On July 15, 2025, Amp, the coding agent spun out of Sourcegraph, deleted its $20 monthly platform fee. It also eliminated the BYOK token fee. A free Hobby tier now sits at the top of its pricing page. The announcement arrived without a published cap on concurrent tasks, Orb runtime, context window, or model access. That is the first red flag. Ledger balances do not lie; they only wait. In my decade of auditing token launches and DeFi protocols, I have learned that "free" is never a price. It is a transfer of cost to an undisclosed liability. Amp's move is not a charity. It is a customer acquisition strategy dressed as developer love. The question is not whether developers save $20 per month. The question is what they surrender in data, lock-in, and operational risk. Based on my audit experience, I treat every zero-fee announcement as a deferred invoice. The invoice always arrives. Sometimes it is paid in tokens. Sometimes it is paid in dependency. Amp's Hobby tier is a contract with missing terms. And missing terms are where the risk lives.
Amp is not a model lab. It does not train frontier models. It is an orchestration layer. The agent reads code, modifies code, executes commands, and runs tests. That is the standard agent loop. The technical differentiation, if any, comes from Sourcegraph's code index and Orb's remote execution environment. Sourcegraph built a business on code search across repositories. Orb provides cloud sandboxes that keep running after your laptop closes. Amp combines these into a coding agent that can work in parallel. The market context is brutal. Cursor, GitHub Copilot, Windsurf, Claude Code, OpenAI Codex, Cline, Roo Code, OpenHands, and Aider all compete for the same developer attention. Most charge subscription fees. Some are open source. Some are bundled with model subscriptions. Amp's previous pricing was $20 per month plus a BYOK token fee. Users still paid their model provider separately. That double-charge was always a friction point. The new model is free platform access, bring your own key, pay for Orb usage, and likely pay for enterprise features. This is classic open core. Give away the orchestration layer. Monetize the cloud and compliance. The strategy is rational. The execution risk is high. Hype evaporates; receipts remain.
Let me be precise about what Amp is and what it is not. Amp does not innovate at the model architecture layer. It does not announce a new attention mechanism, a new training method, or a new tokenizer. It packages third-party models into a workflow. That is engineering innovation, not scientific innovation. The distinction matters for valuation. A model lab can defend a moat through compute, data, and talent. An orchestration layer defends through switching costs, integrations, and context management. Amp's claimed support for OpenRouter, Amazon Bedrock, Azure Foundry, Ollama Cloud, and custom model endpoints suggests a model-agnostic posture. That is good for users. It is bad for margin. When you do not own the model, you do not own the pricing power. You are a reseller of compute or a router. Routers are commoditized quickly.
The Orb infrastructure is more interesting. Orb gives each task a remote cloud environment. The agent can continue working when your computer is off. It can run multiple tasks in parallel. That is a real workflow change. But it also introduces a cost center. Cloud sandboxes are not free. They consume CPU, memory, storage, and network. If Amp charges only for Orb usage, then the free Hobby tier must limit Orb access. Otherwise, the unit economics break. The undisclosed limits are likely concurrent tasks, runtime hours, context window size, and model access. In my experience auditing SaaS pricing, the free tier is a funnel. The funnel is designed to convert users to paid Orb usage or enterprise contracts. The free tier is not a product. It is a sample.
Now consider the BYOK token fee elimination. Amp previously charged a fee on top of your model costs. That was a tax on using your own key. Removing it signals that the fee was causing churn. Developers hate double-charging. They also hate being locked into a single model provider. BYOK is a compliance feature for enterprises. A bank or a healthcare company cannot send code to a third-party model API without a data processing agreement. BYOK with Bedrock or Azure Foundry allows them to use their own cloud contracts. Ollama Cloud allows private model hosting. This is not developer love. This is enterprise sales enablement. Amp is building a bridge to regulated industries. That is smart. It is also expensive to support. Every model endpoint has different APIs, authentication, rate limits, and failure modes. The support burden grows with the matrix.
