%PDF-1.4 %âãÏÓ 1 0 obj << /Type /Catalog /Pages 2 0 R >> endobj 2 0 obj << /Type /Pages /Count 5 /Kids [5 0 R 7 0 R 9 0 R 11 0 R 13 0 R] >> endobj 3 0 obj << /Type /Font /Subtype /Type1 /BaseFont /Helvetica >> endobj 4 0 obj << /Type /Font /Subtype /Type1 /BaseFont /Helvetica-Bold >> endobj 5 0 obj << /Type /Page /Parent 2 0 R /MediaBox [0 0 595.28 841.89] /Resources << /Font << /F1 3 0 R /F2 4 0 R >> >> /Contents 6 0 R >> endobj 6 0 obj << /Length 5745 >> stream BT /F2 22 Tf 0.06 0.08 0.12 rg 1 0 0 1 46 789.89 Tm (How Are Software Companies Cutting AI Costs) Tj ET BT /F2 22 Tf 0.06 0.08 0.12 rg 1 0 0 1 46 762.89 Tm (Without Slowing Down Development?) Tj ET BT /F2 11 Tf 0.72 0.14 0.18 rg 1 0 0 1 46 725.89 Tm (TechRounder PDF Edition) Tj ET BT /F1 9.5 Tf 0.36 0.39 0.46 rg 1 0 0 1 46 709.89 Tm (Live article:) Tj ET BT /F1 9.5 Tf 0.36 0.39 0.46 rg 1 0 0 1 46 697.39 Tm (https://www.techrounder.com/ai/how-are-software-companies-cutting-ai-costs-without-slowing-down-development/) Tj ET q 0.82 0.85 0.9 RG 1 w 46 678.89 m 549.28 678.89 l S Q BT /F1 10 Tf 0.24 0.27 0.32 rg 1 0 0 1 46 666.89 Tm (By Vipin PG | Published August 8, 2026 | Updated August 8, 2026 | Format: Deep Dive | 9 min read) Tj ET BT /F2 13 Tf 0.72 0.14 0.18 rg 1 0 0 1 46 643.89 Tm (In brief) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 623.89 Tm (Software companies are balancing AI cost and development speed by routing tasks to different AI) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 608.89 Tm (models based on difficulty \(using cheap models for simple work and expensive ones only when) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 593.89 Tm (needed\), turning on prompt caching to avoid paying for the same context twice, setting hard spending) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 578.89 Tm (limits before a project starts, and measuring actual output like merged code instead of just how much) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 563.89 Tm (the AI was used. The companies getting this right treat AI spend the same way they treat cloud spend:) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 548.89 Tm (something you track, cap, and review every week, not something you find out about when the bill) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 533.89 Tm (arrives.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 508.89 Tm (In April 2026, Uber's CTO told reporters that the company had burned through its entire annual AI) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 493.89 Tm (coding budget in four months. Not because the tools didn't work - engineers liked them. The problem) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 478.89 Tm (was that nobody had modeled what would happen when 5,000 developers started running AI agents that) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 463.89 Tm (read, re-read, and reasoned over large codebases dozens of times a day. Microsoft ran into a similar) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 448.89 Tm (wall around the same time. These aren't edge cases. A 2025 survey of 372 enterprises found that) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 433.89 Tm (only 15% of companies could forecast their AI costs within 10% of what they actually spent. Most) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 418.89 Tm (missed by 11% to 25%, and nearly a quarter missed by more than half.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 396.89 Tm (At the same time, the pressure to move faster with AI hasn't gone anywhere. Boards want to see the) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 381.89 Tm (productivity gains they've been promised. So software companies are stuck between two forces) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 366.89 Tm (pulling in opposite directions - ship faster, spend less - and the ones handling it well aren't doing) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 351.89 Tm (anything mysterious. They're just being deliberate about a few specific decisions. Here's what that) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 336.89 Tm (actually looks like in practice.) Tj ET BT /F2 15 Tf 0.08 0.1 0.14 rg 1 0 0 1 46 308.89 Tm (Why AI Development Costs Get Out of Hand So Quickly) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 284.89 Tm (To understand the fix, it helps to understand why the bill grows the way it does. When a developer uses) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 269.89 Tm (a simple chatbot, they ask one question and get one answer. AI coding agents don't work that way. An) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 254.89 Tm (agent reads relevant files, forms a plan, writes some code, checks whether it worked, and loops back) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 239.89 Tm (if it didn't - and at every single step, it resends the entire accumulated context as part of the request.