Back to list
Weekly 2026 Week 22

AI Pulse Weekly | 2026-W22

AnthropicOpenAIMetaMicrosoftNVIDIADeepMindCognitionMistralAI AgentValuationClaude

This Week in Focus: Anthropic Nears $1T Valuation, AI Agents Go OS-Native, Claude Solves Century-Old Math Problem

Extracting real signal from the noise

Another explosive week for the AI industry. Anthropic closed a $65 billion Series H funding round, pushing its valuation toward the trillion-dollar mark — the largest single funding round in AI startup history. Claude Opus 4.8 officially launched, surpassing GPT-5.5 on most benchmarks with a fast mode that’s 3x cheaper. OpenAI’s Codex can now autonomously operate Windows PCs, evolving from a coding assistant into an OS-level Agent. Claude Mythos reportedly solved OpenAI’s landmark Erdős math problem. Meta’s leaked internal memo revealed a full strategy spanning AI pendants, supersensing glasses, and enterprise wearables.

This is a week where “Agent-native” is accelerating into reality.

Anthropic Raises $65 Billion in Series H, Nearing Trillion-Dollar Valuation

On May 28, The Decoder reported that Anthropic raised $65 billion in its Series H round, pushing its valuation toward the trillion-dollar threshold.

This is one of the largest single funding rounds in AI startup history. Anthropic’s strong enterprise performance with the Claude series has attracted sustained investment from top global capital.

AI Pulse View: A $65 billion single round surpasses almost all historical tech company funding records. It signals that the market has extremely high expectations for Anthropic’s positioning in the Claude ecosystem. A trillion-dollar valuation means Anthropic has entered discussions at the same level as Google, Meta, and Microsoft. But high valuations bring high expectations — Anthropic needs to prove its commercialization capabilities can sustain this number over the next 12-18 months.

Source: The Decoder

Anthropic Ships Claude Opus 4.8: Tops GPT-5.5 on Most Benchmarks, Fast Mode 3x Cheaper

On May 28, The Decoder and VentureBeat reported that Anthropic released Claude Opus 4.8, described as a “modest but tangible improvement” that tops GPT-5.5 on most benchmarks, with a fast mode that’s 3x cheaper.

Opus 4.8’s positioning is interesting — not a revolutionary leap, but a steady, perceptible upgrade while dramatically reducing cost.

AI Pulse View: With Claude Mythos already occupying the performance peak, Opus 4.8’s strategy is pragmatic: don’t chase the “most powerful” narrative, instead deliver “good enough and cheap.” A fast mode that’s 3x cheaper means enterprises can deploy Claude capabilities at significantly lower cost. This may be Anthropic’s differentiated response to low-cost competitors like DeepSeek — not competing on absolute price, but on cost-performance ratio.

Source: The Decoder | VentureBeat

Sam Altman and Dario Amodei Walk Back AI Job Apocalypse Predictions

On May 27, The Decoder reported that OpenAI CEO Sam Altman and Anthropic CEO Dario Amodei both walked back or softened their previous predictions about AI causing mass unemployment.

The attitude shift from two top AI company CEOs reflects a reassessment of AI’s speed and timeline for replacing human labor.

AI Pulse View: This is a telling signal. Previously, AI leaders universally pushed the “AI will replace大量 jobs” apocalyptic narrative, and now they’re walking it back. This could be because actual enterprise deployment data shows AI replacement is slower than expected, or due to political and social pressure. But the more likely explanation: they’re discovering AI is more of a “capability amplifier” than a “job replacer” — consistent with what we’re seeing in the Agent space, where AI is expanding human capability boundaries rather than simply replacing workers.

Source: The Decoder

OpenAI Codex Can Now Operate Your Windows PC Autonomously

On May 30, The Decoder reported that OpenAI’s Codex can now autonomously operate Windows PCs, independently hunting bugs and testing applications.

Codex is evolving from a coding assistant into an OS-level autonomous Agent — it no longer just generates code, but can execute tasks within an operating system like a human would.

AI Pulse View: This is a key step for AI Agents moving from “tool” to “colleague.” When AI can autonomously operate an entire OS, its role shifts from “write what you tell it” code generator to “go look, go fix” autonomous engineer. This means AI’s penetration into software engineering will accelerate further. But it also raises new security questions: an AI Agent with OS-level access needs entirely redefined permission boundaries and safety guardrails.

Source: The Decoder

Meta’s Leaked Memo Reveals AI Pendant, Supersensing Glasses, and Enterprise Wearables Strategy

On May 30, The Decoder and TechCrunch reported that a leaked internal memo from Meta revealed the company’s full strategy for AI pendants, supersensing glasses, and enterprise wearables.

Zuckerberg has finally put a price tag on all that AI hardware spending — Meta is going all-in on AI wearables.

AI Pulse View: Meta’s ambition is broad: from pendant (possibly a standalone AI device like Rabbit R1 or Humane Pin) to supersensing glasses (capabilities far beyond Google AI glasses) to enterprise-grade wearables. This shows Meta believes the next AI interaction interface isn’t the phone or the computer — it’s devices worn on the body. But Meta has had hardware missteps before (Ray-Ban Stories, Portal), and this time’s success depends on whether AI capabilities are strong enough to support a genuinely new interaction paradigm.

Source: The Decoder | TechCrunch

Microsoft and Nvidia Team Up on AI PCs That Run Actual Agents Instead of Copilot

On May 30, The Decoder reported that Microsoft and Nvidia are collaborating on AI PCs that will run true autonomous agents rather than the current Copilot assistants.

