A prospective client asked us last month why our proposal didn't show an AI-driven discount. Their other bids did. The honest answer is that the industry does not yet have the data to know whether AI tools make software projects cheaper, and anyone offering a firm discount based on AI productivity gains in 2026 is making a bet, not a measurement.
Not because we don't use AI tools. We do, across different phases of every project. But we don't yet have the data to say AI is a net positive on total project cost. Some things are faster: scaffolding, boilerplate, and first drafts of repetitive code. Some things are slower as a direct result of that speed: debugging AI-generated logic, reviewing hallucinated patterns, and managing tool costs that change quarter to quarter.
This post is the context behind that answer. Here is what AI actually costs the teams building your software, why the pricing is less stable than it looks, and what to ask before trusting a number that nobody can fully substantiate yet.
AI Coding Tool Pricing Is Shifting in 2026
Three things happened in April 2026 that anyone evaluating a software proposal should know about.
On April 4, Anthropic cut off third-party tool access from Claude subscriptions. Developers who had been running AI coding agents through external tools like OpenClaw on a flat $20 or $200/month subscription were told those tools would now bill separately at pay-as-you-go API rates. Industry analysts estimated the price gap at more than five times, with heavy agentic users facing increases up to 50x their previous monthly spend. Boris Cherny, Anthropic's Head of Claude Code, was direct about the reason: "Our subscriptions weren't built for the usage patterns of these third-party tools."
On April 27, GitHub announced that all Copilot plans would move to usage-based billing on June 1. The sticker price stays the same ($10/month for Pro, $19/user/month for Business), but usage is now metered by tokens. Code completions remain free on paid plans. Everything else (Chat, CLI, the cloud agent, third-party coding agents) consumes credits from a monthly allotment. In a separate post explaining why it was tightening individual plan limits, GitHub was unusually direct: "It's now common for a handful of requests to incur costs that exceed the plan price." The developer reaction was predictable and blunt: the low introductory pricing built dependency, and the repricing arrives after the workflows are too embedded to easily switch.
On April 28, the Wall Street Journal reported that the four largest cloud providers are projected to spend a collective $674 billion on AI infrastructure in 2026, more than double their 2024 levels. That spending is being funded in part by workforce reductions: 45,800 tech layoffs were announced in March alone, the worst month in at least two years. This matters for tool pricing because hyperscaler costs flow downstream. Model providers like Anthropic and OpenAI price their APIs against the underlying compute. Tool vendors like GitHub, Cursor, and Copilot price against those API rates. When the top of the stack gets more expensive, every layer below recalibrates eventually. The Copilot and Claude repricing did not happen in a vacuum.
The pattern across all three: flat-rate AI pricing was a customer acquisition strategy, not a sustainable business model. The correction is happening now, and the development teams using these tools are adjusting in real time.
What Engineering Teams Actually Pay for AI Coding Tools
The sticker price on AI coding tools is $10 to $20 per month. The actual per-developer spend at engineering teams right now is significantly higher.
The Pragmatic Engineer's 2026 survey of engineering teams found that companies commonly pay $100 to $200 per month per engineer for "max" tier plans across tools like Claude Code, Cursor, and Codex. One chief product and technology officer reported running up $600/month bills on Cursor across their team. A founder at a seed-stage company questioned whether the pricing could last: "If we assume that the third-party inference providers are doing so at a sustainable price, the much more expensive Opus model cannot be sustainable, never mind profitable at these plan costs." Most survey respondents expected AI tool prices to rise further.
The typical developer stack compounds quickly. A common configuration: Cursor Pro ($20) + Claude Pro ($20) + ChatGPT Plus ($20) + GitHub Copilot ($10) = $70/month as a floor. That is before any overages, API usage, or premium model access. Morph's analysis of real Claude Code usage found that one developer consumed 10 billion tokens over eight months. At API rates, that would cost over $15,000. The Max subscription costs $800. The subscription math is favorable today, but that is precisely the math that Anthropic and GitHub are now changing.
For teams using AI aggressively, the numbers climb further. SitePoint's cost analysis modeled a 30-developer team with 10 heavy agentic users and found monthly costs ranging from $5,000 in a quiet month to $14,000 during intensive refactoring sprints. Vantage's analysis of agentic coding costs makes an important counterpoint: their hypothetical $1,200/month developer also had the second-lowest cost per PR and nearly double the throughput of anyone else on the team. Per-developer spend without a denominator is meaningless. The question is not what AI costs but what it costs per unit of output, and almost nobody is tracking that yet.
Why Agentic AI Costs Compound Faster Than You Expect
The cost gap between using AI as an autocomplete tool and using it as a coding agent is not linear. It is multiplicative.
A coding agent resends the full conversation history with every API call. As the context window fills up, costs compound. A 200K-token conversation costs 10x what a 20K-token one does, and long agentic sessions routinely hit that range. For context on the magnitude: Anthropic's engineering team documented that a single-agent loop in their research system consumes roughly four times more tokens than a standard chat interaction, and multi-agent systems consume approximately 15 times more. Those figures are from research agents running web searches in parallel, not coding agents specifically, but the dynamics (context accumulation, retry loops, multi-step tool use) are the same.
When things go wrong, the costs multiply further. Morph documented a case where a $0.50 fix became a $30 bill through 47 agent iterations. Studies consistently find that 60 to 80% of tokens in agentic coding sessions go to navigation and search (reading files, finding code, building context), not to writing code. The agent is spending most of your money looking for things, not building things.