The hidden risk is the ChatGPT subscription integration. The parsed source mentions "using ChatGPT subscription to access Amp" without explaining the mechanism. This is a critical omission. If Amp is using OpenAI's Codex OAuth login or some unofficial passthrough, then it is exposed to terms-of-service risk. OpenAI can change its policies overnight. It can throttle third-party access. It can require enterprise agreements. Anthropic has similar control over Claude Code. If Amp's free tier depends on a subscription passthrough that violates provider terms, then the entire funnel can be shut down with a single API policy update. That is not a theoretical risk. It is a known pattern. In 2023, several third-party clients lost access to model APIs after providers changed their terms. The same can happen here.
Let me quantify the competitive pressure. Cursor charges $20 per month for Pro. GitHub Copilot charges $10 to $19 per month. Windsurf has a free tier and paid tiers. Claude Code is bundled with Anthropic's Max plans. OpenAI Codex is bundled with ChatGPT Plus, Pro, and Team. Cline and Roo Code are open-source extensions that are free, but you pay model costs. Amp's new free tier competes directly with Cline and Roo Code on price. It differentiates through Orb and Sourcegraph's code index. The question is whether developers will pay for Orb when they can run local agents for free. The answer depends on parallel execution and remote runtime. Some developers will. Most will not. The free tier will attract hobbyists. Hobbyists do not generate revenue. They generate support tickets.
The enterprise opportunity is real. Sourcegraph already sells to enterprises. Amp can inherit that sales motion. Enterprise features like SSO, audit logs, permissions, and private deployment are not free. They are high-margin. The free tier is a lead generation tool for enterprise. That is the only coherent business model. If Amp tries to monetize individual developers through Orb usage alone, it will struggle. Individual developers are price-sensitive. They will switch to a local agent if Orb costs exceed their perceived value. The switching cost is low. The code index is the only sticky asset. If Amp can index your entire monorepo and provide better context than a local agent, then it has a moat. But that requires uploading code to Amp's servers or connecting to your Sourcegraph instance. Data governance becomes the gate.
Security is the elephant in the room. A coding agent that executes commands has the same permissions as a developer. It can read secrets, modify production configurations, and exfiltrate data. Orb runs in a cloud sandbox, but the agent still needs access to your repository. If the sandbox is not properly isolated, a prompt injection attack could turn the agent into a remote access trojan. The source article does not discuss the permission model. It does not mention sandbox technology. It does not explain how secrets are stored. These are not minor details. They are the difference between a toy and an enterprise tool. In my 2020 DeFi rug pull investigation, I traced malicious contract interactions by following on-chain evidence. The same forensic mindset applies here. You cannot audit what you cannot see. Amp's security posture is a black box. Until it publishes a threat model, enterprises should assume the worst.
Let me apply game theory. The coding agent market is a prisoner's dilemma on pricing. Every player wants to charge subscription fees. But open-source alternatives and BYOK tools create a price ceiling. When one player goes free, others must respond. Cursor might introduce a free tier. Copilot might reduce prices. Windsurf might bundle more usage. The result is margin compression across the category. Amp's move is an attempt to preempt that compression by becoming the free default. But free is not a strategy. It is a tactic. The winner will be the player with the lowest customer acquisition cost and the highest enterprise conversion rate. Sourcegraph has brand recognition, but it does not have the distribution of Microsoft or OpenAI. GitHub Copilot is installed by default in VS Code. OpenAI Codex is one click away for ChatGPT users. Amp must convince developers to install another tool. That is a high-friction ask.
Now consider the blockchain developer angle. Smart contract development is unforgiving. A bug can cost millions. Coding agents are increasingly used to write Solidity, Rust, and Move code. They can also be used to audit code. But an agent that executes commands and installs dependencies is a supply-chain risk. If an agent pulls a malicious package from npm or crates.io, it can compromise the developer's machine. If it runs in a cloud sandbox, it can compromise the cloud account. Blockchain developers often manage private keys and deploy contracts. An agent with access to a wallet or deployment script is a catastrophic liability. Amp's free tier will attract blockchain hobbyists. Some will connect their wallets. Some will lose funds. The source article does not mention wallet integration, but any coding agent that can run forge script or anchor deploy is in the blast radius. This is not hypothetical. In 2024, several AI coding assistants were found to leak API keys through prompt injection. The pattern repeats.