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 224.89 Tm (The model doesn't remember the previous step. It has to be told everything again, every time. That's) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 209.89 Tm (why a single coding task can push anywhere from 400,000 to 2 million cumulative tokens through the) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 194.89 Tm (API, even though the actual output might be a few hundred lines of code.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 172.89 Tm (Code itself is also more expensive to process than plain English - roughly 1.5 to 2.5 times more tokens) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 157.89 Tm (for the same amount of content, because of how code gets broken into tokens. And output tokens \(what) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 142.89 Tm (the model generates\) typically cost three to five times more than input tokens \(what you send it\),) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 127.89 Tm (because generating text requires far more computation than reading it.) Tj ET q 0.86 0.88 0.92 RG 1 w 46 42 m 549.28 42 l S Q BT /F1 8.4 Tf 0.42 0.45 0.5 rg 1 0 0 1 46 30 Tm (TechRounder | Page 1 of 5) Tj ET BT /F1 7.2 Tf 0.42 0.45 0.5 rg 1 0 0 1 46 19 Tm (https://www.techrounder.com/pdf/blog/how-are-software-companies-cutting-ai-costs-without-slowing-down-development.pdf) Tj ET endstream endobj 7 0 obj << /Type /Page /Parent 2 0 R /MediaBox [0 0 595.28 841.89] /Resources << /Font << /F1 3 0 R /F2 4 0 R >> >> /Contents 8 0 R >> endobj 8 0 obj << /Length 6472 >> stream BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 789.89 Tm (Put those two things together and you get numbers like this: Anthropic's own enterprise data shows) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 774.89 Tm (Claude Code costs an average of $13 per developer per active day, or roughly $150 to $250 per) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 759.89 Tm (developer per month. That sounds manageable until you look at the spread - heavy users of agentic) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 744.89 Tm (coding tools regularly hit $500 to $2,000 per developer per month. Multiply that across a few thousand) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 729.89 Tm (engineers, and you understand exactly how Uber's budget disappeared in sixteen weeks.) Tj ET BT /F2 15 Tf 0.08 0.1 0.14 rg 1 0 0 1 46 701.89 Tm (The Speed Gains Are Real, But Smaller Than the Hype Suggests) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 677.89 Tm (Before diving into cost strategies, it's worth pausing on something that changes how companies should) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 662.89 Tm (think about this whole balancing act: AI's actual effect on development speed is messier than most) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 647.89 Tm (marketing suggests, and that matters for cost decisions.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 625.89 Tm (METR, a research group that runs controlled studies on AI's real-world impact, found in mid-2025) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 610.89 Tm (that experienced developers were actually 19% slower when using AI tools on their own established) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 595.89 Tm (codebases - even though those same developers believed, after the fact, that AI had made them 20%) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 580.89 Tm (faster. By early 2026, METR's follow-up work suggested the picture had improved to something) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 565.89 Tm (closer to an 18% speedup, though the researchers themselves noted the data was noisy and the true) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 550.89 Tm (number could range widely.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 528.89 Tm (A separate 2026 study from developer-intelligence platform DX looked at 121,000 developers and) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 513.89 Tm (found that while 93% of them were using AI regularly, actual pull-request throughput across their) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 498.89 Tm (organizations rose only about 10% - nowhere near the 2x or 3x figures often quoted by tool vendors.