Skip to main content

​
Weekly AI Review - 2026-02-17

Anthropic closed a staggering 30billionSeriesGata30 billion Series G at a 380 billion valuation. The second-largest venture deal in history as Claude Code’s revenue doubled to $2.5 billion run-rate in under two months. The open-source frontier exploded with MiniMax M2.5 matching Opus-class performance at 1/20th the cost and GLM-5 shipping a 745B-parameter model trained entirely on Huawei chips under MIT license. OpenAI diversified its inference infrastructure by deploying Codex-Spark on Cerebras chips for 1,000+ token/second coding. Meanwhile, the concept of “cognitive debt” emerged as the industry’s sharpest critique of AI-accelerated development, and India hosted the largest Global South AI summit yet.

​
Highlight of the week

Anthropic raises 30Bat30B at 380B valuation as Claude Code becomes a revenue machine [Anthropic closed a 30billionSeriesG](https://www.anthropic.com/news/anthropic−raises−30−billion−series−g−funding−380−billion−post−money−valuation)ledbyGICandCoatue,nearlydoublingitsvaluationfromthe30 billion Series G](https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation) led by GIC and Coatue, nearly doubling its valuation from the 183 billion Series F just five months prior. The round brings total capital raised to 69.1billion,makingitthe[second−largestventuredealever](https://news.crunchbase.com/ai/anthropic−raises−30b−second−largest−deal−all−time/)behindOpenAI′s69.1 billion, making it the [second-largest venture deal ever](https://news.crunchbase.com/ai/anthropic-raises-30b-second-largest-deal-all-time/) behind OpenAI's 40 billion raise last year. The numbers behind the valuation tell the real story. Anthropic’s run-rate revenue hit [14billion](https://siliconangle.com/2026/02/12/anthropic−closes−30b−round−annualized−revenue−tops−14b/),growingover10xannuallyforthreeconsecutiveyears.Morethan500customersspendover14 billion](https://siliconangle.com/2026/02/12/anthropic-closes-30b-round-annualized-revenue-tops-14b/), growing over 10x annually for three consecutive years. More than 500 customers spend over 1 million annually. Eight of the Fortune 10 are customers. But the standout is Claude Code. Since its May 2025 GA launch, Claude Code’s run-rate revenue has surged past $2.5 billion. More than doubling since January 1, 2026. Weekly active users also doubled in the same period. Business subscriptions quadrupled, with enterprise use now exceeding half of all Claude Code revenue. Claude Code currently authors 4% of GitHub public commits, and Semianalysis projects that number could reach 20% or more by year-end. For context, Goldman Sachs has deployed Claude across $2.5 trillion in assets under supervision, with 12,000+ developers using it. Reported outcomes: 30% faster client onboarding, 20%+ developer productivity gains. The “multi-model enterprise stack” is now the norm: 79% of OpenAI users also pay for Anthropic, reinforcing that this market isn’t winner-takes-all. Anthropic has begun IPO preparations, hiring Wilson Sonsini to advise. The only question is when, not whether.

​
Models & research

​
Open-source frontier models arrive in force

This was the week the open-source tier closed the gap on proprietary models in a way that matters for production workloads. MiniMax M2.5 landed on February 12 as a 230B-parameter MoE model (10B active per forward pass) that scores 80.2% on SWE-Bench Verified, within 0.6 points of Opus 4.6 (80.8%) and ahead of GPT-5.2 (80%). The Lightning variant runs at 100 tokens/second and costs 0.30/0.30/2.40 per million tokens, roughly 1/20th the cost of Opus 4.6. Trained with a novel RL algorithm (CISPO) across 200,000+ real-world environments, MiniMax claims 80% of their newly committed code is generated by M2.5. The model uses Lightning Attention for native 1M+ token context with linear efficiency. Released under a modified MIT license on Hugging Face. GLM-5 from Zhipu AI dropped on February 11 as a 745B MoE model (44B active) trained entirely on Huawei Ascend chips (zero NVIDIA dependency). It scores 77.8% on SWE-bench, beating Gemini 3 Pro and GPT-5.2 while trailing Opus 4.6. Released under MIT license at 1.00/1.00/3.20 per million tokens. The model introduced “Slime,” a novel asynchronous RL training infrastructure that substantially improves throughput. Zhipu’s shares surged 34% on the Hong Kong Stock Exchange following the launch. Simon Willison noted the significance of Zhipu’s “Agentic Engineering” framing, a term gaining traction alongside similar usage by Andrej Karpathy and Addy Osmani. Kimi K2.5 from Moonshot AI continues gaining adoption with its 1T-parameter MoE architecture (32B active), 76.8% on SWE-bench, and a distinctive Agent Swarm feature that orchestrates up to 100 sub-agents across 1,500 coordinated steps. At $0.60 per million input tokens, it’s 10x cheaper than Claude. The takeaway: teams now have multiple open-weight models within a few percentage points of the best proprietary offerings, at a fraction of the cost. The premium for closed models increasingly comes down to ecosystem, tooling, and trust rather than raw capability.

