ETHORITY®GOOD AI PRISM · TOPIC AREA · SOFTWARE DEVELOPMENT
84% use coding AI. 46% distrust it.
Compare coding tools on billing unit, training terms and data location, and measure lead time before trusting any speed claim.
- 24
- AI offerings placed in Software DevelopmentGood AI Prism, Oct 2026
- 84%
- of developers use or plan to use AI toolsStack Overflow, Jul 2025 · survey.stackoverflow.co ↗
- 46%
- actively distrust the accuracy of AI toolsStack Overflow, Jul 2025 · survey.stackoverflow.co ↗
- −19% to +21%
- measured effect of AI on developer speedMETR, Jul 2025; Google RCT · metr.org ↗
Felt speed is not measured speed.
Developers in the METR trial believed AI made them 20% faster while it made them 19% slower.
METR, Jul 2025
Prices come in five units or more.
The 12 profiled coding offerings bill in tokens, credits, seats, plan tiers or compute only.
Good AI Prism Registry, 20 Oct 2026
Models retire under your tools.
OpenAI shuts down its GPT-4o and GPT-4.1 snapshots on this date; tools built on them change too.
OpenAI deprecations, via Good AI Prism Registry
How to use this page
Three tools and one offer turn this page into a decision.
- Open the Software Development map: 24 offerings on one petal of the Good AI Prism, with logos and provider facts.
- Compare up to three offerings: decision facts side by side, each with its source and date.
- Run the Task economics calculator: turn saved hours into capacity you can attribute, before a pilot is approved.
- Offer: AI Decision Brief: two or three coding tools compared fact by fact, with sources, dates and the questions to ask before signing.
Twelve coding offerings, fact by fact
Billing unit, price, training terms, data location and certifications, each with source and date.
Grok Build (grok-code-fast-1)
xAI · Model
xAI's grok-code-fast-1 (now aliased to grok-build-0.1) is a fast reasoning model with native tool calling, built to run multi-step agentic coding loops inside coding tools via an OpenAI-compatible API.
- Published price
- Context window
- Model access
- Deployment requirements
Morph Fast Apply
Morph · Model
Model that replaces the file-writing layer in AI coding agents and IDE plugins, available as an MCP tool for Claude Code, Cursor and similar environments.
- Published price
- Context window
- Licence and access rights
- Model access
Mellum
JetBrains · Model family
JetBrains' 4B-parameter multilingual model, trained from scratch for code completion, powering cloud-based completion in JetBrains IDEs.
- Published price
- Licence and access rights
- Model access
- Deployment requirements
voyage-code
Voyage AI (MongoDB) · Model family
Voyage AI's voyage-code family (code-4 current; code-3, code-2 older) turns source code into Matryoshka-sized, quantisable embedding vectors so coding agents and code search can retrieve relevant files.
- Published price
- Context window
- Licence and access rights
- Model access
Auggie CLI
Augment Code · Agent
Terminal agent from Augment Code that brings its codebase context engine to the command line, running interactively or in --print mode for CI pipelines such as GitHub Actions PR reviews.
- Published price
- How it is billed
- Does it train AI on your data?
- Certificates and audit reports
CodeRabbit
CodeRabbit · Product
AI pull-request reviewer for GitHub, GitLab, Bitbucket and Azure DevOps that automatically reviews every PR, ranks incoming PRs via Triage and opens security fixes as PRs.
- Published price
- How it is billed
- Does it train AI on your data?
- Connects with
Bolt
StackBlitz · Product
StackBlitz tool, from the team behind WebContainers, that builds full-stack web applications using a wide range of JavaScript web frameworks.
- Published price
- How it is billed
- Does it train AI on your data?
- Certificates and audit reports
Tricentis Tosca AI agents
Tricentis · Product
According to Tricentis: “AI agents built for the Tosca environment generate complete, end-to-end tests from plain-language descriptions.”
- How it is billed
- Where data is stored
- Does it train AI on your data?
- Certificates and audit reports
GitLab Duo Agent Platform
GitLab · Platform
GitLab's agent layer for GitLab.com, Self-Managed and Dedicated where agents and flows turn issues into merge requests, review code and fix CI/CD pipelines, billed via GitLab Credits.
- Published price
- How it is billed
- Does it train AI on your data?
- Certificates and audit reports
Snyk DeepCode AI
Snyk · Product
According to Snyk: “Snyk's DeepCode AI code analyzer was designed to find, autofix, and prioritize vulnerabilities”
- Published price
- How it is billed
- Where data is stored
- Does it train AI on your data?
SonarQube AI CodeFix
Sonar · Product
According to Sonar: “AI CodeFix then proposes an edit that resolves the problem without changing the code's functionality.”
- Published price
- How it is billed
- Where data is stored
- Certificates and audit reports
Zencoder
Zencoder · Platform
Zencoder runs multi-model coding agents in VS Code and JetBrains IDEs and as headless CI workers, indexing several repositories to turn Jira, Linear or GitHub Issues tickets into multi-file changes.
