What are the main business models for AI code assistant companies?
In our AI code assistant market deck, you will find everything you need to understand the market
SUMMARY
The main business models for AI code assistant companies are now hybrid SaaS-and-usage models: customers pay for software access, while heavier AI work is metered, prepaid, pooled, or passed through separately. Pure unlimited subscriptions look much less durable once autonomous agents can consume large amounts of model inference and infrastructure on behalf of one developer.
Per-seat pricing is not disappearing, but the seat is changing meaning. It increasingly pays for access, administration, security and a bundled allowance of computation rather than genuinely unlimited AI.
Agentic coding is weakening the link between headcount and cost. One engineer can launch several background agents at once, so the expensive unit is gradually becoming machine workload rather than the number of humans with licenses.
Enterprise customers look like the strongest profit pool because they pay for far more than model access. SSO, auditability, privacy controls, procurement support, context, policy enforcement and spend management add revenue without increasing inference cost in the same proportion.
Low-priced individual subscriptions still matter a lot, but mostly as distribution. They create habits, seed bottom-up adoption and help tools enter companies, while heavy individual users can be among the least attractive customers economically.
Incumbents have a structural pricing advantage that independent startups do not. Microsoft, OpenAI, Amazon, Google and JetBrains can treat coding AI as one part of a wider cloud, IDE or AI relationship, while a standalone vendor usually needs the coding product itself to carry the economics.
There are two opposite but credible ways to protect gross margin. A vendor can own enough of the model stack to make common coding work cheaper, or it can let customers bring and pay for their own models while charging for the software layer around them.
Codebase context, code review, orchestration and engineering governance are becoming separate revenue layers rather than minor features inside an assistant. As agents produce more code, companies need more automated systems to understand, review and control that output.
Credits are becoming common, but the important point is not the word “credit.” What matters is the structure: a recurring fee, an included allowance, clear overage rules, pooled consumption and often prepaid commitments that make agent spending easier to budget.
The strongest standalone businesses will probably combine recurring enterprise software revenue with one or more of three things: paid agent consumption, cheaper proprietary inference, or customer-funded model usage. Freemium and low-priced subscriptions can still be excellent growth engines, but they are increasingly the front door rather than the whole business.
Why are AI code assistant prices changing so fast right now?
AI code assistant pricing is changing so fast because long-running coding agents have made the old unlimited $10 or $20 subscription increasingly hard to sustain.
Early AI coding tools mostly completed a line, generated a function or answered a short question. The cost of serving one developer was reasonably predictable. Today's agents can search an entire repository, load large amounts of context, call tools, run commands, test their own work and keep going for minutes without the developer doing anything.
Replit offers one of the clearest examples of how quickly the economics changed. Its first Agent could generally work for only a couple of minutes, so Replit charged $0.25 per checkpoint. When Agent v2 became capable of working autonomously for up to 20 minutes, Replit said some runs could cost the company as much as $10. The fixed $0.25 unit stopped making sense, and Replit moved to effort-based pricing tied more closely to the computing resources consumed.
Augment Code ran into the same problem from another direction. It originally priced usage around messages, then explained that one message might trigger a tiny edit while another could launch dozens of tool calls. It moved customers to credits instead.
GitHub has now gone further. Copilot Business still costs $19 per user per month and Enterprise costs $39, but most AI activity is measured in AI credits based on the underlying model and token consumption. Ordinary completions remain unlimited on paid plans, while chat, CLI use, cloud agents and third-party coding agents consume credits.
This is happening across enough unrelated vendors that it is no longer just pricing experimentation around the edges. AI coding is being repriced around the amount of work the software actually performs.
What does an AI code assistant company actually sell?
AI code assistant companies now make money from a mix of software access, AI computation and developer infrastructure rather than from one simple SaaS subscription.
The first part still looks familiar. A developer or company pays to access Cursor, GitHub Copilot, JetBrains AI, Tabnine or another coding product. The subscription includes the interface, integrations, administration and some level of AI usage.
The second part behaves much more like cloud computing. Every serious agent task consumes tokens and often additional infrastructure. Longer context, more expensive models, repeated testing and parallel agents all raise the vendor's cost.
Then there is a third business that becomes easy to miss: the infrastructure around the model. Companies can charge for repository indexing, organizational context, code search, security controls, agent orchestration, virtual environments, code review, CI/CD execution and governance.
