NextFin

Amp Puts AI Agents to Work Beyond Code, Signaling the Next Phase of the Developer-Tool Race

Summarized by NextFin AI
  • Amp, the AI coding agent spun out of Sourcegraph, is expanding beyond software engineering into the full product-building lifecycle including marketing, design, and documentation using OpenAI's newest models.
  • Some Amp users now consume 10 billion tokens a week, and the company's 20-person team runs an agent devoted to marketing instead of hiring a human for it.
  • Cursor crossed $2 billion in annualized revenue in Q1 2026 and forecasts $6 billion by year-end, while Cursor, Cognition, Replit, and Lovable raised nearly $2 billion in five months.
  • Amp's pricing is shifting from seat licenses to token pass-through, betting that removing the mental accounting of counting tokens will unlock greater agent autonomy across workflows.

NextFin News - Amp, the AI coding agent spun out of Sourcegraph, is pushing its agents past software engineering into the full product-building lifecycle - marketing, design, documentation, planning - using OpenAI's newest models, a move that reframes the coding-agent market as a race for the whole software-creation workflow rather than the code editor alone. Quinn Slack, Amp's CEO, told OpenAI that some Amp users now consume 10 billion tokens a week, and that the company's own 20-person team runs an agent "devoted" to marketing instead of hiring someone for it.

The customer story, published by OpenAI on October 14, 2026, shows Amp integrating Sign in with ChatGPT, GPT-6 Astra, the Ultrafast service tier, and GPT-Live-1 - a stack that lets builders hand agents larger tasks without watching every token. It lands as AI-native developer tools have become a category of their own, with Cursor, Claude Code, GitHub Copilot, Devin, Replit, and Lovable raising nearly $2 billion among them in a five-month stretch, according to coverage of the spinout.

The Deal: What Amp Is Doing With OpenAI's Models

Amp chooses and recommends models based on what works best for its users, and OpenAI models account for the majority of tokens used across the product. Slack's central argument is behavioral economics as much as technology: when builders count every token, they underuse the agent. "Sign in with ChatGPT means that people can have that feeling where they're not counting every single token," he said. "And once you actually let the agent go and have a longer leash, then you realize just how much it can do."

Four OpenAI capabilities anchor the integration. Sign in with ChatGPT lowers the cost of building by giving Amp users access to more tokens and to frontier models including GPT-6 Astra, which OpenAI released in September 2026. GPT-6 Astra handles the writing and design work - Amp's 20-person team uses it to produce documentation and design pages that meet the company's standards. GPT-Live-1 powers "Calling Puck," a real-time voice interface to Puck, Amp's built-in conversational assistant that launches and coordinates other agents; users can talk to Puck from a watch. And Ultrafast mode, which OpenAI says delivers up to 8x faster speeds than its standard service tier, is being tested with GPT-6 Astra inside Amp's workflows.

Slack's own experiments illustrate how far the models have moved. He fed GPT-6 Astra the plans for his house and used it to build a 3D model in Blender, which he and his children then turned into games. "And that was not possible just a couple months ago," he said.

"Our users really appreciate the consistency with which OpenAI models have been getting better and better."

- Quinn Slack, CEO and member of technical staff, Amp

The transparency point matters because Amp is betting its users will let an agent run more of their workflow only if they understand the guardrails. "OpenAI is very transparent, and when there are safeguards applied, users know and they understand. That makes it really easy to use OpenAI models. You can do all of your work in Amp with OpenAI models and be really happy," Slack said. For a company whose heaviest users burn 10 billion tokens a week, that trust is the bottleneck, not the model.

Why This Matters: The Agent Race Is Moving Beyond the Code Editor

The obvious read of the OpenAI post is that another coding agent uses OpenAI models. The more important read is what Amp is choosing to automate next. Amp is not just writing code; it is trying to absorb the work before and after coding - deciding what to build, writing documentation, designing pages, running marketing, and managing other agents through Puck. "Amp wants agents to help with all the work that goes into building and growing a product," the company said.

