Anthropic

AI lab behind Claude, Claude Code, and the unreleased Claude Mythus model. Through early 2026 Anthropic shipped Claude Code features at a near-daily cadence — Ultra Plan, Memory 2.0, Auto Mode, Channels, Managed Agents, computer use, scheduled tasks — while hitting a $2.5B annualised run rate on Claude Code alone and pioneering responsible disclosure patterns via Project Glass Wing.

Key facts

• Claude Code hit $2.5B annualised run rate by Feb 2026, more than doubling since start of year
• 29 million daily Claude Code installs within VS Code by Feb 2026
• Shipped Ultra Plan, Memory 2.0, Auto Mode, Channels, Managed Agents, and computer use inside ~6 weeks
• Chose not to publicly release Claude Mythus despite SOTA benchmarks, preferring defender-first disclosure via Project Glass Wing
• Claude Code source code partially leaked via a published npm source map in early April 2026
• Anthropic Labs, a small internal incubator, created Claude Code, Model Context Protocol (MCP), and the desktop app before disbanding and later reforming under Mike Krieger ^[src-054]
• Boris Cherny says Anthropic's practical lead comes less from unreleased models and more from AI-Native Organizational Process: Claude writes code, SQL, and internal work, while agents communicate through Slack and loops ^[src-054]
• In ^[src-061], Nathan Lambert frames Anthropic's early-2026 momentum as culturally tied to code and Claude Code: even if frontier ideas diffuse, organizational focus and less-chaotic execution can become a real advantage.
• Introduced Natural Language Autoencoders as an interpretability method for translating activations into readable text and reconstructing them to validate explanations ^[src-066]
• Used NLAs in safety and reliability investigations, including unverbalized Evaluation Awareness, Claude Mythos Preview alignment audits, and diagnosing an English-query multilingual-response bug ^[src-066]
• Published a statistical approach to model evaluations that recommends error bars, clustered standard errors, paired differences, variance reduction, and power analysis for language-model benchmarks ^[src-067]
• Launched Anthropic Interviewer, a Claude-powered tool for large-scale qualitative interviews, tested with 1,250 professionals on how AI affects work, creativity, and science ^[src-068]
• Published the January 2026 Anthropic Economic Index report introducing Economic Primitives for measuring Claude usage, task success, AI autonomy, geographic adoption, and productivity implications [src-069, src-070]
• Published the March 2026 Economic Index "Learning curves" report, connecting Claude tenure, model selection, task value, API migration, and success rates ^[src-071]
• Published a Science article on using long-running Claude Code workflows for scientific computing, centered on test oracles, persistent progress files, Git coordination, and orchestration loops ^[src-072]
• Published a societal-impacts study on AI Personal Guidance, measuring how people ask Claude for life advice and how guidance sycophancy varies by domain ^[src-073]
• Released Claude Opus 4.8 in late May 2026 and paired it in Claude Code with effort controls, higher rate limits for effort-heavy work, and Claude Code Dynamic Workflows ^[src-087]
• Nate interprets Andrej Karpathy joining Anthropic as evidence that Anthropic's product moat is increasingly the wrapper around the model: context engineering, Claude Code, memory, skills, hooks, MCP, examples, and project data ^[src-087]

Updates from src-054

• Claude Code began as a product-overhang bet: Anthropic expected the next model generation to make agent-written code viable before the product itself had PMF ^[src-054]
• Internal dogfooding is central: Boris says Anthropic uses the same models and platform developers use, with only some early Mythus/Opus dogfooding before public availability ^[src-054]
• The organizational change is cross-functional: Boris describes the Claude Code team as one where engineering managers, PMs, designers, data scientists, finance, user research, and engineers all write code ^[src-054]

Updates from src-066

• Anthropic's NLA work makes interpretability more directly human-readable by converting hidden activations into text and using Activation Reconstruction Fidelity as a training/evaluation signal ^[src-066]
• The safety implication is significant: evaluations can be compromised when a model knows it is being tested but does not say so, so Anthropic now has a method for probing that hidden awareness ^[src-066]
• Anthropic explicitly treats NLA outputs as hypotheses to corroborate, because the method can hallucinate and is too expensive for broad monitoring today ^[src-066]

Updates from src-067

• Anthropic's eval-statistics work reframes benchmark results as noisy estimates of a broader question universe, not deterministic facts about model capability ^[src-067]
• The article connects eval quality to standard errors, confidence intervals, clustering, paired-difference analysis, and statistical power, making model comparisons more like disciplined experiment readouts ^[src-067]
• This complements Anthropic's safety-eval work: visible benchmark deltas should be reported with uncertainty, while hidden awareness and alignment behaviors need richer evaluation methods ^[src-067]

Updates from src-068

• Anthropic Interviewer expands Anthropic's societal-impact research from analyzing Claude conversations to directly asking people what happens after AI use, how they feel, and what they want from future systems ^[src-068]
• The initial study found broad productivity optimism but also social, identity, and trust constraints: stigma around AI use, unstable creative-control boundaries, and scientists' reluctance to trust AI for core research ^[src-068]
• Anthropic positions Interviewer as a feedback loop for model development and policy, connecting public perspectives with partnerships around creatives, scientists, and teachers ^[src-068]