We can see all three layers in Cursor's current Teams product. A Standard seat costs $40 per user per month. That seat includes the editor, team administration, privacy controls, agents, Bugbot and cloud workflows. It also includes model usage, while additional third-party model consumption is billed separately. Cursor even adds a $0.25-per-million-token platform charge when Teams customers use eligible third-party models.
Customers may think they are buying a “coding assistant,” but the vendor can earn revenue from several different layers underneath it.
This market map, featured in our AI code assistant market deck, highlights top companies and startups in the AI code assistant market
Is per-seat pricing still the main AI code assistant business model?
Per-seat pricing is still the main way AI code assistant companies sell access to engineering teams, but the seat increasingly comes with a usage allowance rather than unlimited AI.
GitHub charges $19 per user per month for Copilot Business and $39 for Copilot Enterprise. Cursor charges $40 per month for a Standard Teams seat and $120 for a Premium seat with five times more included usage. Tabnine charges $39 per developer per month for Code Assistant and $59 for its broader Agentic Platform. Devin currently charges $40 per month for a full Teams seat, although occasional users can join through free flex seats and consume shared credits instead.
The seat survives because companies already know how to buy software this way. If a bank has 5,000 developers, procurement can understand 5,000 licenses. SSO, user provisioning, audit logs, security controls and departmental budgeting also map naturally onto people.
What has changed is what the seat represents. With conventional SaaS, paying for a seat usually gives the user nearly unlimited use because the incremental cost of another click is tiny. An AI coding seat increasingly buys access, governance and a prepaid amount of computation.
That is why seat pricing survives even while the underlying economics become more usage-based.
| AI coding product | Current team pricing | How usage is handled |
|---|---|---|
| GitHub Copilot Business | $19/user/month | 1,900 pooled AI credits per user, then paid overage |
| GitHub Copilot Enterprise | $39/user/month | 3,900 pooled AI credits per user, then paid overage |
| Cursor Teams Standard | $40/user/month | Included model pools, then on-demand usage |
| Cursor Teams Premium | $120/user/month | 5× Standard included usage |
| Tabnine Agentic Platform | $59/user/month | Software fee plus separate model economics when Tabnine supplies the LLM |
Why are AI code assistants moving to subscription plus usage?
Subscription plus usage is currently becoming the default AI code assistant model because it gives customers a predictable base price without forcing vendors to subsidize unlimited agent work.
The pattern is striking. GitHub now converts model and token consumption into AI credits. Cursor gives each plan an included usage pool and charges on demand after that. JetBrains AI Pro and AI Ultimate include monthly AI credits and let customers buy additional credits. Replit Core includes $25 of monthly credits, while Replit Pro includes $100. OpenAI has also moved Codex toward flexible usage: ChatGPT plans still include access, but customers can buy credits after they exhaust their included limits.
These systems are designed differently, so comparing one “credit” with another would be meaningless. What we can compare is the structure. In several cases, the subscription price and included compute allowance are now deliberately linked.
GitHub makes that relationship unusually explicit. Copilot Business costs $19 per month and contributes 1,900 AI credits to the organization's shared pool. Each AI credit is worth one cent. Enterprise costs $39 and contributes 3,900 credits. A customer can then set budgets at the user, cost-center, organization or enterprise level.
Cursor has recently pushed the model even further. The company introduced separate pools for its cheaper first-party models and external frontier models, while adding a $120 Premium team seat for the relatively small group of developers who consume much more AI than everyone else. Cursor said its own usage data showed that a minority of power users drives most of the unpredictable spend.
A flat subscription works well until usage becomes wildly unequal. Subscription plus metering keeps the simple entry price while charging the expensive tail of the distribution more accurately.
If you want more recent data on this point, please see our latest AI code assistant market report.
As this chart shows, and as featured in our AI code assistant market deck, search interest in AI code assistants has increased significantly
Are $20 individual AI coding subscriptions actually good businesses?
Individual AI coding subscriptions are great distribution products, but current evidence suggests that heavy consumer-style usage can be a poor business on its own.
Cursor gives us an unusually revealing case. The company reached roughly $2 billion in annualized revenue in February 2026, about $3 billion by late April and approximately $4 billion by June, according to reporting by Bloomberg and Forbes. Very few developer products have ever monetized this quickly.
Yet the economics underneath that growth were much less impressive. TechCrunch reported that Cursor had operated at negative gross margins until recently. After introducing proprietary models and routing more traffic toward cheaper inference, the company reached only slight overall gross-margin profitability. Large-enterprise accounts were already gross-margin positive, while individual developer accounts were still losing money.