That expansion is structural, not cyclical. A cyclical shift would be a temporary surge in agent usage that reverts when budgets tighten. What Amp is describing is a change in the unit of automation: the agent is no longer a pair programmer inside an IDE, it is a production teammate that owns a slice of the product lifecycle. Amp's marketing example is the cleanest evidence - "We don't have someone devoted to it, but we have an agent devoted to it," Slack said. When an agent replaces a hiring decision rather than a keystroke, the displacement is durable.

The token numbers back the behavioral claim. At 10 billion tokens a week for some users, Amp's heaviest customers are running agent fleets, not chat sessions. That volume only makes sense if the agent is doing work that previously required humans across multiple functions - research, drafting, design iteration, coordination. The constraint is no longer whether the model can do the task; it is whether the builder trusts the agent with a long enough leash.

The timing is also deliberate. Amp became an independent company in December 2025, when Sourcegraph and Amp announced their separation into two businesses: Sourcegraph focused on enterprise code search and code understanding, and Amp Inc. built to explore the frontier of coding agents. Dan Adler stepped into the CEO role at Sourcegraph; Slack and co-founder Beyang Liu founded Amp Inc. with the rest of the Amp team. The same investors - Craft, Redpoint, Sequoia, Goldcrest, and Andreessen Horowitz - sit on both boards, but the two companies now optimize for different cadences. Code search has a steady, infrastructure-heavy rhythm; agents move on cycles shaped by model releases and interaction models that change month to month.

The Economics: Token Pass-Through vs. the Seat-License Model

Amp's bet on Sign in with ChatGPT also signals a pricing shift. Traditional developer tools sell seats - a flat monthly fee per user. Agent-native tools increasingly pass token costs through at cost, with Amp having launched an ad-supported free tier in October 2025 and later moving to pass-through API billing with zero markup alongside flat monthly subscriptions. The logic is that seat pricing caps usage: a user on a fixed plan stops when they hit the limit. Token pass-through with generous included allowances removes the mental accounting that Slack says holds users back.

This creates a different competitive axis. The winner in the next phase may not be the model with the best coding benchmark; it may be the environment that makes token consumption feel free enough that users hand it bigger problems. OpenAI's Ultrafast tier - up to 8x faster than standard - matters here because speed changes the interaction model. "People have not seen a model that's this good, that's this fast," Slack said. Faster responses let users juggle fewer agent tasks, or give agents more time to check their own work before returning a result.

The counter-pressure is that token pass-through makes costs variable and opaque to the buyer. A seat is a known line item; a token bill scales with usage and can spike. That is why Amp pairs the integration with safeguards and transparency. For the CFO now sitting in on AI-tool decisions, the question is not the headline model capability - it is whether the bill is predictable enough to budget.

The Competitive Landscape: Who Is Chasing the Same Ground

Amp's expansion beyond code is not happening in isolation. GitHub Copilot, with 4.7 million paid subscribers as of early 2026 and deployment at roughly 90% of the Fortune 100, has been extending past code completion into analysis and document-based work. Cursor crossed $2 billion in annualized revenue in the first quarter of 2026 and is forecasting $6 billion by year-end, according to market coverage. Cognition acquired Windsurf for roughly $250 million and is consolidating Devin and Windsurf under one roof. Replit and Lovable are raising capital to turn natural-language prompts into deployed applications.

The common thread: every major player is trying to own more of the workflow that surrounds coding. Amp's differentiation is that it started from Sourcegraph's code-search infrastructure and is now an independent company built to move at agent speed. The spinout was a tacit acknowledgment that the AI-native workflow no longer fits neatly inside traditional developer-tool companies. "We've realized over the last few months that the two products have totally different distribution engines and target audiences," Slack and Adler wrote in their joint announcement. "Amp needs to remain on the frontier, exploring what's possible with models and tools and agents by building for the devs and teams who want to be a year ahead of everyone else."

The capital behind the category underscores how much is at stake. Cursor, Cognition, Replit, and Lovable raised nearly $2 billion among them in a five-month stretch in 2026, according to coverage of the Sourcegraph-Amp separation. That money is not funding better autocomplete; it is funding the race to own the layer between the developer and the finished product.