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 483.89 Tm (McKinsey's research tells a more nuanced version of the same story: AI saves close to 46% of the) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 468.89 Tm (time on routine, repetitive tasks, but under 10% on genuinely complex work. And a large-scale analysis) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 453.89 Tm (by Opsera found that while AI cut time-to-pull-request by up to 58%, those same AI-written pull) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 438.89 Tm (requests then sat in code review 4.6 times longer, and carried security vulnerabilities at nearly triple) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 423.89 Tm (the rate of human-written code.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 401.89 Tm (None of this means AI isn't useful - it clearly speeds up the right kind of work. But it explains why) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 386.89 Tm (companies that just throw AI at everything and hope for the best often end up with a bigger bill and a) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 371.89 Tm (similar delivery timeline. The gains are concentrated in specific, well-scoped tasks, and the cost) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 356.89 Tm (strategies that work best are the ones built around that reality rather than around the assumption that AI) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 341.89 Tm (makes everything faster by default.) Tj ET BT /F2 15 Tf 0.08 0.1 0.14 rg 1 0 0 1 46 313.89 Tm (Strategy 1: Route Tasks to the Right Model, Not the Biggest One) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 289.89 Tm (This is the single most effective lever companies are pulling in 2026, and the logic is simple: not every) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 274.89 Tm (coding task needs your most powerful, most expensive AI model. Writing a unit test, formatting a) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 259.89 Tm (function, classifying a bug report, or generating boilerplate code doesn't require the same reasoning) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 244.89 Tm (depth as architecting a new system or debugging a subtle race condition.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 222.89 Tm (A "model router" checks how complex a task is and sends it to the cheapest model that can handle it) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 207.89 Tm (well, saving the expensive frontier models for the work that genuinely needs them. The price) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 192.89 Tm (difference between tiers is large enough that this alone can reshape a company's entire AI budget:) Tj ET BT /F1 10 Tf 0.18 0.2 0.24 rg 1 0 0 1 46 170.89 Tm (Model Tier: Budget / small models | Best For: Formatting, classification, simple extraction, boilerplate code |) Tj ET BT /F1 10 Tf 0.18 0.2 0.24 rg 1 0 0 1 46 157.89 Tm (Relative Cost: Often 10-15x cheaper per task) Tj ET BT /F1 10 Tf 0.18 0.2 0.24 rg 1 0 0 1 46 140.89 Tm (Model Tier: Mid-tier models | Best For: Everyday feature work, standard bug fixes, routine reviews | Relative) Tj ET BT /F1 10 Tf 0.18 0.2 0.24 rg 1 0 0 1 46 127.89 Tm (Cost: Balanced workhorse tier) Tj ET BT /F1 10 Tf 0.18 0.2 0.24 rg 1 0 0 1 46 110.89 Tm (Model Tier: Frontier models | Best For: System design, complex debugging, security-sensitive code, multi-step) Tj ET BT /F1 10 Tf 0.18 0.2 0.24 rg 1 0 0 1 46 97.89 Tm (planning | Relative Cost: Highest cost, reserved for hard problems) Tj ET q 0.86 0.88 0.92 RG 1 w 46 42 m 549.28 42 l S Q BT /F1 8.4 Tf 0.42 0.45 0.5 rg 1 0 0 1 46 30 Tm (TechRounder | Page 2 of 5) Tj ET BT /F1 7.2 Tf 0.42 0.45 0.5 rg 1 0 0 1 46 19 Tm (https://www.techrounder.com/pdf/blog/how-are-software-companies-cutting-ai-costs-without-slowing-down-development.pdf) Tj ET endstream endobj 9 0 obj << /Type /Page /Parent 2 0 R /MediaBox [0 0 595.28 841.89] /Resources << /Font << /F1 3 0 R /F2 4 0 R >> >> /Contents 10 0 R >> endobj 10 0 obj << /Length 5584 >> stream BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 789.89 Tm (Companies that have implemented this kind of tiered routing report cost reductions between 40% and) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 774.89 Tm (85%, depending on how much of their workload is routine versus complex. Even a moderate shift -) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 759.89 Tm (routing 60% to 70% of tasks to a cheaper model instead of a frontier one - tends to cut the input-token) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 744.89 Tm (bill by roughly two-thirds, without a noticeable drop in output quality, as long as the routing logic is) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 729.89 Tm (tuned properly and checked regularly against real results.) Tj ET BT /F2 15 Tf 0.08 0.1 0.14 rg 1 0 0 1 46 701.89 Tm (Strategy 2: Stop Paying for the Same Context Twice) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 677.89 Tm (Here's a detail most engineering teams don't realize until someone points it out: if your AI coding tool) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 662.89 Tm (sends the same system instructions, the same file context, or the same project background with every) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 647.89 Tm (single request, you're paying full price to process that information again and again, even when nothing) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 632.89 Tm (in it has changed.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 610.89 Tm (Prompt caching solves this. Major providers now let you cache a chunk of context and reuse it across) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 595.89 Tm (requests at a steep discount - Anthropic charges roughly 90% less for cached reads, and OpenAI) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 580.89 Tm (charges about half price. For teams running agentic workflows with long, repeated system prompts,) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 565.89 Tm (this single setting can meaningfully cut the bill without touching the model or the workflow at all.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 543.89 Tm (The other half of this is context management - being deliberate about what gets sent to the model in the) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 528.89 Tm (first place. Pasting an entire file when only ten lines matter, keeping stale conversation history alive) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 513.89 Tm (across many turns, or letting an agent re-read the whole codebase on every step all add up fast. Some) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 498.89 Tm (coding tools now handle this automatically by trimming and managing context behind the scenes, which) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 483.89 Tm (is worth checking for when comparing tools, since it can be the difference between a predictable bill) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 468.89 Tm (and a surprising one.) Tj ET BT /F2 15 Tf 0.08 0.1 0.14 rg 1 0 0 1 46 440.89 Tm (Strategy 3: Set Budgets and Watch Spend in Real Time - Not After the) Tj ET BT /F2 15 Tf 0.08 0.1 0.14 rg 1 0 0 1 46 421.89 Tm (Invoice) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 397.89 Tm (One pattern shows up again and again in the companies that overspent badly: they had no owner for AI) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 382.89 Tm (costs and no alert system until the bill was already large. A 2026 survey of 700 engineering leaders) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 367.89 Tm (found that more than half of organizations still have no clear owner for AI spending. That's the gap) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 352.89 Tm (causing most of the pain, not the technology itself.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 330.89 Tm (Companies handling this well treat AI spend the way they treat cloud infrastructure spend - with the) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 315.89 Tm (same discipline, not less. That typically includes a token budget for each team or workflow, automatic) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 300.89 Tm (alerts when usage jumps well above someone's normal pattern, and hard shutoff limits on agentic) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 285.89 Tm (workflows that could otherwise run unattended and rack up charges without anyone noticing. Zapier,) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 270.89 Tm (for example, flags any employee whose AI usage runs five times higher than their peers, then checks) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 255.89 Tm (whether that usage is genuinely productive or simply wasteful, rather than assuming either by default.