That gap tells us more about the market than Cursor's revenue alone. Individual developers include exactly the kind of power users who can run agents continuously, load huge repositories and deliberately choose expensive frontier models. They also compare competing products constantly and can cancel with little organizational friction.
Replit offers a useful contrast. CEO Amjad Masad said in 2026 that Replit had been gross-margin positive for more than a year. Replit prices complex agent work according to effort and also sells hosting, databases, deployments and other infrastructure around the generated application. The company therefore has more ways to monetize a project than the AI interaction itself.
Low-priced individual plans are best understood mainly as acquisition and habit-building products. Some vendors can absolutely make money on them, but they are not the safest foundation for an agent-heavy business.
Is enterprise AI coding becoming the real profit pool?
Enterprise AI coding currently looks like the strongest profit pool because companies pay for security, administration, reliability and organizational deployment alongside the AI itself.
Claude Code provides the strongest demand evidence. Anthropic disclosed that Claude Code had passed $2.5 billion in annualized run-rate revenue after reaching $1 billion only a few months earlier. More than half of Claude Code revenue was already coming from enterprise use, while business subscriptions had quadrupled since the beginning of 2026.
GitHub shows the same shift at a much larger installed base. Microsoft reported 4.7 million paid GitHub Copilot subscribers, up 75% year over year, and later said nearly 140,000 organizations were using Copilot. Enterprise subscribers had nearly tripled year over year.
Cursor has moved in the same direction. By the time its annualized revenue reached roughly $2 billion, corporate customers were already generating around 60% of revenue according to Bloomberg reporting. A few months later, Cursor reportedly had more than 3,000 customers each spending at least $100,000 annually.
The economics are straightforward. A large company may care about model quality, but it also needs SSO, SCIM, privacy guarantees, auditability, procurement support, spending controls, deployment choices and policy enforcement. Vendors can charge for those features without increasing inference costs proportionally.
Enterprise contracts also spread usage more efficiently. GitHub pools included AI credits across the entire billing entity. Cursor reserves pooled usage for Enterprise customers. Devin Teams combines fixed full seats with shared on-demand credits. A company with thousands of developers can therefore offset light and heavy users instead of pricing every person around the worst-case user.
Today, serious AI coding businesses are increasingly following enterprise software economics even when their products originally spread through individual developers.
If you want more recent data on this point, please see our latest AI code assistant market report.
This chart, featured in our AI code assistant market deck, illustrates yearly VC funding for AI code assistant startups
Are coding agents changing what AI coding companies charge for?
Coding agents are pushing AI coding companies away from charging only for people and toward charging for actual software work.
Autocomplete naturally maps to a developer seat because one developer sits at one keyboard. Agentic coding breaks that relationship. One engineer can now launch several tasks in parallel, leave agents running and return later to completed code, tests or pull requests.
Replit already lets Pro users run as many as ten agents in parallel. Cursor's current Teams product includes cloud agents and automations. GitHub Copilot can run cloud agents against repositories. Devin was designed from the beginning around autonomous sessions rather than constant human interaction.
Once one person can generate five or ten simultaneous workloads, headcount becomes a weak proxy for cost.
Devin's current pricing is particularly interesting because Cognition mixes both ideas in one product. A regular Teams user can have a $40 full seat with an included quota. Occasional users can receive free flex seats and simply draw from a shared on-demand credit pool. Devin Review consumes those credits separately from the normal seat quota.
Replit reached the same conclusion when long-running Agent sessions made its old $0.25 checkpoint price economically useless.
This shift should get stronger as agents become more autonomous. The industry may still call these products “developer tools,” but a growing share of revenue will follow machine workload rather than human headcount.
Can pure usage-based pricing work for AI coding agents?
Usage-based pricing works well for expensive AI coding workloads, but pure pay-as-you-go is unlikely to replace subscriptions completely because engineering teams still want predictable bills.
The attraction is obvious. If a coding agent consumes ten times more model inference and infrastructure, the vendor earns more revenue instead of absorbing the extra cost.
OpenAI tested a particularly direct version of this with Codex-only seats for teams. Those seats had no fixed seat fee and billed usage according to token consumption. OpenAI later stopped offering new Codex-only seats to ChatGPT Business customers, although flexible usage remains part of the wider Codex and Enterprise pricing structure.
Devin now combines an $80 monthly Teams minimum with $40 full seats and shared on-demand credits. A team with no full seats effectively puts the entire minimum into prepaid usage. A team with occasional users can give those people free flex seats and pay only when they actually use Devin.
