Amp drops per-token pricing and turns compute into its real business

By Zendoric · September 13, 2026.

Amp, Sourcegraph's coding agent, has changed its pricing model: from now on it is free for anyone who brings their own compute and their own model subscriptions or keys. A monthly plan is no longer needed to connect a ChatGPT subscription to Amp, and there is no fee or cap on BYOK (bring your own key) usage, except at the Enterprise tier.

The only thing Amp still charges for is the use of "orbs": remote computers owned by the company where coding agents run independently and in parallel. Anyone who prefers to run their agents on their own hardware, via local "runners," can do so at no cost. And if you choose to pay for model inference through Amp, the company says it does so with no markup on the provider's price.

The new plan structure is as follows: a free Hobby tier, two paid tiers for individual use—Megawatt and Gigawatt, with average discounts of 60% and 65% on Amp usage, according to the company's own figures—a Teams tier at no additional cost in which each member can be on a different plan—Hobby, Megawatt or Gigawatt—and draw on shared workspace credits, and a custom Enterprise tier. Paying users also get early access to new BYOK providers—OpenRouter, Amazon Bedrock, Google Cloud Agent Platform (formerly Vertex AI), Azure Foundry, Vercel AI Gateway, Cloudflare AI Gateway, Ollama Cloud, OpenCode Go and custom endpoints—and Sourcegraph is now formally extending its minimal or zero data retention policy to all users, not just Enterprise.

As sector context, Amp's move is not an isolated case: over the past year the market for AI coding agents (Cursor, GitHub Copilot, Windsurf, Claude Code, Devin, Amp itself) has gone from competing over which model each tool integrated to competing over how much friction can be removed to connect any model. The proliferation of cheap providers and competitive open-weight models—GLM, Qwen, DeepSeek and Kimi, among others—has made it unsustainable to charge a premium for mere access to the model when developers can get it cheaper elsewhere.

Our reading is that the value of these tools is shifting from the model layer to the infrastructure and corporate-trust layer: the winner is not whoever extracts the most margin per token, but whoever offers the most efficient execution environment—parallel, isolated, fast compute—and the guarantees a large company demands: SSO, MCP access control, audit logs, spending limits, data retention. Amp is explicitly walking away from the business of reselling tokens to keep the business of selling compute and trust; it is a bet that the margin no longer lies in reselling intelligence, but in the plumbing that makes it useful inside an organization.

In the short term this is good news for individual developers and small teams: less friction, lower fixed costs, more freedom to choose a model to suit the task. But it is also evidence of fierce competition compressing margins across the AI coding tools segment; players without a clear niche—their own compute, deep integration or a loyal Enterprise customer base—will struggle, and it is reasonable to expect consolidation in the coming quarters. It does fit, however, with the underlying trend that interests us most: the cheaper and more accessible AI-assisted programming becomes, the closer we are to that productivity surplus translating into tools, and in the long run into abundance, within reach of anyone and not just those who can pay a provider's toll.