- Published price
- How it is billed
- Does it train AI on your data?
- Certificates and audit reports
Compare baseline: three units, one team
Per seat, per credit and per month with credits: convert all three to cost per merged change.
| Offering | Price | Billing | Training | Source |
|---|---|---|---|---|
| CodeRabbit | $24–72 /user/mo | Per user per month | Not used for training | CodeRabbit pricing page |
| GitLab Duo Agent Platform | $1 per credit | GitLab Credits (usage) | Not used for training | GitLab Duo pricing page |
| Snyk DeepCode AI | Free · from $25 /mo | Monthly · credits (Ent.) | Not used for training | Snyk DeepCode AI documentation |
Registry facts as stated on each provider's own pages, read 21 Oct 2026. Add your own volumes in Compare before deciding.
Rules and dates
Model retirements and price changes reach your tools first.
- 2 Aug 2025
Obligations for general-purpose AI models apply.
Reg. (EU) 2024/1689
- 2 Aug 2026
Art. 50: people must be told when they interact with an AI system.
Reg. (EU) 2026/1744
- 30 Nov 2026
Anthropic retires the claude-sonnet-4-5-20250929 model snapshot.
Anthropic model deprecations
- 1 Dec 2026
Gemini 4 Argon list price doubles to US$4 / US$20 per million tokens.
Google Gemini API pricing
- 14 Jan 2027
OpenAI shuts down its GPT-4o and GPT-4.1 snapshots; tools built on them change.
OpenAI deprecations
Read the evidence
Global edition · EN
961 AI offerings. One map. Decide now what you will use in 2027.
The Good AI Prism helps you find useful AI and compare the evidence behind your options. For organisations already using AI, the question is moving from whether to use AI to which AI, how safely and whether it pays. This Global edition brings together 961 AI offerings across 24 topic areas, 8 data types and 6 technology layers, with sources, dates and limits to guide your next decision.
Version 2.6 · 90 pages
Open the study →
Global edition · EN
One model is not enough. Large enterprises already run three or more.
81% of large enterprises use three or more model families. Who builds AI models, what they cost, how long they last, and the rules by jurisdiction.
Version 1.5 · 38 pages
Open the study →
Germany & United States Focus · EN
Germany's mid-sized companies meet Silicon Valley: one AI market, two different questions.
The Mittelstand needs know-how; US firms need cost control. What German and US businesses need from AI now, sector by sector, function by function.
Version 2.6 · 72 pages
Open the study →
Global edition · EN
Compute that fits the work you run.
Capacity, region and energy terms count as much as chip performance. Where AI chips and capacity come from, what the market offers and the energy terms by jurisdiction.
Version 1.5 · 40 pages
Open the study →Related insights
961 AI offerings. Start with the job, not the tool.
A method for going from one business need to a shortlist of AI offerings you can defend.
See the map on one page →
Saved hours are not AI value.
Separate a better task, freed-up time and a business result, so you can show the value AI really delivers.
See the map on one page →
Check every vendor claim before you rely on it.
Tie every vendor claim to the exact product, version, contract and use before you rely on it.
See the map on one page →
AI does not scale by itself. Redesign the work.
Start with one objective, choose the simplest approach for each task and measure the cost of every result you accept.
See the map on one page →Implement it with ETHORITY
AI Decision Brief
€195 net · one-time
Puts the facts that decide between two or three AI offerings side by side, each sourced and dated, with the questions to ask each provider.
Request scope and delivery
See what this package includes →Enterprise AI Opportunity Case
Priced after scoping
Compares real options, including doing nothing, models the economics and the assumptions that can break them, and recommends: start, change, test first or stop.
Scope and delivery by agreement
See what this package includes →AI Literacy & Adoption
€2,400 one-time · net
Builds AI literacy and everyday adoption in your teams, in step with the AI-literacy duty of the EU AI Act.
Scope and delivery by agreement
See what this package includes →Buyer questions
Do AI coding tools make developers faster?
Sometimes. Controlled studies range from 19% slower (METR trial, Jul 2025) to about 21% faster (Google RCT). Measure lead time and change failure rate for your own team before and after.
Is our code used to train the provider's models?
It depends on the plan. Six of the 12 profiled offerings state customer code is not used for training; one allows training on free-account data. Read the terms of the plan you will buy, as the Zencoder registry card shows.
How do we compare prices billed in tokens, credits and seats?
Convert every option into cost per accepted change at your own volume, including review time. The buyer's guide to 961 offerings sets out the method.
Which tools keep data in the EU?
Three of the 12 profiled cards name an EU region or EU default: Snyk, SonarQube and Tricentis. The SonarQube AI CodeFix card lists Frankfurt as its default.
What happens when the model under a tool retires?
Behaviour, price or quality can change. Record the model behind each tool and its retirement date; the AI Models technology brief lists the current dates.
The evidence at a glance
ETHORITY · LONG READ
Coding AI: 19% slower to 21% faster. Measure yours.
