From Aryan Vatsa | Product & Market Analysis
Factory AI Review: What Droids Do, What They Cost, What the $5B Pitch Rests On
On this page
Factory AI went from a $1.5 billion valuation in April 2026 to $5 billion in September. Its pitch is that Droids, its coding agents, work with any model, so your team never bets on one lab. The documentation backs that claim more firmly than most vendor marketing does. The cost-saving and customer numbers attached to it are thinner than the headlines suggest.
Key takeaways
- Factory AI's model-agnostic claim is real in the docs. Its models page lists Anthropic, OpenAI, Google, xAI and open-weight models under one rate-limit system, which neither Claude Code nor Codex offers natively.
- The plan price is not the real price. Factory bills each model through a multiplier that runs from 0.04x to 12x, so the same $20 Pro plan buys roughly 300 times more work on the cheapest model than on the most expensive.
- The 63% router saving is one week of production data against a list price. Factory compares routed sessions with what the work would have cost at unnamed frontier rates, and publishes no production quality measure alongside it.
- The strongest reason to buy Factory is governance across vendors. Autonomy caps, model access controls and on-premise options matter more to a regulated buyer than any benchmark lead, and those features sit in the Business and Enterprise tiers.
What Factory AI actually sells
Factory is a coding agent company founded in 2023 by Matan Grinberg and Eno Reyes. Its agents are called Droids. You give a Droid a task, and its documentation says it "plans, implements, tests, and reviews changes end to end while approval workflows keep you in control".
That description fits almost every coding agent on the market in 2026. What differs is where the Droid runs and which model sits underneath it. Factory ships the same agent through a terminal CLI, a desktop app, IDE integrations and a headless mode for pipelines.
For a head of delivery, the useful question is not whether Droids can write code. All the serious agents can. The question is which surfaces your team will actually use, and who controls what the agent is allowed to touch.
The surfaces Droids run on
The CLI is the core product. It supports MCP tools, hooks, plugins, skills and custom subagents, which Factory calls Custom Droids. Missions, a mode for long multi-step work, coordinates several agents over hours or days. Factory's documentation states that Missions require the highest autonomy level, which tells you how much trust the feature assumes.
Delegations lets your team hand work to a Droid from Slack, Linear, Jira and Microsoft Teams. The docs list it as Private Preview. Pull requests opened that way are authored by a bot account through Factory's GitHub App, so your audit trail shows the agent, not the person who asked.
Droid Exec and code review in CI
Droid Exec is the headless mode. It runs a task once and exits, which is what you want inside a CI pipeline. Factory's docs say it is read-only by default and needs an explicit flag before it can change files. That is the correct default, and it matches the choice both Anthropic and OpenAI made.
Automated pull request review ships as a packaged GitHub Action that runs on runners you control. It can approve a pull request when it finds no issues. I would switch that setting off on day one. An agent that can approve code is a second approver your branch protection rules were not designed for.
Factory AI funding: from $1.5 billion to $5 billion in five months
Factory announced a $150 million Series C led by Khosla Ventures on 16 April 2026. Sequoia Capital, Blackstone, Insight Partners, Evantic Capital, 20VC, NEA and Mantis VC joined. The company put its valuation at $1.5 billion.
Five months later, on 15 September 2026, Factory said it had raised $200 million at a $5 billion valuation, bringing total funding to over $400 million. The announcement names no lead investor. Most third-party reviews still describe the April round as the latest, so check the date on anything you read about Factory.
The named customers, and how the list moved
The April announcement said Droids are used daily by "hundreds of thousands of developers" at Nvidia, Adobe, EY, Palo Alto Networks and Adyen. The September announcement named Nvidia, Blackstone, Royal Bank of Canada, Palo Alto Networks, Adobe and T-Mobile.
Two details are worth noticing. Blackstone appears as both an investor and a customer, which is common and is still a disclosure you should weigh. EY dropped from the headline list between April and September. That may mean nothing, and it is still a question to ask in a reference call.
Factory's customers page states that "every number on this page comes from a published customer story". Some of those numbers are specific and modest, such as Groq's 3x faster feature development on medium complexity tasks. Others are not attached to any named customer. The enterprise page cites "7x faster" feature delivery and a "96.1% reduction" in migration time with no source.
The page also contradicts itself once. A 40% cut in incident response time is credited to Tilt in one grid and to Empower in a case study title on the same site. None of this is unusual for a vendor page. It is exactly why no figure from it should go into your business case unverified.