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 233.89 Tm (The goal isn't to restrict developers from using AI. It's to catch the runaway cases - a poorly) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 218.89 Tm (configured agent looping endlessly, or a workflow accidentally sending far more context than it needs) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 203.89 Tm (- before they turn into a five- or six-figure surprise at the end of the month.) Tj ET BT /F2 15 Tf 0.08 0.1 0.14 rg 1 0 0 1 46 175.89 Tm (Strategy 4: Match Your Pricing Plan to How Your Team Actually Works) Tj ET q 0.86 0.88 0.92 RG 1 w 46 42 m 549.28 42 l S Q BT /F1 8.4 Tf 0.42 0.45 0.5 rg 1 0 0 1 46 30 Tm (TechRounder | Page 3 of 5) Tj ET BT /F1 7.2 Tf 0.42 0.45 0.5 rg 1 0 0 1 46 19 Tm (https://www.techrounder.com/pdf/blog/how-are-software-companies-cutting-ai-costs-without-slowing-down-development.pdf) Tj ET endstream endobj 11 0 obj << /Type /Page /Parent 2 0 R /MediaBox [0 0 595.28 841.89] /Resources << /Font << /F1 3 0 R /F2 4 0 R >> >> /Contents 12 0 R >> endobj 12 0 obj << /Length 5743 >> stream BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 789.89 Tm (AI coding tools are priced in a few fundamentally different ways, and picking the wrong one for your) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 774.89 Tm (team's usage pattern is its own hidden cost. Flat per-seat subscriptions \(commonly $20 to $40 per) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 759.89 Tm (developer per month\) are predictable but can throttle heavy users or push them into paying API rates) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 744.89 Tm (once they hit a usage cap. Pure usage-based billing offers no ceiling and scales with exactly what you) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 729.89 Tm (use, but it also has no ceiling if something goes wrong. A growing number of providers now offer a) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 714.89 Tm (hybrid - a modest seat fee plus metered usage on top - which tends to fit most teams best because it) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 699.89 Tm (keeps light users cheap while still tracking heavy usage transparently.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 677.89 Tm (The practical move here is simple: look at your actual daily usage data for a few weeks before) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 662.89 Tm (committing to a plan, and revisit that choice every quarter. GitHub Copilot's move to usage-based "AI) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 647.89 Tm (Credits" billing in June 2026, and similar shifts from other providers, mean the plan that made sense) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 632.89 Tm (six months ago may not be the cheapest option anymore.) Tj ET BT /F2 15 Tf 0.08 0.1 0.14 rg 1 0 0 1 46 604.89 Tm (Strategy 5: Don't Let Speed Create Costs You'll Pay for Later) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 580.89 Tm (This is the part that's easy to overlook when the goal is "move faster." AI-generated code that ships) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 565.89 Tm (quickly but poorly reviewed doesn't actually save money - it just moves the cost downstream.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 550.89 Tm (Research analyzing over 200 million lines of code found that AI-heavy codebases see roughly four) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 535.89 Tm (times more code duplication, and the amount of time spent on refactoring dropped from around a) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 520.89 Tm (quarter of engineering work to under 10%. Separately, AI-generated pull requests have been found to) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 505.89 Tm (carry security vulnerabilities at close to three times the rate of human-written code.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 483.89 Tm (Companies avoiding this trap treat automated testing, linting, and structured code review as) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 468.89 Tm (non-negotiable guardrails that run alongside AI-generated code, not as optional extras. Several are also) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 453.89 Tm (adopting AI-powered code review tools specifically built to catch what a human reviewer might rush) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 438.89 Tm (past when the volume of AI-generated code is high. The logic is straightforward: a bug caught in) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 423.89 Tm (review costs a few minutes. The same bug caught in production, after it's already shipped and) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 408.89 Tm (possibly caused an incident, costs vastly more - in both time and reputation.) Tj ET BT /F2 15 Tf 0.08 0.1 0.14 rg 1 0 0 1 46 380.89 Tm (Strategy 6: Measure What the AI Actually Produced, Not Just How Much It) Tj ET BT /F2 15 Tf 0.08 0.1 0.14 rg 1 0 0 1 46 361.89 Tm (Was Used) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 337.89 Tm (Given how mixed the productivity data has turned out to be, one of the most important shifts companies) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 322.89 Tm (are making is simply measuring the right thing. Usage dashboards showing how many prompts were) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 307.89 Tm (sent or how many seats are active tell you almost nothing about whether AI is actually helping. What) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 292.89 Tm (matters is output: merged pull requests, defect rates on AI-assisted code, review confidence, and -) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 277.89 Tm (critically - cost attributed to what actually got shipped, not just tokens consumed.