"People have not seen a model that's this good, that's this fast."

- Quinn Slack, CEO and member of technical staff, Amp

The Counter-Thesis: Agents Beyond Code May Be Harder Than They Look

The strongest argument against Amp's thesis is that code is the easy part. Code has a compiler, tests, and a clear definition of correctness. Marketing copy, design, and product decisions have no compiler - success is fuzzy, subjective, and slow to measure. An agent that can write a function and have it pass a test suite is operating in a closed loop; an agent that runs marketing is operating in an open loop where feedback arrives weeks later and attribution is murky.

There is also the question of whether "beyond code" is a distraction. The most valuable automation may remain inside the development loop, where engineers already spend their days and where the willingness to pay is proven. Cursor's revenue trajectory - $2 billion annualized and climbing, by market estimates - suggests the code-first wedge is still far from exhausted. If Amp's bet on pre- and post-code work dilutes its focus on the core coding experience, it risks losing the users who matter most to a competitor that stays narrow and deep.

The falsifying signal is concrete: if Amp's heaviest users - the ones burning 10 billion tokens a week - are still spending the majority of those tokens on code generation and review rather than on design, documentation, marketing, and agent coordination a year from now, then the "beyond code" thesis is a feature expansion, not a regime change. Token telemetry, not roadmap announcements, will settle it.

What's Next: Three Horizons to Watch

Short term (next quarter): Watch whether the Sign in with ChatGPT integration moves Amp's token consumption per user upward. If the "longer leash" effect is real, average tokens per active user should rise noticeably after the integration ships, because the friction of counting is removed. The OpenAI post frames the integration as giving users "access to more tokens," which is a direct attempt to lift that ceiling.

Medium term (6-12 months): Watch the token mix. The share of tokens spent outside the code editor - on design, docs, marketing, planning - is the metric that validates or refutes the beyond-code thesis. A rising share means Amp is becoming a product-building agent; a flat share means code remains the core use case. This is the single number that separates a regime change from a feature list.

Long term (structural): The pricing model that wins. If token pass-through with generous allowances becomes the category standard, seat-based developer tools will face margin pressure as users migrate to usage-based agents. If opaque token bills scare buyers back to predictable seats, the subscription and ad-supported models Amp has been testing will prove more durable. Either way, the winner is the company that makes the economics legible to the person signing the check.

Base case: agents keep expanding outward from code, but code remains the largest token category for at least the next year. Upside case: the friction removal from Sign in with ChatGPT unlocks a step-change in agent autonomy, and Amp's whole-product agent becomes the default way small teams build. Downside case: the beyond-code work proves too fuzzy to automate reliably, and Amp retrenches to a best-in-class coding environment.

The central judgment: Amp's OpenAI integration is less about any single model and more about who owns the workflow around the model. The company betting correctly on where the agent's job begins and ends - not on whose benchmark is highest - will define the next generation of developer tools. The code editor was the first battlefield; the product lifecycle is the second, and it is already open.

Explore more exclusive insights at nextfin.ai.

Insights

What is Amp's core business shift?

How does Amp use OpenAI models?

Why move beyond code editors?

Why choose token pass-through billing?

Who are Amp's main market competitors?

How much funding did rivals raise?

What risks face non-code agents?

Why is user trust a key bottleneck?

What tasks does GPT-6 Astra handle?

How does Puck coordinate other agents?

What changed after Sourcegraph split?

Will seat licenses survive AI tools?

Metrics validating beyond-code work?

How fast is OpenAI Ultrafast tier?

Why do users count every token?

What is the agent race next phase?

Can agents replace marketing hires?

Who defines workflow ownership next?

How does Amp handle model safeguards?

What is Amp's worst downside case?

What is Amp's core business model?

How did Amp separate from Sourcegraph?

Which OpenAI models does Amp use?

Why do users count every token?

Why use Sign in with ChatGPT integration?

How does token pass-through pricing work?

Who are Amp's main competitors?

What defines the AI agent race?

Why is code easier than design work?

What risks face beyond-code agents?

Search, ask or analyze anything...