Factory AI pricing, read line by line
Factory changed its pricing model during 2026. Older reviews describe monthly token allowances. The current pricing page and documentation describe rolling rate limits over 5-hour, 7-day and 30-day windows instead, with no token count published.
| Plan | Price | Usage and notable inclusions |
|---|---|---|
| Pro | $20 per month | Base rolling rate limits |
| Plus | $100 per month | About 5x Pro usage, Factory-managed cloud computers |
| Max | $200 per month | About 10x Pro usage, early access to features |
| Teams | $60 per team plus $40 per seat, monthly | Pro limits per seat, up to 10 seats, central billing, no SSO listed |
| Business | Custom | SSO, SAML and SCIM, zero data retention, audit logs, model and autonomy controls |
| Enterprise | Custom | Dedicated compute, on-premise options, data residency, full admin controls |
The marketing page and the docs disagree on one point: the docs list managed cloud computers under Pro, the marketing page lists them only from Plus. Read on 10 October 2026.
The multiplier is the real Factory AI price
Every model on Factory carries a multiplier against your plan's usage. On the models page, Claude Opus 5.5 bills at 1.6x, Claude Sonnet 5.5 at 0.8x and GPT-6.1 Sol at 0.8x. GPT-5.5 Pro bills at 12x, while GPT-6 Luna bills at 0.04x.
That means "about 10x the usage of Pro" on the Max plan is a statement about units, not about work. A team running the most expensive model exhausts the same plan roughly 300 times faster than a team on the cheapest, if each task used the same tokens. Several multipliers are promotional, with GPT-5.6 Sol rising from 1.6x to 2x after 22 November 2026.
This structure is the same credit abstraction we examined in why credits and tokens have made AI bills hard to forecast. It is not dishonest. It does move the forecasting burden from Factory to you.
Droid Core, BYOK and the Teams gap
When standard usage runs out, individual plans fall back to Droid Core, a pool of open-weight models with separate limits. You can also buy prepaid Extra Usage from $10. Bring-your-own-key is free "up to an allowance" that Factory does not size publicly.
The bigger catch is the Teams plan. A 10-seat team costs $460 a month, and the docs comparison table shows no SSO, governance, data controls or deployment options for it. Those sit in Business and Enterprise at custom prices. For a regulated buyer, the published price is a starting point for a sales call, not a budget line.
The model-agnostic bet, and the Factory Router claim behind it
Factory's case rests on one argument. Model leadership changes every quarter, so the agent harness should not be tied to one lab. The September announcement frames this as giving enterprises control over "which models it uses, and where it runs."
The docs support it. Factory lists native access to Anthropic, OpenAI, Google and xAI models plus open-weight models, and BYOK covers Bedrock, OpenRouter-style endpoints and local models through Ollama or LM Studio. In September 2026 Factory also joined both Anthropic's and OpenAI's marketplaces, selling through the two labs it hedges against.
The weak point sits in Factory's own fine print. Its BYOK page warns that "only Anthropic and OpenAI models accessed via their official APIs are fully tested and benchmarked." Model-agnostic in the catalogue does not mean model-agnostic in quality. That gap is the subject of what routing to a cheaper model costs in output quality.
What the 63% router saving measures
Factory Router picks a model per task and escalates to a stronger one when needed. Factory says sessions using it "cost 63% less in aggregate than the same workload would have cost at published frontier-model rates." That figure is one week, ending 13 September 2026.
Three things qualify it. The baseline is a counterfactual price, not a measured run, and the frontier model is not named. Over the surrounding two weeks the aggregate saving was 60%, not 63%. Factory also attributes part of the gain to "lower inference prices," which a router does not create.
Most importantly, no production quality measure accompanies the saving. A router that sends more work to cheap models will always save money. Whether the work still passes review is the number that matters, and it is not published for production use.
What the June benchmark actually held constant
The June launch post is more careful. On Terminal-Bench 2, routed runs reached 99% of Claude Opus 4.7's pass rate at 20% lower cost. On Factory's own Legacy-Bench, they reached 96% at 25% lower cost.
The same post shows the trade-off honestly. The most aggressive routing setting cut cost to 56% of Opus on Terminal-Bench 2 and dropped the pass rate to 81%. On Legacy-Bench, 30% of the cost came with a 49% pass rate. Factory published those numbers itself, and it deserves credit for that.
