Overview and positioning
Amp is no longer a Sourcegraph product. Amp's December 2025 announcement says it spun out into Amp Frontier Corporation, an independent research lab, and the current manual describes Amp as a multi-model coding agent for terminal and editor workflows. Its low, medium, high, and ultra modes currently draw on GPT-5.6, Claude Fable 5, and faster models, while the product chooses how to apply that mix. The result is a deliberately model-agnostic experience: buyers choose an effort mode and workflow, not a single model provider for every task.
Claude Code is Anthropic's first-party coding agent and keeps the Claude model family at the center of the experience. It runs locally, in Anthropic-managed cloud environments, and through Remote Control, while interfaces include the terminal, VS Code, JetBrains IDEs, claude.ai/code, Slack, and CI/CD pipelines. Claude Code can read project files, execute shell commands, inspect Git state, and carry persistent instructions through CLAUDE.md plus automatic memory. That breadth makes it less a terminal-only utility than a Claude-native engineering layer deployed across several surfaces.
Agent capabilities and extension model
Amp's current capability set is built around delegation and inspectable work. Subagents handle isolated tasks, the Oracle supplies a second opinion from a different frontier model, and threads preserve prompts, context, tool calls, and results as shareable artifacts. Amp also supports MCP servers, plugins, agent skills, remote runners, and orbs for durable cloud execution. The four effort modes create a clear control surface: low and medium target cheaper or balanced work, while high and ultra reserve stronger reasoning for open-ended tasks. Teams that want the agent to select models and split work automatically get that behavior without maintaining a provider-routing layer themselves.
Claude Code's extension surface is broader and more explicitly composable. CLAUDE.md provides always-loaded project conventions; skills package repeatable instructions; MCP connects external services; hooks can run scripts, HTTP requests, prompts, MCP tools, or subagents on lifecycle events; and plugins distribute these pieces together. Custom subagents isolate specialized work, while agent teams coordinate independent sessions with shared tasks. These primitives let an organization encode policy and workflow around the agent instead of relying only on conversational prompting, which is the strongest structural reason to prefer Claude Code for standardized engineering environments.
Developer workflow and collaboration
Amp emphasizes handing work to agents and reviewing the resulting thread. Its CLI connects to VS Code and VS Code-based editors, JetBrains, Neovim, and Zed; the web interface can search and share threads, remote-control local runs, and supervise orb-backed work. Workspaces pool credits and expose thread visibility controls, while Multiplayer allows workspace members to join eligible orb threads, send messages, and access the shared terminal and files for a bounded session. This workflow suits teams that treat an agent run as a collaborative unit of work and want parallel remote execution to remain visible from a common feed.
Claude Code emphasizes bringing the same agentic loop to the environment a team already uses. Local sessions retain project and Git context; cloud sessions offload work to Anthropic-managed machines; Remote Control keeps execution local while moving the interface to the browser. The VS Code integration exposes native diffs, editor selections, diagnostics, and controlled notebook execution, while JetBrains, Slack, and CI/CD provide additional entry points. Persistent CLAUDE.md instructions, automatic memory, hooks, and permissions make repeatability the center of the workflow, especially when many developers need the agent to follow the same repository rules.
Pricing and cost controls
Amp now combines free credits, subscriptions, and metered usage. Its free program states that credits replenish hourly up to $10 per day, although Amp labels the program experimental. The Megawatt beta subscription costs $20 per month and includes $20 of agent usage, 750 hours of small orbs, and low and medium modes. Gigawatt costs $200 per month, includes $200 of agent usage, 1,000 hours of large orbs, and unlocks high and ultra. Subscriptions are optional: users can still buy pay-as-you-go credits, with the manual stating zero markup on provider API prices for individuals and non-enterprise workspaces and a $5 minimum purchase.
Claude also combines included subscription usage with optional metered continuation. Pro costs $20 monthly or about $17 per month on the $200 annual plan and includes Claude Code. Max starts at $100 per month for 5x Pro usage, with a $200 monthly 20x tier. Team Standard costs $20 per seat monthly when billed annually or $25 month-to-month; Premium costs $100 annually or $125 monthly. Usage is shared across Claude products and subject to rolling and weekly limits. Paid users can enable usage credits after those limits, and Pro, Max, and Team customers can buy discounted bundles, so Claude should not be described as an unconditional flat monthly ceiling.
Which teams should choose each
Choose Amp when multi-model routing, shared agent threads, and remote orb execution are central requirements. It is the cleaner fit for teams that want GPT and Claude models available through one agent, prefer effort modes over manual model selection, and expect reviewers to inspect or join the agent's working thread. The $20 and $200 subscriptions make recurring spend easier to plan than Amp's older usage-only model, while pay-as-you-go credits remain useful for bursty workloads. Amp is also attractive when linking an existing ChatGPT subscription for additional GPT-5.6 usage materially improves the team's economics.
Choose Claude Code when the organization wants a Claude-native workflow that can be standardized through persistent instructions, skills, MCP, hooks, permissions, subagents, agent teams, and plugins. It is better suited to repositories where conventions must load every session, deterministic automation must run around tool use, or the same agent needs to appear in terminal, IDE, browser-controlled local sessions, Slack, cloud environments, and CI. Claude's plan limits and optional usage credits still require spend controls, but Team and Enterprise administrators can apply organizational limits rather than treating subscription price alone as the budget boundary.
Verdict
Claude Code wins for the broadest professional audience, but not because Amp lacks subscriptions or because Claude offers a perfectly fixed bill. The deciding advantage is Claude Code's cohesive extension and deployment layer: CLAUDE.md, automatic memory, skills, MCP, hooks, subagents, agent teams, plugins, IDE integrations, cloud execution, Remote Control, Slack, and CI all wrap the same first-party agentic loop. That gives engineering leaders more ways to encode repeatable behavior and place the agent inside existing delivery controls, which matters more than nominal entry price once usage credits and overages exist on both sides.
Amp remains the stronger specialist choice for teams that value model diversity, shareable execution threads, and orb-centered collaboration above Claude-specific workflow depth. Its independent-company status, current four-mode model mix, Oracle, subagents, Multiplayer, and newly added $20/$200 subscriptions make it far more complete than stale usage-only comparisons imply. Buyers should shortlist Amp when avoiding a single-model commitment is a primary requirement. For a default recommendation across individual developers and governed teams, however, Claude Code's integrated extension surface and wider set of execution interfaces justify the concrete winnerTool=claude-code decision.