Claude Code Launch: Terminal Agent Capabilities and Limits

One announcement, two different products

On February 24, 2025, Anthropic announced Claude 3.7 Sonnet and, in the same post, a separate command-line tool called Claude Code. The two were bundled in one announcement but were not the same kind of release: Claude 3.7 Sonnet was a new model, rolled out immediately to all Claude users and API customers, while Claude Code was a limited, invite-based research preview distributed as an npm package (npm install -g @anthropic-ai/claude-code) with first-come-first-served access rather than a general product launch. [1][2]

Anthropic called Claude 3.7 Sonnet its first hybrid reasoning model: a single model that answers normally by default, or can be told to spend extra tokens on a visible chain of reasoning before producing a final answer. The reasoning tokens, which Anthropic calls extended thinking, are shown to the user rather than hidden, which Anthropic contrasted with OpenAI's o1, where the reasoning trace is not exposed. API users could set a thinking-token budget up to 64,000 tokens by default, or up to 128,000 by adding a beta header (anthropic-beta: output-128k-2025-02-19). Pricing was $3 per million input tokens and $15 per million output tokens, with thinking tokens billed as output tokens, more expensive per token than dedicated reasoning models on the market at the time such as OpenAI's o3-mini ($1.10 / $4.40 per million) or DeepSeek's R1 (55 cents / $2.19 per million). [1][2][3]

What the research preview could and could not do

At launch, Claude Code ran entirely from a local terminal session: it could search and read files across a repository, edit files directly, write and run tests, execute arbitrary command-line tools, and perform git operations including committing and pushing to GitHub. It had no separate IDE, browser interface, or hosted execution environment; it read and wrote against the developer's own filesystem and used the developer's own local git configuration and shell, calling out to Anthropic's API only to generate the next step in the loop. Anthropic described the tool as keeping the user in the loop at every step, which in practice meant a permission model: rather than executing every action unattended, the agent surfaced proposed file edits and shell commands for the developer to approve before they ran. Authentication used OAuth against a developer's Anthropic account, which Anthropic's own writeup called a first for its API at the time, tying usage to the developer's existing billing rather than a bundled subscription price. [1][3]

There was no flat price for Claude Code itself at launch: usage was metered against the same per-token API pricing as Claude 3.7 Sonnet. Developer and Anthropic advisor Simon Willison, testing it on release day, reported that an extended working session could run up single-digit dollars of API spend and described the tool as an extremely powerful agent for coding while cautioning that its token consumption during long sessions could surprise a first-time user. Anthropic's own follow-up guidance two months later, aimed at internal and early external users, still centered on iterative workflows such as customizing a project's environment file and running multiple Claude Code instances in parallel, rather than around a finished, polished product. [3][4]

Coding agents were not a new idea in February 2025

Atlas interpretation: Claude Code did not introduce the idea of an AI system that edits code and runs commands with limited supervision. Cognition Labs had announced Devin on March 12, 2024, marketed explicitly as the first AI software engineer, capable of planning and executing multi-step coding tasks with comparatively little human intervention. GitHub announced a technical preview of Copilot Workspace, a Copilot-native environment for turning a task description into a pull request, on April 29, 2024. Cursor, an AI-native fork of VS Code from Anysphere, had been public since March 2023 and had an established developer user base well before February 2025. What distinguished Claude Code from that field was less the underlying idea than how it was presented and delivered: a plain command-line tool that worked inside a developer's existing terminal, shell, and git setup rather than a new IDE or a cloud-hosted autonomous run, shipped explicitly as a limited preview rather than a finished product, and paired with a model release rather than sold as a standalone company. Devin's early reception had already turned skeptical by the time Claude Code shipped: independent testing reported by The Register in January 2025 found some evaluators judged Devin poor at the job it was marketed for, which set a backdrop of lowered expectations for autonomous coding claims that Claude Code's more modest, permission-gated pitch arrived into. [5][6][7]

Why this release counts as a turning point

Atlas interpretation: Claude Code stayed a limited preview for three months. Anthropic made it generally available on May 22, 2025, the same day it introduced Claude Opus 4 and Claude Sonnet 4, models Anthropic sold partly on how long they could run unattended on a coding task, with early testers cited as reporting multi-hour autonomous sessions. Anthropic's own general-availability announcement tied Claude Code's expansion, into IDEs and a background SDK mode, directly to that model release, treating the terminal tool from February as the seed of a product line rather than a one-off SDK release. By late 2025, third-party analyst estimates, compiled by research firm Sacra from Anthropic's disclosures to investors, put Claude Code's annualized revenue run rate at roughly $1 billion by November 2025, six months after general availability, climbing to about $2.5 billion by February 2026, with enterprise customers accounting for more than half of that revenue. These are reported estimates rather than audited figures Anthropic has published directly, but the scale and speed of the reported growth is what commentators cite as evidence the February 2025 preview marked a shift in how coding agents were used, from a demo capability into a paid, workflow-embedded tool. Anthropic later built on the same premise directly: Agent Skills, announced October 16, 2025, packaged reusable, file-based expertise for agents including Claude Code, extending the same terminal-agent model rather than replacing it. [7][8][9]