If you have been asked to prove the AI coding budget pays off, spend nothing on a measurement platform for the first quarter. The free GitHub Copilot and Claude Code dashboards now attribute merged pull requests to their own tool, and a quarterly developer survey covers what telemetry misses. To compare AI-assisted pull requests with the rest across several AI tools, buy Swarmia: it publishes its price, names Copilot, Cursor, Claude Code and Codex as supported out of the box, and sells the AI impact module alone. Buy DX on Atlassian for survey-based time savings your CFO will accept, and Jellyfish if finance wants software capitalization and R&D tax credit reports. LinearB fits worst: most of its seat price buys AI code review and workflow automation.
All prices below were checked on each vendor's live pricing page on November 21, 2026. The workload is the same for every row: 200 engineers on GitHub Cloud, using two AI coding tools, on an annual contract.
| Method | Attributes code to an AI tool? | Published price (Nov 21, 2026) | List cost at 200 engineers/yr | Pick it if | Skip it if | |
|---|---|---|---|---|---|---|
| Vendor dashboards (Copilot, Claude Code) | Tool telemetry matched to merged PRs | Yes, per tool, for that tool only | Included in the tool's seat | $0 extra | You run one or two AI tools on github.com | You need one view across tools, or delivery outcomes |
| Swarmia | Git and PR telemetry plus AI tool integrations | Compares AI-assisted vs non-assisted PRs | Standard $45, AI impact plan $23, per developer/month, annual | $108,000 (Standard) or $55,200 (AI impact plan alone) | You want a published price and cross-tool PR comparison | You need finance-grade cost allocation |
| DX | Developer surveys plus system data | Tracks AI-generated code by commit, PR and agent, per DX | Quote only, via demo request | Contact sales | You are standardized on Atlassian and want survey-based time savings | You will not run surveys every quarter |
| Jellyfish | Git, planning and workflow data, plus a DevFinOps module | Derives AI signals from SDLC data; method not published | Quote only, by seats and modules | Contact sales | Your CFO needs software capitalization and R&D tax credit reports | You only need to answer "did delivery change" |
| LinearB | Git and PR telemetry plus AI review credits | "Unified view" of AI and human delivery; method not published | Essentials $29, Enterprise $59, per user/month, annual | $69,600 (Essentials) or $141,600 (Enterprise) | You also want AI code review and PR automation on one bill | Measurement is the only job |
Why Adoption Dashboards Cannot Answer the ROI Question
An adoption dashboard tells you who used the tool; it cannot tell you whether more work shipped, and the best independent study of AI coding agents shows the two numbers moving very differently. Across 720,542 workers at 655 companies, agents raised pull requests opened by 35.7% while resolved Jira issues rose only 13.7%, a borderline result, according to the November 2026 revision of the Harvard working paper built on Jellyfish data. Time from pull request submission for review to merge rose 56.1%. We covered that study in detail in our write-up of the Harvard Jellyfish paper.
The survey evidence points the same way. Google's 2025 DORA report, drawn from nearly 5,000 technology professionals, found 90% use AI at work and more than 80% believe it raised their productivity. It also found AI adoption now has a positive relationship with delivery throughput and a negative one with delivery stability. Faster output with more broken changes is exactly what a dashboard of acceptance rates and lines accepted cannot see.
Self-report is the weakest evidence of all. In METR's 2025 randomized trial, 16 experienced open-source developers took 19% longer to finish issues when AI was allowed, and afterwards still believed it had sped them up by 20%. METR's February 2026 update called its newer data "only very weak evidence" for the size of any speedup, because some developers refused to work without AI and 30% to 50% withheld some tasks from the study. That is a warning about every measurement program: the developers who like a tool also decide which work it gets used on, which skews any before-and-after comparison.
Code quality gives you a third check. GitClear's analysis of 211 million changed lines found refactoring fell from 25% of changed lines in 2021 to under 10% in 2024, while copy-pasted lines rose from 8.3% to 12.3%. GitClear sells code analytics, so weigh it as a vendor's research, but the direction matches DORA.
Use these four as your benchmarks. If a vendor's case study shows a gain far outside them, ask how the comparison group was chosen.
What the Free Vendor Dashboards Already Give You
The built-in dashboards now attribute merged code to their own tool. Their limit is that each one sees only its own tool.
Claude Code has a dashboard for Team and Enterprise plans that counts "PRs with Claude Code," defined as merged pull requests containing at least one line written with Claude Code, per Anthropic's analytics docs. It matches session edits to merged diffs using sessions from 21 days before to 2 days after the merge, labels matched PRs claude-code-assisted on GitHub, and describes its counts as "deliberately conservative." The contribution metrics are in public beta, cover github.com repositories only, and are unavailable to organizations on Zero Data Retention. For raw events, Claude Code can export OpenTelemetry metrics such as sessions, lines changed, commits, pull requests and cost to a collector you run, per its monitoring docs.