Let me address the BYOK token fee more directly. The fee was likely a percentage of model usage or a flat monthly charge. Removing it means Amp forgoes revenue from model resale. Why would a company do that? Because the model resale business is terrible. Model providers charge per token. The margins are thin. Price wars are constant. Users compare token costs across providers. A reseller cannot compete with the provider's own distribution. By moving to BYOK, Amp shifts the model cost to the user and focuses on its own value-add. That is a rational retreat. It also removes the conflict of interest where Amp might favor a model that pays higher margins. BYOK aligns incentives. The user chooses the best model. Amp provides the orchestration. This is the correct architecture for a multi-model world.
But the free Hobby tier creates a new problem. If users bring their own keys and use the free tier, Amp earns nothing. The only upsell is Orb. Orb is a cloud environment. Cloud costs are linear with usage. If a user runs long tasks, Amp loses money. The free tier must therefore be throttled. The throttle might be a queue, a time limit, or a context limit. The source article does not disclose these limits. That is a material omission. In my regulatory compliance audits, I have seen companies hide throttles in acceptable use policies. The result is user frustration and churn. Amp should publish the limits. Transparency builds trust. Opacity invites suspicion.
Let me also examine the Sourcegraph spin-out. Sourcegraph started as a code search company. It raised significant venture capital. It pivoted to code intelligence and then to coding agents. The spin-out of Amp suggests a need for speed and focus. Sourcegraph's legacy business may have different economics. Code search is a mature market. Coding agents are a growth market. By separating, Amp can raise capital at a higher valuation and offer different compensation to employees. That is standard corporate strategy. But it also means Amp must stand on its own. It cannot rely on Sourcegraph's balance sheet forever. The free tier will burn cash. The question is how long the runway lasts. The source article does not provide financials. That is expected for a private company. But it means the free tier is a bet on future fundraising. If the market turns, the free tier could disappear overnight. Developers should treat free tiers as temporary. Hype evaporates; receipts remain.
Let me consider the actual user experience. A developer installs Amp. They connect their OpenRouter API key. They open a repository. They ask the agent to fix a bug. The agent reads files, edits code, runs tests, and submits a pull request. That is the promise. The reality is messier. The agent might hallucinate. It might break the build. It might consume tokens unpredictably. The cost of a bad agent run is not just the API fee. It is the developer's time. Time is the real currency. If Amp's free tier saves $20 per month but costs two hours of debugging, the developer loses. The value proposition must be measured in hours saved, not dollars saved. The source article focuses on pricing. It does not provide benchmarking data. I have audited productivity claims before. They are usually inflated. The only metric that matters is cycle time. Amp has not published cycle time data. That is a red flag.
Let me apply the forensic code verification standard. I want to see the actual contract for the free tier. I want to see the Orb pricing page. I want to see the data retention policy. I want to see the security whitepaper. None of these are in the source article. That is not the article's fault. It is a news brief. But as an auditor, I cannot assign a confident rating without primary sources. My confidence in the commercial analysis is B-minus. My confidence in the technical analysis is C-plus. My confidence in the competitive analysis is B. The overall picture is clear enough: Amp is making an aggressive land grab. The details are hidden. The risk is real.
Let me think about the bull case. Bulls will say that free BYOK is a masterstroke. It commoditizes the model layer and forces competitors to compete on agent quality. Amp's code index and Orb environment are genuine differentiators. Sourcegraph's enterprise relationships provide a ready customer base. The coding agent market is growing fast enough to support multiple winners. Free is a customer acquisition cost. If Amp converts even a small percentage of free users to enterprise contracts, the math works. The bulls also say that model providers will not cut off third-party access because they want distribution. OpenAI and Anthropic need developers to use their models. Blocking Amp would reduce token consumption. That is a fair point. The model providers are in a land grab too. They want their models in every tool. So the ToS risk may be overstated.