Replit Pro goes in another direction. Instead of charging every collaborator separately, the $100 monthly plan includes up to 15 builders and $100 in pooled credits. Customers can pre-purchase larger credit packages, and the company offers discounts as those commitments increase.
Those designs solve the biggest weakness of raw metering: people hate opening an agent without knowing whether the task will cost $0.50 or $50. Replit learned that painfully when a pricing-calculation error affected roughly 6% of paying users during its first effort-based rollout and forced the company to issue refunds and extra credits.
The strongest version of usage pricing today is prepaid or committed consumption with clear limits. Customers get budget control, while vendors stop pretending that every agent run costs the same.
If you want more recent data on this point, please see our latest AI code assistant market report.
This chart, featured in our AI code assistant market deck, breaks down Anyshpere’s playbook in AI code assistants
Is freemium a real business model for AI code assistant companies?
Freemium is currently a powerful way to distribute AI code assistants, but free users only make economic sense when enough of them later convert into paid developers or enterprise accounts.
GitHub Copilot has a free tier. Cursor offers Hobby for free. Devin has a free individual plan. Replit gives Starter users daily Agent credits. JetBrains offers a small free AI allowance.
Developer software is unusually well suited to this approach because adoption often begins before procurement. One engineer can try a tool, use it on real work and then ask the company to pay once it becomes useful enough. GitHub's own numbers show how large that funnel can become: Microsoft said GitHub Copilot had 26 million total users in late 2025, compared with 4.7 million paid subscribers reported a few months later.
Free AI does have a cost that free conventional SaaS usually does not. Every generated answer can consume model inference. Vendors therefore limit free agent requests, restrict model choice, use cheaper models or provide small daily credit allocations.
GitHub's current pricing makes this distinction very visible. Paid plans can still offer unlimited code completions because short completions are relatively cheap and predictable, while heavier AI interactions are tracked through credits.
So freemium remains important, especially for bottom-up enterprise sales. It is better understood as a distribution model than as a standalone economic model.
Why can Microsoft, OpenAI and JetBrains price AI coding differently from startups?
Microsoft, OpenAI, Amazon, Google and JetBrains can afford to make AI coding part of a larger product relationship, which gives those companies much more freedom on price than a standalone startup has.
GitHub Copilot can help Microsoft make GitHub more valuable, strengthen Visual Studio and pull more development activity toward Azure. Amazon Q Developer can make AWS easier to build on. Gemini Code Assist supports Google Cloud. JetBrains AI strengthens paid IDE subscriptions that developers already use.
OpenAI has taken the bundling idea furthest. Codex is included inside several ChatGPT plans, so a user buying ChatGPT for research, writing, data analysis and other work can also receive coding capability. Additional Codex consumption can then be monetized through credits.
This creates a difficult competitive benchmark for independent companies. A startup may need $40 of revenue from its coding product because coding is the entire business. Microsoft can accept weaker standalone Copilot economics if the product improves GitHub retention, Azure usage or the value of its wider developer ecosystem.
Microsoft's own financial comments show that this cross-subsidy is not theoretical. In its fiscal third-quarter call, management said Microsoft Cloud gross margin was being pressured partly by increased GitHub Copilot usage. In the same discussion, Microsoft described Copilot's move toward pricing tied more closely to usage and value.
Independent companies can still win. Cursor's rise from a small developer tool to roughly $4 billion in annualized revenue proves that developers will pay separately when the experience is better enough. The hurdle is simply higher: an independent AI coding company needs a product developers actively choose over something their employer may already be paying for.
| Type of vendor | Typical AI coding model | Economic advantage |
|---|---|---|
| Microsoft / GitHub | Paid seats + AI credits | Can strengthen GitHub, Visual Studio and Azure |
| OpenAI | Codex bundled into ChatGPT + flexible usage | Monetizes the same customer across many AI workflows |
| JetBrains | IDE relationship + AI subscription/credits | AI increases the value of existing developer software |
| Independent vendor such as Cursor | Standalone subscription + usage | Must win directly on coding workflow and product quality |
This chart, featured in our AI code assistant market deck, illustrates yearly funding for AI code assistant startups
Does an AI code assistant company need its own model to make good margins?
An AI code assistant company does not need to own every model, but having a cheap proprietary model is becoming a major economic advantage for high-volume coding workloads.
Cursor again gives us the best evidence because we can compare the company's economics before and after deeper vertical integration. Cursor initially built its product largely on models supplied by companies such as Anthropic. TechCrunch reported that this dependence helped push Cursor into negative gross margins.