Factory Droids vs Claude Code vs Codex on the same tasks
The three products now solve the same job in similar ways. Each has a planning step, a graded permission system, a headless mode for CI and a GitHub integration. The real differences are model reach, governance and price structure. A wider adoption comparison sits in the Claude Code, Cursor and Copilot adoption verdict.
| Job | Factory Droid | Claude Code | OpenAI Codex |
|---|---|---|---|
| Plan before editing | Spec Mode, read-only, can use a separate planning model | Plan mode, read-only until you approve | Read-only sandbox or approval mode |
| Permission levels | Off, Low, Medium, High, plus allow, ask and block rules | default, acceptEdits, plan, auto, dontAsk, bypassPermissions | Ask for approval, Approve for me, Full access |
| Headless CI run | droid exec, read-only by default | claude -p, with turn and budget caps | codex exec, read-only sandbox by default |
| Models natively | Anthropic, OpenAI, Google, xAI, open-weight | Claude models, via Anthropic or major clouds | OpenAI models, plus local and custom providers by config |
| Entry price | $20 per month | $20 per month (Claude Pro, monthly) | $20 per month (ChatGPT Plus) |
| Where the rival wins | Cross-vendor admin caps | Simplest setup if you only use Claude | Bundled with ChatGPT seats you may already own |
Autonomy controls are Factory's clearest edge
All three tools let a user approve or skip actions. Factory goes further for administrators. Enterprise admins can set a default autonomy level and a maximum level that members cannot exceed. A block rule rejects a command with no approval path at all.
That matters for the scenario most security teams fear, which is a developer switching the agent to full access to save time. Similar controls exist elsewhere in managed settings. Factory makes them the centre of its enterprise pitch, which is where they belong.
Where Claude Code and Codex win outright
If your team has standardised on one lab, the native tool is simpler and often cheaper. A 10-seat Claude Team on standard seats costs $250 a month billed monthly, against $460 for Factory Teams. ChatGPT Business includes Codex at $25 per user per month billed monthly, so many teams already own it.
The native tools also get new model features first, because the lab ships both the model and the harness. For a team that is happy with one vendor, I would not add Factory as a middle layer. The switching cost you are insuring against is real but modest, as we argued in what it actually costs to switch model providers.
Factory's benchmark record, and why it has aged
Factory's April announcement says it "maintained our status as #1 across the leading software development agent benchmarks." It names no benchmark and gives no score. The most specific public claim is older.
In September 2025, Factory reported a score of 58.75% on Terminal-Bench Core v0.1.1, an 80-task set. In the same post it said Droid with Opus beat Claude Code with Opus at 43.2%, and Droid with GPT-5 beat Codex CLI at 42.8%.
That result is now three benchmark versions old. The public Terminal-Bench leaderboard currently shows version 4.0. A year-old score on a superseded test is not evidence about the product you would buy today.
The 2025 result does show something durable. The same model scored very differently in different harnesses, which supports Factory's argument that the agent layer adds value. It also means any benchmark you see is a test of one harness, one model and one configuration on one date.
The Groq case study illustrates the problem. Its 5x faster claim compares Droid running Kimi K2 on Groq hardware with GPT-5 on Codex and Claude Code, on an internal Groq benchmark. Groq sells that inference. It is a fair speed claim about Groq's chips and a weak claim about Droids. Our wider critique of these comparisons is in the three scoreboards AI coding tools are judged on.
Where this Factory AI review is weakest
This review compares documentation, announcements and published prices. It does not report a hands-on test of Droids on our own repository. A documented capability and a capability that works on your codebase are different things, and we have only assessed the first.
Most of the numbers here come from Factory itself. A company announcement is the primary source for its own funding and pricing, and that is why we cite it. It is a weak source for performance claims, and we found no independent measurement of Droids' output quality from 2026.
The strongest counter-case is real. Model rankings changed several times in 2026, and an enterprise with many teams may genuinely want different models for different work. For that buyer, one governance layer across vendors is worth a premium, and our scepticism about the router does not touch that argument. The general risk of a single-model bet is set out in the build versus buy decision for coding agents.
Finally, pricing is moving fast. Factory changed its billing model within 2026, several multipliers are promotional, and it bought Mentlio, a token allocation company, on 9 October 2026. Treat every price in this post as correct on 10 October 2026 and not after.
Who should buy Factory AI, and who should skip it
Buy Factory if three things are true. Your engineering organisation uses models from more than one lab, your security team needs central caps on agent autonomy, and you will pay for the Business or Enterprise tier where those controls live.
Skip it if your team is small, uses one lab and buys on the published price. In that case the native agent does the same job with one fewer vendor, one fewer data processor and one fewer bill to forecast. The Teams plan, without SSO, is the hardest tier to justify.