This matters for proposals because agentic workflows are where the productivity gains are largest, and also where the hidden costs live. A developer billing at $150/hour who saves 100 minutes on a task saves $250 in labor. If the agent burns $30 in tokens doing it, that is a clear win. But when the agent loops 47 times on a bug a developer could have fixed in 15 minutes, the $30 in tokens lands on top of the developer's time, not instead of it. The token cost is always additive to the hourly rate. It only pays for itself when the time saved exceeds the tokens spent, and nobody is tracking that ratio consistently yet.
The "AI Discount" Problem in Software Proposals
Some development shops are advertising AI-powered discounts on their proposals. The question buyers should ask is: discounted against what baseline?
If the shop raised their rates over the past year and then discounted them back, that is marketing. If they genuinely reduced their estimates based on AI-driven productivity gains, ask for the data. How many projects have they completed with AI tools? What was the actual delivery time compared to the estimate? Did the total cost come in lower than comparable pre-AI projects? If they cannot show that data, the discount is a bet on a thesis they have not validated.
The productivity data that does exist is mixed. The Pragmatic Engineer's survey found that developers report spending more time debugging and fixing AI-generated code than they expected. Some report a sense of "identity loss" from no longer doing hands-on work, which is not a cost that shows up on an invoice but affects retention and team health. A 2025 METR study cited by Codebridge found that experienced open-source developers were 19% slower with AI tools, despite believing they were 20% faster, a gap of nearly 40 percentage points between perceived and measured productivity. METR has since published follow-up data on late-2025 tools showing the effect may have narrowed, though they note significant measurement caveats, including selection bias among participants.
None of this means AI tools are not valuable. They are. But the value is uneven, it varies by task type and developer experience, and it is offset by costs that most teams are not yet tracking rigorously. A development partner who tells you "AI makes us faster, so your project costs less" without the project data to back it up is selling a feeling, not a finding.
How AI Costs Show Up in Software Development Pricing
AI tool costs are still small relative to the fully loaded cost of a developer. A software engineer in the U.S. costs $12,000 to $20,000 per month when you include salary, benefits, overhead, and management. A $200/month AI tool budget is 1 to 2% of that. At current prices, most development shops are absorbing AI tool costs into their existing margins without adjusting their rates.
The trajectory is what matters. If agentic workflows push per-developer AI costs to $500 to $1,000 per month (which the SitePoint modeling suggests is already happening for heavy users), that becomes 3 to 8% of developer cost, which starts affecting margins. At that point, development partners have three options.
- Absorb it. Margin compression. The shop eats the cost and hopes productivity gains offset it. This works until it does not, and it is not transparent to the buyer.
- Pass it through. AI tool costs appear as a line item on the invoice, similar to cloud hosting or third-party API fees. This is transparent but uncommon today.
- Offset it. AI genuinely makes the team faster, so the project costs less overall despite higher tool costs. This is the outcome everyone wants, but it requires measurement that most shops have not done.
The third option is the one worth paying for. But it is the hardest to verify, and the pricing instability described above makes it harder. A shop that built its productivity assumptions around $20/month, Claude Pro is now dealing with $100 to $200/month Max plans, or per-token API billing with no ceiling. A shop that estimated projects assuming GitHub Copilot's flat-rate model is recalculating for usage-based billing that starts June 1. The estimates they gave you three months ago may not reflect their costs today.
Questions to Ask Your Development Partner About AI Costs
If you are evaluating proposals for a custom software project in 2026, the AI cost question is worth asking directly. Not to catch anyone out, but to understand how your partner thinks about a cost structure that is changing in real time.
- If you are offering an AI-driven discount, can you show me project data? Completed projects, estimated versus actual timelines, and total costs compared to pre-AI baselines. If the answer is "we just know we are faster," that is a belief, not evidence.
- What happens to your pricing if your AI tool vendor changes billing models? This is not hypothetical. It happened twice in April. A partner who has thought about this will have an answer. A partner who has not is exposed to cost surprises that will eventually reach you.
- How are you using AI tools in development, and how does that affect your estimates? The answer should be specific. "We use AI" is not an answer. "We use Claude Code for scaffolding and code review, Copilot for inline completions, and we track token spend per project" is an answer.
- Are AI tool costs included in your rates or passed through separately? Either approach is fine. Not knowing which one they are doing is a red flag.
- How do you manage token costs on agentic workflows? Agent loops can burn through hundreds of dollars in tokens on a single task if they are not monitored. Ask whether the team has spending caps, usage dashboards, or model routing strategies. These are signs of a team that has done the math.
The Honest Answer
The cost of AI in software development is real, growing, and unstable. For buyers, the risk is not that your development partner uses AI tools. You want them to. The risk is that nobody has done the math on what those tools actually cost and whether the productivity gains justify the spend.
The teams that will win this transition are the ones tracking both sides: the time saved and the money spent. The ones who can show you, not tell you, that AI made your project better. We are tracking estimated versus actual hours, token spend per project, and cost per deliverable across our engagements now. We expect the data to be compelling when there is enough of it. But we are not going to put a number on a proposal that we cannot defend with evidence.
That is a harder sell than a 30% discount. But in a market where developers believed they were 20% faster with AI tools while controlled studies measured them at 19% slower, we would rather be the team that measures than the one that guesses.
If you are planning a software project and want to have this conversation before committing to a proposal, that is how we work. Get in touch.