​
OpenAI Codex-Spark on Cerebras: the first move beyond NVIDIA

OpenAI released GPT-5.3-Codex-Spark on February 12, marking its first production deployment on non-NVIDIA hardware. Running on Cerebras Wafer-Scale Engine 3 (a single chip the size of a dinner plate with 4 trillion transistors), Codex-Spark delivers over 1,000 tokens per second, enabling near-instant feedback during live coding. This is part of a $10 billion multi-year partnership with Cerebras to bring 750MW of compute online through 2028. Currently in research preview for ChatGPT Pro users. OpenAI was careful to note that GPUs remain its primary inference platform, positioning Cerebras as complementary for latency-critical workloads. Still, it signals meaningful diversification of the AI inference stack beyond NVIDIA’s dominance.

​
Coding agents & dev tools

​
The multi-agent IDE era arrives

The AI coding market has converged on multi-agent parallel execution as the defining capability of 2026. Every major player now supports running multiple AI agents simultaneously:
  • Cursor 2.0 shipped its own Composer model (4x faster than comparable models) and supports up to 8 agents in parallel via Git worktrees. The new agent-centric interface makes agents, plans, and runs first-class objects. However, Gergely Orosz rightly observes that orchestrating parallel agents requires tech-lead-level skills, creating an accessibility gap.
  • Windsurf Wave 13 added multi-agent sessions with Git worktree integration, an Arena Mode for blind model comparison, and free SWE-1.5 (950 tokens/second). At $15/month, it ranks #1 in LogRocket’s February AI dev tool power rankings.
  • OpenAI Codex App launched February 2 for macOS as a command center for orchestrating multiple coding agents in parallel. Over 1 million developers used it in the first month. The Automations feature lets agents run on schedules in the background with results queued for review.
  • Google Antigravity, Google’s free AI IDE built from the Windsurf acquisition ($2.4B), offers both Editor and Manager views, the latter serving as a control center for parallel agent orchestration. Currently free in public preview.
  • VS Code 1.109 shipped multi-agent orchestration in February, letting users run Claude and Codex agents alongside GitHub Copilot. This changes the Cursor value proposition since teams can now get multi-model agent support within their existing editor.

​
GitHub Agentic Workflows: “Continuous AI”

GitHub previewed Agentic Workflows (gh-aw), a system where you define AI-powered automation tasks in markdown that execute as GitHub Actions. The agent can be Copilot, Claude Code, or OpenAI Codex. Use cases include issue triage, CI failure investigation, test coverage monitoring, and repo health reports. The key insight: these aren’t replacements for CI/CD (which must be deterministic) but a new “Continuous AI” layer alongside it. Workflows run read-only by default with explicit approval gates for writes. Sample workflows include Issue Triage, CI Doctor, and Dependabot PR Bundler. For teams already on GitHub, this is the lowest-friction path to automated AI-driven repo maintenance.

​
Kimi Code: open-source terminal agent

Kimi Code CLI (v1.12, 6,400+ GitHub stars, Apache 2.0) is Moonshot AI’s open-source terminal coding agent. It competes directly with Claude Code and Gemini CLI, supports images and video input, auto-discovers existing MCP tools, and integrates with VSCode, Cursor, and Zed. For teams wanting an open-source alternative to proprietary coding agents, this is the strongest option available.

​
The reality check on AI coding productivity

Amazon wants 80% of its developers using AI coding tools weekly. But there’s a twist: internal engineers are pushing back against being steered toward Kiro (Amazon’s in-house tool) when ~1,500 employees endorsed Claude Code in an internal forum. Despite Amazon’s $8 billion investment in Anthropic, employees need explicit authorization to use Claude Code on production projects. A February Forbes analysis found that leading AI coding models succeed less than 50% of the time at fully generating requested features. The practical guidance: spec-driven development, test-first workflows, and outcome-based validation remain essential. MIT Technology Review reports AI now writes ~30% of Microsoft’s code and over 25% of Google’s, with Meta targeting majority AI-generated code in the near future. The industry consensus on productivity gains has settled at a realistic 20-30%, concentrated in specific workflows rather than the 10x improvement initially promised.