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 255.89 Tm (Some companies now track cost per merged pull request as a core metric, which immediately reveals) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 240.89 Tm (whether a team's AI spend is translating into real progress or just generating a lot of activity. This) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 225.89 Tm (kind of tracking also makes the model-routing and budget decisions above much easier, because) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 210.89 Tm (you're no longer guessing - you can see exactly which workflows are worth the spend and which) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 195.89 Tm (ones need to be reined in.) Tj ET BT /F2 15 Tf 0.08 0.1 0.14 rg 1 0 0 1 46 167.89 Tm (A Simple Starting Point for Smaller Teams) Tj ET q 0.86 0.88 0.92 RG 1 w 46 42 m 549.28 42 l S Q BT /F1 8.4 Tf 0.42 0.45 0.5 rg 1 0 0 1 46 30 Tm (TechRounder | Page 4 of 5) Tj ET BT /F1 7.2 Tf 0.42 0.45 0.5 rg 1 0 0 1 46 19 Tm (https://www.techrounder.com/pdf/blog/how-are-software-companies-cutting-ai-costs-without-slowing-down-development.pdf) Tj ET endstream endobj 13 0 obj << /Type /Page /Parent 2 0 R /MediaBox [0 0 595.28 841.89] /Resources << /Font << /F1 3 0 R /F2 4 0 R >> >> /Contents 14 0 R >> endobj 14 0 obj << /Length 2262 >> stream BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 789.89 Tm (Not every company has the resources to build a full AI cost-governance program on day one, and) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 774.89 Tm (that's fine. The companies that get this right usually started small. A reasonable first step is picking) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 759.89 Tm (one or two high-volume, repetitive workflows - code review comments or test generation are) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 744.89 Tm (common choices - and routing just those to a cheaper model while keeping everything else on the) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 729.89 Tm (current setup. From there, turning on prompt caching wherever the same context repeats is close to a) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 714.89 Tm (free win, since it requires no change to how developers actually work. Once those two things are in) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 699.89 Tm (place, adding a basic weekly spend review - even a simple spreadsheet pulling numbers from the) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 684.89 Tm (provider's dashboard - catches most runaway costs long before they become a real problem.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 662.89 Tm (The companies that struggled hardest in 2025 and 2026 weren't the ones that adopted AI too slowly.) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 647.89 Tm (They were the ones that adopted it without asking, at every step, "is this actually necessary, and are) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 632.89 Tm (we tracking what it costs?" That single habit - treating AI spend as something to actively manage rather) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 617.89 Tm (than something to discover later - is really the whole strategy, dressed up in different tools and) Tj ET BT /F1 11 Tf 0.14 0.16 0.2 rg 1 0 0 1 46 602.89 Tm (techniques.) Tj ET q 0.86 0.88 0.92 RG 1 w 46 42 m 549.28 42 l S Q BT /F1 8.4 Tf 0.42 0.45 0.5 rg 1 0 0 1 46 30 Tm (TechRounder | Page 5 of 5) Tj ET BT /F1 7.2 Tf 0.42 0.45 0.5 rg 1 0 0 1 46 19 Tm (https://www.techrounder.com/pdf/blog/how-are-software-companies-cutting-ai-costs-without-slowing-down-development.pdf) Tj ET endstream endobj xref 0 15 0000000000 65535 f 0000000015 00000 n 0000000064 00000 n 0000000147 00000 n 0000000217 00000 n 0000000292 00000 n 0000000434 00000 n 0000006230 00000 n 0000006372 00000 n 0000012895 00000 n 0000013038 00000 n 0000018674 00000 n 0000018818 00000 n 0000024613 00000 n 0000024757 00000 n trailer << /Size 15 /Root 1 0 R >> startxref 27071 %%EOF