GitHub Copilot offers a usage dashboard, a code generation dashboard that "breaks down code generated by users and agents," an impact dashboard that connects adoption depth with pull request output, and repository insights, per GitHub's Copilot metrics docs. The metrics REST API returns daily repository reports that break out Copilot coding agent and Copilot code review activity, keeps history for up to one year, and has data from October 10, 2025 onward. The API requires the enterprise's Copilot usage metrics policy to be set to "Enabled everywhere."
Each additional tool, such as Cursor or Codex, reports through its own console with its own definition of "AI-assisted," and none of those views knows your cycle time or change failure rate. The four platforms below exist to close that gap.
When the free option is enough: you run one or two AI tools, your code is on github.com, and the question is "are people using it and is it in merged code." Add a short quarterly survey asking time saved per week and satisfaction, using the questions in DX's published AI Measurement Framework, which is free to read and organizes measurement around utilization, impact and cost. Pull change failure rate and PR cycle time from your existing CI and Git host. For most teams under about 100 engineers, that answers the board's question.
Swarmia Is the Default Paid Pick
Swarmia is the platform to buy first: it is the only one of the four that publishes a price for a standalone AI impact plan, and it says it compares AI-assisted pull requests directly against the rest. Its pricing page says it can "compare how AI-assisted PRs perform against non-assisted ones across cycle time, throughput, and batch size," and its AI impact page says it connects with Claude Code, GitHub Copilot, Cursor and Codex out of the box. It also tracks review agents by agent review rate and findings per PR, which matters given that review is where the Harvard study found gains stall.
On Swarmia's pricing page, checked November 21, 2026, Standard is $45 and Enterprise $55 per developer per month, billed annually. Narrower plans include AI adoption and cost at $5, developer surveys at $9, software capitalization at $18 and productivity and AI impact at $23, each per developer per month, billed annually. Swarmia is free for companies with up to 9 developers. Its billing FAQ says the seat count is either team members added in Swarmia or active authors with a non-excluded PR in the last 30 days, depending on the contract, so ask which one you are signing.
At 200 engineers, the AI impact plan alone lists at $55,200 a year and Standard at $108,000.
Who should not buy Swarmia: a company whose CFO wants engineering spend broken down by product line for capitalization and tax credits. Swarmia sells a capitalization plan, but Jellyfish puts more of its product into finance reporting. Also skip it if no manager will act on per-team cycle time data.
DX Fits Atlassian Shops That Will Run Surveys
DX measures through developer surveys combined with system data, which makes it the strongest choice for estimating hours saved and the weakest for teams that will not survey consistently. Its product page says it can "track AI-generated code by commit, PR, team, agent, and repo." The framework treats agents as extensions of the teams that run them, so team PR throughput includes agent-authored PRs, and it recommends measuring time saved per developer per week alongside DX's Core 4 delivery metrics.
DX cites customer results on its framework page: Booking.com, with more than 3,500 engineers, reported a 16% increase in throughput. That is a vendor case study, with no control group published alongside it.
The ownership matters for your contract. Atlassian agreed to buy DX for about $1 billion in cash and restricted stock in September 2025, when DX had more than 350 enterprise customers including GitHub. Atlassian said DX would be integrated into its product suite, and TechCrunch's report did not say whether support for rival toolchains would continue. If you track work outside Jira, get continued support for your integrations written into the order form.
DX's pricing page lists no prices; the only path is a demo request.
Who should not buy DX: a team without the discipline to run a survey every quarter and act on it. Survey-led measurement loses its value when response rates fall, and a single annual survey cannot show whether a tool rollout changed anything.
Jellyfish Fits Finance-Led Measurement Programs
Jellyfish is the right buy when finance needs AI spend and engineering cost reporting in the same tool, and a poor one when you only need to know whether delivery changed. Its AI Impact page says it will "link AI usage to throughput, quality, and delivery speed using real SDLC signals" and covers adoption, impact, AI spend and multi-tool comparison. It derives AI signals from Git, planning tools and workflow data without tagging, but does not publish how it matches AI usage to specific pull requests. Until a vendor shows you that matching, assume you are looking at correlation: teams with high usage compared against teams with low usage.
Jellyfish's research credibility is real. Its dataset underpins the Harvard paper above, though the paper's own note says Jellyfish reviewed the manuscript to check it disclosed nothing proprietary. The paper shows the data set is deep. Whether the product attributes your own PRs correctly is something to test during the trial.
Jellyfish's pricing page says pricing "is based on number of seats and specific modules selected," with AI Impact, Developer Productivity (which includes AI Impact) and DevFinOps sold as modules. No prices are published.