This marks a fundamental shift in Microsoft’s AI PC positioning — from “assistant tool” to “autonomous executor.”

AI Pulse View: Copilot is fundamentally “you ask, it answers,” while a true agent is “give it a goal, it does the work.” If Microsoft and Nvidia integrate this capability at the PC hardware level, it’s a platform-level transformation. Nvidia provides local inference compute, Microsoft provides the Agent framework and OS integration — a complete local AI Agent stack. This also means AI Agent deployment will accelerate from cloud to edge.

Source: The Decoder

Claude Mythos Reportedly Solves OpenAI’s Landmark Erdős Math Problem

On May 26, The Decoder reported that Claude Mythos reportedly solved OpenAI’s landmark Erdős math problem with an “elegant and simple proof.”

If true, this is another milestone breakthrough in AI’s pure mathematical reasoning capabilities.

AI Pulse View: The Erdős problem is a classic in combinatorial mathematics, and OpenAI used it as a benchmark for testing AI mathematical reasoning. Claude Mythos solving it with an “elegant and simple proof” suggests its reasoning has moved beyond “brute-force search + pattern matching” into “genuinely understanding mathematical structures.” Combined with DeepMind’s AlphaProof Nexus solving decades-old math problems for just a few hundred dollars, 2026 is shaping up to be a breakout year for AI mathematical capabilities.

Source: The Decoder

Cognition (Devin’s Maker) Doubles Valuation to $26 Billion in Under 9 Months

On May 27, The Decoder reported that Cognition, the company behind AI coding agent Devin, has more than doubled its valuation to $26 billion in less than 9 months.

This is one of the brightest stars in the AI Agent赛道. Devin, as one of the first autonomous coding agents to capture widespread attention, has proven the market is willing to pay massive premiums for “AI programmers.”

AI Pulse View: A doubled valuation in 9 months shows that AI Agent commercialization in coding is far outpacing market expectations. But a $26 billion valuation also means Cognition needs to deliver convincing productization and revenue growth. With OpenAI Codex, Claude Code, and competitors iterating rapidly, whether Devin can sustain its first-mover advantage is a key question.

Source: The Decoder

Mistral Rebrands LeChat as Vibe, Bets on Full-Blown Work Agent

On May 28, The Decoder and VentureBeat reported that Mistral AI rebranded LeChat as Vibe, announced expansion into industrial AI, and revealed data center expansion plans to challenge OpenAI.

This marks Mistral’s strategic transformation from “open-source model company” to “full-stack AI Agent platform.”

AI Pulse View: The name “Vibe” is interesting — it suggests a natural, intuitive work experience rather than a traditional “chatbot.” Mistral’s strategy is clear: maintain open-source competitiveness at the model layer, build an enterprise Agent platform with Vibe at the application layer, and reduce supply chain dependency through self-built data centers. This “three-pronged” strategy is particularly attractive under the European AI sovereignty narrative.

Source: The Decoder | VentureBeat

One Company Spent $500 Million on Claude in One Month After Failing to Cap AI Usage

On May 29, The Decoder reported that one company spent $500 million on Claude in a single month after failing to set AI usage caps.

The number is staggering — it exposes a critical issue in enterprise AI deployment: cost control.

AI Pulse View: A $500 million monthly AI bill shows that the “cost runaway” risk companies face when adopting AI Agents is real. When AI Agents start autonomously executing tasks (coding, data analysis, customer interactions), usage grows exponentially. This creates a new market demand: AI usage management and cost optimization. Pinterest’s simultaneous report of cutting AI costs 90% by gutting a frontier model’s vision layer shows there’s massive room for optimization.

Source: The Decoder

DeepMind’s AlphaProof Nexus Solves Decades-Old Math Problems for Just a Few Hundred Dollars

On May 27, The Decoder reported that Google DeepMind’s AlphaProof Nexus solved math problems that had stumped mathematicians for decades, at a cost of only a few hundred dollars.

This is another breakthrough in AI mathematical reasoning, at remarkably low cost.

AI Pulse View: The “few hundred dollars” figure is more significant than the problem-solving itself. It means the cost of AI solving mathematical problems has dropped low enough for widespread use. This isn’t a lab demo — it’s math reasoning capability that can be scaled. Combined with Claude Mythos solving the Erdős problem, AI mathematical reasoning is moving from “can do some” to “can do a lot.”

Source: The Decoder

Robinhood Lets AI Agents Trade Shares and Make Credit Card Purchases for Customers

On May 27, The Decoder reported that Robinhood now allows AI Agents to autonomously trade shares and make credit card purchases for customers.

This is a significant step for AI Agents entering finance — moving from “decision support” to “autonomous execution.”

AI Pulse View: Letting AI Agents autonomously trade stocks and make purchases crosses a major trust threshold. Robinhood’s move means they believe AI Agent decision quality can independently bear financial consequences. But this also raises new regulatory questions: when an AI Agent’s autonomous trading causes losses, who’s liable? This is a question fintech must answer.

Source: The Decoder

Other Notable Developments

One-Line Summary

Anthropic nears trillion-dollar valuation with $65B funding, Claude Opus 4.8 and Mythos advance on dual fronts, OpenAI Codex evolves into OS-level Agent, Meta and Microsoft bet on hardware and local AI PCs respectively. AI Agents are moving from “coding assistant” to “autonomous executor,” from “cloud” to “edge,” from “tool” to “colleague.”