​
Cognitive debt: the emerging risk

The most important conceptual contribution of the week comes from Margaret-Anne Storey, who coined “cognitive debt” as the silent cost of AI-accelerated development. Unlike technical debt that lives in code, cognitive debt is the erosion of shared understanding in developers’ minds. Teams move fast with AI but lose the mental models of their systems: what the code does, why architectural decisions were made, how to change things safely. Simon Willison called it “the best explanation of the term I’ve seen.” This concept should be on every engineering leader’s radar. As AI agents generate more code, the risk isn’t just code quality; it’s whether anyone on the team actually understands the system anymore.

​
Web development & frameworks

​
AI-native development platforms mature

The “app builder” category (tools like Cursor, Replit, and “lovable”-style platforms that let non-experts build full web apps from instructions) is now compressing frontend and backend development cycles enough to shift the full-stack role toward architecture, integration, and review. Builder-style platforms now ship AI assistants for code search, doc generation, and interactive API sandboxes tightly integrated into component systems. AI-powered security platforms are bundling agent-based “fixers” that propose or apply patches directly in CI/CD. Snyk’s Agent Fix auto-generates security patches; Sweep.dev creates automated bug-fix PRs. The debate now is how much autonomy to grant these agents in production pipelines. On the content side, RAG-powered Q&A embedded directly into sites is replacing traditional search-engine-driven traffic patterns. Content teams are rethinking SEO as “answer-in-assistant” flows grow.

​
Industry & business

​
Infrastructure: power becomes the bottleneck

AI infrastructure discussions this week centered on power, not GPUs. Peak XV invested in Indian startup C2i focused on data-center power and cooling bottlenecks. The transition to exaFLOP-class GPU fabrics (racks with 70+ accelerators drawing 30-80kW) is pushing adoption of direct-to-chip liquid and immersion cooling. Nebius reported a sharp rise in capex from GPU and data center purchases. AMD’s MI450 “Helios” rack-scale platform, expected late 2026, integrates EPYC CPUs, Instinct GPUs, and networking into a unified stack to challenge NVIDIA’s dominance.

​
Interesting GitHub repositories

​
Portless

Replaces port numbers with stable, named .localhost URLs (e.g., myapp.localhost:1355). Designed for both humans and AI agents navigating multi-service local dev environments. Solves real pain in monorepos where AI agents guess or hardcode wrong ports, and where cookies bleed across apps on different ports. Works with Next.js, Vite, and most frameworks that respect the PORT env var. TypeScript, Node.js 20+.

​
Roam

CLI tool that pre-indexes codebases into semantic graphs (symbols, dependencies, call graphs) in SQLite. AI agents query the graph instead of repeatedly grepping files. 55 commands across 7 categories including blast radius analysis, codebase health scoring, and cross-repo support. Supports 22 languages via Tree-sitter. This is the kind of infrastructure that makes AI coding agents actually useful on large codebases. Python 3.9+.

​
gh-aw (GitHub Agentic Workflows)

GitHub’s official tool for defining AI-powered automation as markdown files that compile to GitHub Actions YAML. Read-only by default, sandboxed execution, supply chain security with SHA-pinned dependencies. Sample workflows for issue triage, CI investigation, and dependency management. Go/JavaScript, MIT license.

​
Skill Compose

Platform for building AI agents from reusable, versioned skill components rather than workflow graphs. Describe an agent in natural language, and it finds existing skills, creates missing ones, and wires them together. Container-based isolation, GitHub integration for version history and rollback. Python 3.11+ backend, Next.js 14 frontend, Apache 2.0.

​
Claude SEO

Universal SEO skill for Claude Code: full website audits with parallel subagent delegation, E-E-A-T analysis, schema markup validation, Core Web Vitals assessment, and AI search optimization for Google Overviews and ChatGPT search. Python 3.8+, MIT license.

​
Quick Bits

​
Sources

​
Anthropic & Claude

​
Open-source models

​
OpenAI & Codex

​
Coding agents & dev tools

​
Cognitive debt & developer experience

​
Industry & business

​
Infrastructure

​
AI safety & policy

​
GitHub repositories

Last modified on April 14, 2026