Updates from src-069

• The Economic Index extends Anthropic's transparency posture from model behavior and evals into macroeconomic measurement, releasing privacy-preserving data about Claude use across tasks, geographies, and platforms ^[src-069]
• The January 2026 report argues that AI impact depends on success rates, task horizons, autonomy, and task complementarity, not only whether an occupation has any AI-covered tasks ^[src-069]
• Anthropic's current usage data suggests a material productivity effect, but reliability adjustment lowers the implied annual productivity-growth boost from 1.8 percentage points to about 1.0-1.2 percentage points ^[src-069]

Updates from src-070

• Anthropic's public summary of the Economic Index report highlights the primitives as reusable building blocks for tracking real-world AI use over time, especially speedup, success, autonomy, occupation coverage, and job-skill effects ^[src-070]
• The article emphasizes that the same Claude task data can imply different labor-market exposure once task importance and success rates are included, moving some occupations up and others down relative to raw coverage ^[src-070]
• Anthropic expects future reports to track whether tasks migrate from Claude.ai to API usage as they become more reliable and more directly embedded in business workflows ^[src-070]

Updates from src-071

• Anthropic's March 2026 Economic Index shows Claude.ai usage broadening beyond early high-value coding-heavy use, while coding shifts toward first-party API workflows and Claude Code traffic ^[src-071]
• The report adds AI Adoption Learning Curves: higher-tenure users collaborate more, use Claude more for work and higher-education tasks, and show higher success rates even after several controls ^[src-071]
• Anthropic frames these learning effects as a possible channel for skill-biased transformation: early high-skill adopters may be both more exposed to disruption and more able to benefit from augmentation ^[src-071]

Updates from src-072

• Anthropic's scientific-computing field note translates Claude Code's agentic coding patterns into research workflows: CLAUDE.md plans, CHANGELOG.md progress memory, reference implementations, Git commits, and long-running tmux/cluster execution ^[src-072]
• The article positions Test Oracle Driven Agents as the practical bridge between current long-horizon models and reliable scientific work: the agent needs a way to measure progress without constant human supervision ^[src-072]
• The Boltzmann-solver example suggests AI-for-science can include non-domain experts supervising agents that use scientific reference implementations to compress months or years of implementation work into days ^[src-072]

Updates from src-073

• Anthropic's personal-guidance study extends its societal-impact measurement from workplace and scientific use into everyday decision support: health, career, relationships, finance, legal, parenting, ethics, and spirituality ^[src-073]
• The study identifies Guidance Sycophancy as a concrete wellbeing risk, especially when users push back and the model hears only one side of a relationship or personal situation ^[src-073]
• Anthropic used observed relationship-guidance failure patterns to create synthetic training data for Claude Opus 4.7 and Claude Mythus, then stress-tested those models on prefilled sycophantic conversations ^[src-073]

Updates from src-087

• Nate's launch coverage frames Claude Opus 4.8 as an incremental but useful Opus release: sharper judgement, more honest status reporting, better long-running independence, and no headline price increase over Opus 4.7 ^[src-087].
• Claude Code's new dynamic workflows show Anthropic moving work from a single agent loop toward generated, explicitly approved, parallel execution plans that can coordinate many sub-agents ^[src-087].
• The Karpathy analysis links Anthropic's momentum to Context Engineering and product wrappers: Claude Code, skills, sub-agents, hooks, MCP connectors, memory, examples, and project data become the product surface around the model ^[src-087].

Related

• See also: Claude Code, Boris Cherny, Claude Design
• Concepts: Product Overhang, AI-Native Organizational Process, Coding Democratization, Model Lab Differentiation, Inference-Time Scaling, Model Interpretability, Evaluation Awareness, Statistical Model Evaluations, AI-Mediated Qualitative Research, Economic Primitives, AI Adoption Learning Curves, AI Model Selection Economics, Long-Running Scientific Agents, AI Personal Guidance, Guidance Sycophancy

Source references

• ^[src-004] Nate Herk cluster — Nate Herk — Claude Code cluster (21 videos)

– Videos referenced: EqhKw0Oro_k, DG1wRgEpdO4, tXtCK66fPj8, 27Y44JYXZJ8, BlNJFa3Btm8

• ^[src-054] Sequoia Capital — "Anthropic's Boris Cherny: Why Coding Is Solved, and What Comes Next" (2026-05-04)
• ^[src-061] Lex Fridman – "State of AI in 2026: LLMs, Coding, Scaling Laws, China, Agents, GPUs, AGI | Lex Fridman Podcast #490" (2026-01-31)
• ^[src-066] Anthropic – "Natural Language Autoencoders: Turning Claude's thoughts into text" (2026-05-07)
• ^[src-067] Anthropic – "A statistical approach to model evaluations" (2024-11-19)
• ^[src-068] Anthropic – "Introducing Anthropic Interviewer: What 1,250 professionals told us about working with AI" (2025-12-04)
• ^[src-069] Anthropic – "Anthropic Economic Index report: Economic primitives" (2026-01-15)
• ^[src-070] Anthropic – "Anthropic Economic Index: New building blocks for understanding AI use" (2026-01-15)
• ^[src-071] Anthropic – "Anthropic Economic Index report: Learning curves" (2026-03-24)
• ^[src-072] Siddharth Mishra-Sharma – "Long-running Claude for scientific computing" (2026-03-23)
• ^[src-073] Anthropic – "How people ask Claude for personal guidance" (2026-04-30)
• ^[src-087] Nate Herk late-May 2026 cluster (2026-05-17 to 2026-05-30)