Whichever way you lean, do not accept the router saving as a forecast. Run it on your own work for two weeks and measure review rejections alongside cost. That is the test Factory's September numbers leave out.
Frequently asked questions
What is Factory AI?
Factory AI is a San Francisco company founded in 2023 that sells autonomous coding agents called Droids. Droids run in a terminal, a desktop app, IDEs and CI pipelines, and they can plan, edit, test and open pull requests for human approval. Factory raised $200 million at a $5 billion valuation in September 2026, after a $150 million Series C in April.
How much does Factory AI cost?
Individual plans cost $20, $100 and $200 per month, with the higher two giving about 5x and 10x the usage of Pro. Teams costs $60 per team plus $40 per seat for up to 10 seats. Business and Enterprise are custom priced. Each model also carries a usage multiplier, from 0.04x to 12x, which changes how much work a plan actually buys.
Is Factory AI better than Claude Code?
Not for every team. Factory's advantage is model choice across Anthropic, OpenAI, Google, xAI and open-weight models, plus central admin caps on agent autonomy. Claude Code is simpler and cheaper for a team that only uses Claude. Factory's benchmark lead over Claude Code dates from September 2025 on an older Terminal-Bench version, so it says little about current performance.
What models do Factory Droids support?
Factory's models page lists Anthropic Claude models, OpenAI GPT models, Google Gemini, xAI Grok and a pool of open-weight models called Droid Core, including GLM, Kimi, Qwen, DeepSeek and Mistral. You can also bring your own keys or run local models. Factory's own docs say only Anthropic and OpenAI models via official APIs are fully tested and benchmarked.
Does Factory Router really cut AI coding costs by 63%?
The 63% figure is Factory's weekly aggregate for one week ending 13 September 2026, compared with what the work would have cost at unnamed frontier list prices. Over two weeks it was 60%. No production quality measure was published with it. Factory's June benchmark tests showed 20% to 25% savings at 96% to 99% of Claude Opus 4.7's pass rate.
Which companies use Factory AI?
Factory has named Nvidia, Adobe, EY, Palo Alto Networks, Adyen, Blackstone, Royal Bank of Canada and T-Mobile in its 2026 funding announcements. Its customers page adds companies such as DoorDash, HP, Wipro, Groq and Chainguard. Blackstone is also a Factory investor. Usage figures for named customers are not disclosed, so verify scale in a reference call.
Where to start this week
Pull last month's usage from whatever coding agent your team already runs, broken down by model. Then price that same mix against Factory's multiplier table. If most of your work runs on one lab's mid-tier model, the native tool is almost certainly cheaper and you can stop here.
If the mix spans several labs, ask Factory for a two-week trial on the Business tier with the router switched on. Track two numbers per week: cost, and the share of agent pull requests your reviewers reject. A saving that raises the rejection rate is not a saving.
Related analysis
Before any agent trial, read why coding agents score lower on real repositories than on benchmarks, then compare the vendors on which AI coding assistants indemnify you.
References
- Factory, Factory raises $150M Series C, 16 April 2026. Used for the Series C, investors, April customer list and usage claim.
- Factory, Factory raises $200M at $5B valuation, 15 September 2026. Used for the $5B valuation, total funding and September customer list.
- Factory, Pricing and Individual plans documentation, read 10 October 2026. Used for plan prices, rate limits, Droid Core, Extra Usage and BYOK.
- Factory, Models documentation, read 10 October 2026. Used for all multipliers and promotional rate dates.
- Factory, Factory Router cuts inference costs by 63%, 23 September 2026, and Factory Router, 1 June 2026. Used for all router figures.
- Factory, Droid: the #1 software development agent on Terminal-Bench, 25 September 2025. Used for the Terminal-Bench Core v0.1.1 scores.
- Factory documentation on autonomy levels, Spec Mode, Droid Exec, BYOK, Delegations and security, docs.factory.com, read 10 October 2026. Factory customers and enterprise pages, factory.com, read the same day.
- Anthropic, Claude pricing and Claude Code CLI reference; OpenAI, ChatGPT and Codex pricing, all read 10 October 2026.
The weakest thing about this source base: every performance and savings figure in this post comes from Factory or its customers, and none has been independently replicated. The router saving uses a counterfactual price baseline with no production quality measure. We did not run Droids ourselves. Prices and multipliers were read on 10 October 2026, several are promotional, and Factory changed its billing model during 2026.
Related reading