Cursor then introduced its own Composer models and increased routing toward cheaper alternatives. According to the same reporting, those changes helped bring the company into slightly positive overall gross margins.
The current pricing gap shows why. Cursor lists Composer 2.5 at $0.50 per million input tokens and $2.50 per million output tokens. Third-party frontier models offered through the same product can cost several times more. Cursor has also created separate usage pools so teams receive generous access to first-party models while expensive outside models remain more tightly metered.
A proprietary model gives an AI code assistant another strategic benefit: suppliers can become competitors overnight. Anthropic now sells Claude Code directly. OpenAI sells Codex. A company whose entire product depends on paying those suppliers for every useful interaction risks having both its margin and product roadmap controlled by rivals.
Still, training the biggest frontier model is not required. A more sensible architecture can use cheap proprietary models for common coding work and route the hardest tasks to outside frontier models.
That hybrid approach is currently more convincing than either extreme.
If you want more recent data on this point, please see our latest AI code assistant market report.
Can AI coding companies make money without paying for the underlying model?
AI coding companies can increasingly sell the software layer while letting customers supply or pay for the underlying models themselves, and this is one of the cleanest ways to protect margins.
Tabnine shows the model explicitly. Its Code Assistant costs $39 per user per month and its Agentic Platform costs $59. Customers using their own model endpoint can get unlimited usage from Tabnine's software layer. If they instead want Tabnine to provide model access, Tabnine charges for reserved token consumption based on the underlying model provider's price plus a 5% handling fee.
JetBrains also lets customers connect third-party model providers while continuing to use AI features inside its IDEs. Enterprise organizations can therefore buy the workflow and integration layer without necessarily asking JetBrains to finance every token.
Augment is moving in a similar direction through “bring your own agent” support. Teams that already pay for Claude Code, Codex or other agents can use those providers while consuming Augment's orchestration and context rather than drawing from the same Augment credit pool.
This separation changes the risk profile of the business. The software vendor can make money from code context, interfaces, orchestration, enterprise controls and workflow integration while the customer absorbs more of the variable model cost.
It also exposes the weakness of simple model resale. If an AI coding company is only buying tokens from Anthropic for $1 and selling them to customers for $1.10, there is very little protection when Anthropic cuts prices or sells directly.
The stronger version is to charge for something the model provider does not automatically provide: the development workflow around the model.
This chart, featured in our AI code assistant market deck, compares the main business model options for AI developer tools platforms
Can codebase context and code review become separate AI coding businesses?
Codebase context, code review and other engineering workflows are already becoming separate products because companies will pay for agents to understand and police their software even if they use another tool to write the code.
Sourcegraph is a good example of how valuable the context layer can become. Its current Enterprise product starts at $16,000 and combines organization-wide code search, navigation, monitoring, batch changes and AI capabilities. More importantly, Sourcegraph can work alongside Claude Code, Cursor, Codex, Amp and other agents rather than requiring the customer to standardize on Sourcegraph's own assistant.
Tabnine is taking the same idea even further with its Enterprise Context Engine. The product builds a structured view of a company's architecture, dependencies and internal standards and can provide that context to Cursor, GitHub Copilot, Claude Code, Tabnine agents or internal tools. Tabnine currently lists the Enterprise Context product separately at $5,800.
Code review is also separating from ordinary coding assistance. Cursor sells Bugbot on usage-based billing. Devin Review consumes on-demand credits independently from a user's normal Devin session quota. Tabnine sells headless agents as an add-on for automated CI/CD workflows.
The economics get stronger as more code is written by agents. If one developer can create much more code, companies need more automated review, repository understanding, policy checks and system-level context. Those workloads also run in the background, where charging purely by human seat makes less sense.
Some successful “AI code assistant companies” may end up making a growing share of their money from products that never help a developer type a line of code directly.
So what are the main business models for AI code assistant companies today?
The strongest AI code assistant business model today is a paid software relationship with included AI usage and metered charges for heavier agent work, especially when the company can add enterprise software or infrastructure around the models.
Per-seat subscriptions remain important because organizations still buy access by employee. But unlimited flat-rate AI is fading for serious agent workloads. GitHub's switch to token-linked AI credits, Cursor's split usage pools, JetBrains' credit system, Replit's effort pricing and Devin's mix of seats and shared credits all point in the same direction.