But the bull case has a blind spot. It assumes that model providers will remain neutral platforms. History shows otherwise. Apple blocked third-party app stores. Google restricted Android forks. Microsoft bundled Internet Explorer. Platform owners eventually integrate and compete. OpenAI already offers Codex. Anthropic already offers Claude Code. They have every incentive to make their own agents the best. They can degrade third-party access, raise prices, or bundle features. Amp's free tier depends on their goodwill. That is not a moat. That is a license to operate at the pleasure of a competitor. The bull case also ignores the cost of support. Free users are expensive. They generate tickets, abuse the system, and consume resources. The conversion rate from free to paid in developer tools is typically low. Amp will need a high volume of free users to find a small number of enterprise buyers. That is a numbers game. The numbers are not disclosed.
Let me also consider the regulatory angle. In the EU, the AI Act imposes obligations on general-purpose AI systems. Coding agents that execute code may be classified as high-risk if they are used in critical infrastructure. The source article does not mention compliance. Amp operates in the US and Europe. If it processes code from regulated industries, it must comply with data protection laws. The BYOK model helps because the user's cloud provider handles data residency. But Amp still processes context. It still sees the code. That is a data processing activity. Amp needs a legal basis. It needs a data processing agreement. It needs to respect the right to erasure. These are not optional. They are the cost of doing business in Europe. The free tier complicates this. If users are not paying, Amp may not have the resources to provide enterprise-grade compliance. That is a risk for institutional adoption. In my 2025 MiCA audit, I found that only one of three major exchanges met the technical standards for proof-of-reserve. The same pattern applies here. Many tools claim compliance. Few can prove it. Amp has not published a compliance package. That is a gap.
Let me synthesize the core insight. The core insight is that Amp's fee removal is not a pricing decision. It is a structural response to the commoditization of the coding agent layer. The model layer is consolidating around OpenAI, Anthropic, and Google. The agent layer is fragmenting. The only defensible positions are distribution, enterprise compliance, and unique context. Amp has some of each, but none at scale. The free tier is an attempt to buy distribution. The BYOK strategy is an attempt to avoid competing on model cost. The Orb environment is an attempt to create a usage-based revenue stream. The enterprise features are the real business. The free tier is a marketing expense. The risk is that the marketing expense balloons before the enterprise revenue arrives. If that happens, Amp will either raise prices, cut the free tier, or get acquired. Developers should plan accordingly.
The bulls are right about one thing. The coding agent market is not a winner-take-all market. Developers use multiple tools. They will use Amp for some tasks and Cursor for others. The free tier lowers the barrier to trial. That is good for competition. It forces incumbents to improve. It also validates BYOK as a standard. In the long run, every serious coding agent will support BYOK. Amp is simply early. The bulls also correctly identify that Sourcegraph's code index is a real asset. If Amp can index a monorepo and provide better context than a local agent, it can charge for that. The free tier is a way to get users to connect their repositories. Once connected, the switching cost rises. That is a classic land-and-expand motion. The bulls are not wrong about the strategy. They are wrong about the timeline. The strategy will take years. The market may not wait. The model providers may not wait. The free tier may not survive. The contrarian insight is that Amp's biggest risk is not competition. It is its own cost structure. Free BYOK is cheap to offer. Orb is not. If Orb usage grows faster than enterprise revenue, Amp will be forced to throttle or charge. That will break the free tier promise. The bull case assumes infinite patience from investors. That is not a safe assumption.
The next 18 months will determine whether Amp becomes the Switzerland of coding agents or a footnote in the consolidation wars. Watch three metrics. First, Orb's published pricing and free limits. Second, enterprise SSO and audit log availability. Third, whether OpenAI or Anthropic change their subscription passthrough policies. If Amp publishes transparent limits and lands enterprise contracts, the free tier is a sustainable funnel. If it stays opaque and burns cash, the free tier is a temporary subsidy. In my audit practice, I do not trust announcements. I trust receipts. The receipts will show up in pricing pages, security whitepapers, and enterprise case studies. Until then, developers should enjoy the free tier but keep their exit plan ready. The ledger balances do not lie; they only wait.