Who should not buy Jellyfish: an engineering leader who needs an answer this quarter and has no finance stakeholder asking for allocation. Mapping work to finance categories takes setup time that a cycle-time question does not need.
LinearB Is the Loser for Pure Measurement
LinearB publishes clear prices, but its seat now bundles AI code review credits, so a buyer who only wants to measure AI impact pays for automation they did not ask for. On LinearB's pricing page, checked November 21, 2026, Essentials is $29 per user per month with 1,000 monthly AI credits per user, a 50-user minimum and GitHub Cloud only; Enterprise is $59 per user per month with 1,500 credits and a 100-user minimum, both billed annually. Its credits page says one PR consumes 100 credits, sells extra packs of 5,000 credits for $67.50, and lets you exceed your allocation free in two months per contract year, capped at 120%, after which automations pause until the monthly reset or a credit purchase.
The measurement feature in both plans is "AI Impact Measurement," described as a dashboard "offering a unified view of both AI and human-driven code delivery." The page does not explain how it decides which code is AI-driven, and it does not name Copilot, Cursor or Claude Code. At 200 engineers, LinearB lists at $69,600 a year on Essentials, less than Swarmia Standard, but you are buying a review and automation product with metrics attached.
Who should pick LinearB anyway: a team on GitHub Cloud that wants AI code review and PR workflow automation and would otherwise buy a separate review tool such as CodeRabbit. For that team the metrics come bundled with a tool it was going to buy anyway.
Which Metrics Hold Up and Which Get Gamed
The metrics that game easily are the ones AI inflates for free: lines of code, PR count, commits and acceptance rate. The ones that hold are review time, cycle time and change failure rate, because an agent cannot improve them without the work actually getting better.
Acceptance rate is the clearest case. DX's own CTO, Laura Tacho, called it "the new 'lines of code' measurement: easy to measure, easy to misuse, and largely irrelevant to business value or team productivity." LeadDev's reporting quotes GitLab's CTO describing it as "a measure of engagement." Once you put it in a target, developers accept suggestions to hit it.
PR count fails for the reason the Harvard data shows: agents raised PRs 35.7% and completed work 13.7%. Report PR count only next to resolved issues, so the gap is visible.
Three metrics are harder to fake:
- Review time per PR. If it rises after an agent rollout, your bottleneck has moved to senior reviewers, and buying more agent seats will make it worse.
- Change failure rate. DORA found AI adoption tied to lower stability. If this rises, part of any throughput gain is rework.
- Cycle time from first commit to production, split by AI-assisted and non-assisted PRs. This is the comparison Swarmia sells and the one the free dashboards cannot produce.
Never use any of these to rank individual engineers. DX's framework advises against it, and per-person leaderboards are the fastest way to get the gaming you are trying to avoid.
How to Decide Without Regretting It
The criterion that best predicts regret is whether the platform can show you its attribution method on your own repositories during the trial.
Ask every vendor these four questions in the demo:
- How do you decide a PR is AI-assisted? Session matching, a label, a commit trailer, or usage by team?
- What happens to a PR touched by two tools?
- Can I see change failure rate and review time split by AI-assisted status?
- What is the billable unit: seats in teams, or active authors in the last 30 days?
Your situation changes the answer:
- With one AI tool and under 100 engineers, use the vendor dashboard plus a quarterly survey, and do not buy a platform.
- With two or more tools and 100 to 1,000 engineers, start on Swarmia's AI impact plan and move to Standard if managers use it.
- If you are standardized on Atlassian and already survey developers, pick DX.
- If finance needs capitalization and R&D tax credit reporting, pick Jellyfish.
- If you also want AI code review on the same contract, pick LinearB.
If you are still choosing the coding tools themselves, start with our coding agent price comparison, since the measurement bill should be a small fraction of the tool bill.
What to Do This Quarter
This Week: Turn on the Copilot usage metrics policy and, on Claude Team or Enterprise, install the GitHub app for contribution metrics. Export 90 days of data.
This Month: Pull PR review time and change failure rate for the same 90 days from your Git host and CI. Send a five-question survey on time saved per week. Put the three side by side.
Before Q1 Close: If the board wants a cross-tool answer, trial Swarmia's AI impact plan on two teams, one heavy AI user and one light. Ask it to show which PRs it marked as AI-assisted and check ten of them by hand before you sign.
Continue Reading
- AI Coding Agents Lift Pull Requests 35.7%, Harvard Study Finds
- Your Copilot Bill Can Reach $137 a Month at 10 Daily Tasks
- Barclays Picks Claude for Half Its Developers, Copilot for Desks
- Cursor Alternatives That Survive Security Review Start With Copilot
- GitHub Enterprise Adds Copilot, Bills From First Use
- Fable Cost $18 a Fix on Private Code. Flash Cost $8.