Enterprise licensing looks particularly attractive. Claude Code already gets more than half of its revenue from enterprise use. Cursor's large-enterprise accounts reached positive gross margins before its individual accounts did. GitHub is seeing rapidly rising enterprise adoption. Companies will pay more for security, administration, context and predictable deployment than an individual developer will pay for model access alone.
Platform bundling is equally powerful for incumbents. Microsoft, OpenAI, Amazon, Google and JetBrains can make coding AI improve a much larger commercial relationship. Independent companies need a stronger standalone product to compensate.
For startups, there are two especially interesting alternatives. One is to own enough of the model stack to bring inference costs down, as Cursor has been doing. The other is to stop carrying much of the inference cost at all and charge customers for context, orchestration, governance or workflow software while the customer supplies the models.
Autonomous agents are also weakening the relationship between developer headcount and AI consumption. That makes workload-based revenue, pooled credits and prepaid usage commitments much more important than they were during the autocomplete era.
We therefore see six real business models in the market now, but only a few look strong enough to build large standalone companies around.
| Business model | How companies make money | Our view today |
|---|---|---|
| Enterprise seats + metered AI usage | Fixed software fee plus paid consumption above an allowance | Strongest general model |
| Platform bundling | Coding assistant supports a larger cloud, IDE or AI subscription | Extremely strong for incumbents |
| Agent usage / effort pricing | Customer pays according to compute or autonomous work consumed | Likely to grow quickly |
| Software + customer-supplied models | Vendor charges for workflow while customer carries inference cost | Very attractive enterprise model |
| Context, review and engineering infrastructure | Separate fees for code intelligence and automated workflows | Emerging as a major second revenue layer |
| Low-priced individual subscriptions | Monthly access aimed at individual developers | Excellent distribution, weaker economics for heavy users |
The answer to the title is fairly clear now. AI code assistant companies started by selling something that looked like ordinary SaaS, but the durable business is becoming a hybrid of SaaS and metered computing. The companies with the best economics will collect recurring software revenue while making sure that expensive autonomous work is either paid for separately, run on cheaper proprietary models or passed through to the customer.
If you want more recent data on this point, please see our latest AI code assistant market report.
This chart, featured in our AI code assistant market deck, illustrates how market revenue is distributed across customer segments in the AI code assistant market
OUR METHODOLOGY
The economics of AI code assistants are moving quickly, so this analysis breaks the market into the dimensions that actually determine how these companies make money: pricing structure, AI consumption, customer type, cost structure, distribution, model economics, bundling, and additional software or infrastructure revenue.
For each dimension, we prioritized recent market evidence showing what companies are actually doing today. We used current pricing and billing structures, product changes, usage policies, company disclosures, earnings commentary, adoption data, and reported unit economics, with more weight on fresh evidence because agentic coding is changing much faster than the earlier autocomplete market did.
We then compared those patterns across companies rather than treating one pricing change as proof of a market-wide shift. Similar responses appearing independently across GitHub, Cursor, Replit, Devin, JetBrains, Tabnine, OpenAI and Augment were treated as stronger evidence that the economics themselves are moving.
Different source types were used for different questions. Pricing pages and billing documentation were the clearest evidence of how vendors capture revenue and pass through AI costs; product announcements showed how those models are changing; company and earnings disclosures were used for adoption and customer mix; and established financial or technology reporting was used where private companies do not publish details such as gross margins or revenue composition.
We did not turn the evidence into an artificial numerical score. Terms such as “strongest,” “attractive,” or “weaker economics” reflect the weight and consistency of the combined evidence across pricing, cost, customer mix and distribution rather than any single metric. Pricing and product structures were checked against sources available through August 25, 2026.
Key sources used for this analysis include GitHub Copilot plans and GitHub's usage-based billing documentation; Microsoft FY2026 Q2 and FY2026 Q3 earnings materials; Cursor Teams pricing, Cursor's Teams pricing update, Composer 2.5 pricing, and TechCrunch's reporting on Cursor's margins and enterprise growth; Replit's effort-based pricing announcement, its pricing recap, current Replit pricing, and TechCrunch's interview with Amjad Masad; Tabnine pricing, Tabnine Enterprise Context pricing, and Tabnine headless-agent pricing; JetBrains AI licensing and BYOK documentation; Devin billing documentation; OpenAI's Codex flexible-pricing announcement and flexible pricing documentation; and Augment Code's credit-based plans, Context Engine MCP, and BYOA pricing guide.
This chart, featured in our AI code assistant market deck, shows how AI coding assistant technology has evolved over time
Related blog posts
