybersecurity risk classification and committed $10M in API credits for defensive security research. Early versions helped debug training runs and manage deployment, making Codex the first model to actively participate in its own development. (source)
OpenAI launches Frontier platform to manage enterprise 'AI coworkers'
OpenAI unveiled Frontier, an enterprise platform for building, deploying, and managing AI agents across an organization's tech stack. Each agent gets its own identity with scoped permissions, shared business context, and built-in evaluation loops modeled on employee onboarding and performance reviews.
Frontier works with agents from OpenAI, Google, Microsoft, Anthropic, and custom-built ones. HP, Oracle, State Farm, and Uber are early adopters. Pricing remains undisclosed. The launch positions OpenAI directly against Microsoft's Agent 365 and Salesforce's Agentforce in the race to control the enterprise agent orchestration layer. (source)
Software stocks lose $1T+ as Anthropic tools trigger AI disruption fears
Software stocks shed over $1 trillion in market cap this week after Anthropic's Cowork legal plugin and broader AI agent tools spooked investors. Legal-tech firms Thomson Reuters and RLEX dropped ~15% each; ServiceNow lost 25% in a month; Salesforce fell ~40% from its high. Oracle's Larry Ellison saw $49B wiped from his net worth YTD.
Short sellers netted $24B from the rout. JPMorgan analyst Toby Ogg wrote the sector is "now being sentenced before trial." The equal-weight S&P 500 hit a record high, suggesting the pain is concentrated in tech while broader markets rotate into cyclicals and industrials. (source)
Anthropic agent teams build C compiler from scratch: 16 agents, 100K lines of code
Anthropic demonstrated its new Claude Code "agent teams" feature by having 16 AI agents collaborate to build a working C compiler from scratch. The 100,000-line codebase compiles and successfully runs the Linux kernel, showcasing the potential of multi-agent parallel development.
A lead agent decomposes tasks and delegates to specialized sub-agents that work simultaneously. The compiler project took hours rather than the weeks or months a human team would need, though Anthropic noted human oversight remained critical for architectural decisions. (source)
ChatGPT market share drops from 69% to 45% as Gemini and Claude gain ground
New data from Apptopia shows ChatGPT's share of the generative AI chatbot market has eroded from 69% to 45%, driven by growth at Google's Gemini, xAI's Grok, and Anthropic's Claude. The shift coincides with OpenAI's pivot toward monetization through ads and enterprise products.
The report landed alongside Sam Altman's claim that ChatGPT has more users in Texas alone than Claude has nationwide, a defensive response to Anthropic's Super Bowl ad campaign mocking ChatGPT ads. Altman accused Anthropic of elitism for primarily serving paying subscribers. (source)
Claude Opus 4.6 discovers 500+ zero-day vulnerabilities in security audit
Anthropic's internal security evaluation revealed that Claude Opus 4.6 identified over 500 previously unknown zero-day vulnerabilities across widely-used open-source software. The discovery came during standard pre-release safety testing.
The findings highlight a dual-use tension: the same capabilities that make Opus 4.6 effective at code analysis and security research could also be exploited. Anthropic coordinated responsible disclosure with affected maintainers before publishing the results, and cited this as justification for its cautious deployment approach. (source)
Goldman Sachs deploys Anthropic's Claude to automate trade accounting
Goldman Sachs is using Anthropic's Claude to automate parts of its trade accounting workflow, marking one of the highest-profile enterprise AI deployments in finance. The system handles reconciliation and processing tasks that previously required manual review by accounting teams.
The deal underscores how frontier AI models are penetrating regulated industries where accuracy and auditability are paramount. Goldman joins a growing list of financial institutions integrating AI directly into core operations rather than just customer-facing tools. (source)
Meta completes pretraining 'Avocado,' its most capable AI model yet
Meta finished pretraining its new AI model codenamed "Avocado," which outperforms the best freely available base models in knowledge, visual perception, and multilingual performance even before post-training. An internal memo says it's 10x more efficient than Maverick and 100x more efficient than Behemoth.
The milestone marks a potential turnaround after Meta's rocky 2025, which saw Llama 4 delays, manipulated benchmarks, and Yann LeCun's departure. Reports suggest Meta may move away from open-source for Avocado, a significant shift from its Llama strategy. A visual model codenamed "Mango" is also in development. (source)
Nvidia-OpenAI $100B deal unravels, raising questions about circular AI economy
A widely-reported $100B deal between Nvidia and OpenAI appears to have evaporated. Jensen Huang privately told associates the agreement was "non-binding" and publicly said any investment would be "nothing like" $100B. Reuters reported OpenAI is "unsatisfied" with Nvidia's advanced chips and seeking alternatives.
Nvidia's stock dropped 10% this week on the news. Oracle, which is counting on a separate $300B cloud deal with OpenAI, rushed to reassure investors. The unraveling highlights concerns about circular AI investment where chipmakers fund AI companies that buy their own chips. (source)
Microsoft AI CEO Suleyman says vibe coding will make custom apps replace packaged software
Microsoft AI CEO Mustafa Suleyman argued that "vibe coding" will fundamentally reshape software by enabling anyone to build custom applications, making many packaged SaaS products obsolete. He predicted enterprises would increasingly build bespoke tools rather than buy licenses.
The comments amplified the software selloff narrative and drew pushback from incumbents. Suleyman framed it as a natural evolution: "The app was the interface for the internet era. The agent is the interface for the AI era." His remarks align with Microsoft's own pivot toward AI-powered development tools like GitHub Copilot and Codex. (source)
Fundamental raises $255M at $1.2B valuation for Large Tabular Model
AI lab Fundamental emerged from stealth with $255M in funding at a $1.2B valuation to build Nexus, a "Large Tabular Model" for enterprise structured data. Unlike transformer-based LLMs, Nexus is deterministic and can reason over billions of spreadsheet rows without context window limitations.
The Series A was led by Oak HC/FT, Valor Equity Partners, Battery Ventures, and Salesforce Ventures, with angel backing from Perplexity's CEO and Datadog's CEO. Fundamental already has seven-figure Fortune 100 contracts and an AWS partnership for direct deployment. (source)
GPT-5 cuts cell-free protein synthesis costs 40% in autonomous lab experiment
An autonomous lab combining OpenAI's GPT-5 with Ginkgo Bioworks' cloud automation reduced cell-free protein synthesis costs by 40% through closed-loop experimentation. The system designed, executed, and iterated on experiments without human intervention.
The result demonstrates AI's growing ability to accelerate scientific research, with the model optimizing experimental parameters that human researchers had not explored. It marks a significant milestone in AI-driven biological research. (source)
This is enough for 7 min! Read 24 more stories in 7min.ai