84% of developers use AI and 46% distrust it. Controlled studies disagree on the gain. The Good AI Prism shows what to compare instead: billing units, training terms, data location and the models underneath.
Software teams use AI every day. Most of them do not trust it yet. In the Stack Overflow 2025 survey, 84% of developers said they use or plan to use AI tools in their development process. In the same survey, 46% said they actively distrust the accuracy of those tools, against 33% who trust it. Two thirds, 66%, named their biggest frustration: AI solutions that are "almost right, but not quite".
That gap between use and trust is the market finding of 2026 in software development. It decides what a buyer should ask for, what a team should measure and which offerings belong on a shortlist.
Measured effects point both ways
The best evidence on coding AI comes from controlled studies, and they disagree. In a randomised trial published by METR in July 2025, 16 experienced open-source developers worked on 246 real issues in their own repositories. With AI tools allowed, they took 19% longer. Before the study they had expected a 24% speed-up. Afterwards, they still believed AI had made them 20% faster.
A randomised trial at Google found the opposite direction: engineers with AI features finished an enterprise task about 21% faster. The Global Market Study 2026/27 puts both results side by side for a reason. Both are real. They measure different work, different developers and different tools.
−19% to +21%. The measured range of AI effects on software work. Your own number sits somewhere in it, and only your own baseline shows where.
The lesson for buyers is narrow and practical. A supplier's productivity claim is not your baseline. The effect depends on the codebase, the task type and the developer's experience. Teams that measure lead time and change failure rate before and after adoption know where they stand. Teams that rely on how fast the work feels can be badly wrong: in the METR study, felt and measured speed were 39 points apart.
Where the money and the agents go
Software engineering is one of the places where companies report a return. In the McKinsey State of AI 2025, respondents most often reported cost benefits in software engineering, manufacturing and IT. The August 2026 edition adds that organisations scale AI agents most often in IT, knowledge management and software engineering.
Spending follows. Gartner's September 2026 AI spending forecast puts AI application development platforms at US$9.5bn in 2026, up from US$6.9bn in 2025, and generative AI models at US$28.3bn. Coding tools sit on top of those models, so the model layer is part of every coding decision. The AI Models technology brief shows why: 81% of large enterprises already use three or more model families (a16z, Jan 2026).
What the Prism shows about coding offerings
The Good AI Prism places 24 AI offerings in Software Development, from code completion and review to testing agents and code-specific embeddings. Twelve of them carry a full registry card with sourced decision facts, read from provider pages on 20 Oct 2026.
Exhibit: twelve profiled coding offerings, by the facts a buyer needs
| Decision fact | What the 12 cards show |
|---|---|
| Billing unit | At least five units: tokens, credits, seats, plan tiers and compute only |
| Training on your code | 6 state customer code is not used for training; others offer opt-in or opt-out terms |
| EU data location | 3 state an EU region or EU default (Snyk, SonarQube, Tricentis) |
| Self-hosting | 6 offer self-hosted, self-managed, on-prem or open weights |
| Published price | 11 publish a list price |
Source: Good AI Prism Registry, provider pages read 20 Oct 2026.
Three patterns stand out. First, prices come in units that cannot be compared directly. CodeRabbit's registry card bills US$24–72 per user per month. GitLab Duo bills US$1 per credit. Grok's coding model bills US$1 per million input tokens and US$2 per million output tokens. A team of 40 developers needs all three converted into cost per merged change before the comparison means anything.
Second, training terms differ even inside one product. One profiled assistant states in its terms that trial and paid data is used for training only with further consent, while data from free accounts may be used. A pilot on a free plan can therefore set different terms from the contract you later sign.
Third, data location is rarely the headline. Only three of the twelve cards name an EU region or an EU default. SonarQube AI CodeFix states Frankfurt as its EU default. For a European buyer with client code under contract, that single fact can decide the shortlist.
Scenario. A team of 40 developers on CodeRabbit at US$24–72 per user pays US$960–2,880 a month. If the team merges 1,200 changes a month, that is US$0.80–2.40 per merged change before review time. Run the same sum for credits and tokens, then compare.
Models retire, and your tools move with them
Coding tools inherit the life cycle of the models underneath. The registry records that OpenAI retires its GPT-4o and GPT-4.1 snapshots on 14 Jan 2027. Anthropic lists Claude Opus 5.5 at US$4 per million input tokens and US$20 per million output tokens, with retirement not before 22 Sep 2027. A coding assistant pinned to an older model may change behaviour, price or quality when its provider switches.
Ask every supplier which model and which cloud their product depends on, and how they announce a change. The buyer's guide to 961 offerings treats this as a standard question for every shortlist. In software, it matters twice: a model switch can change the code your team ships.
46% of developers actively distrust the accuracy of AI tools, while 84% use or plan to use them (Stack Overflow, 2025).
Pricing moves to usage
Seat licences are giving way to credits, tokens and actions. Gartner, cited by Deloitte in the Global Market Study, expects at least 40% of SaaS spend to be usage-, agent- or outcome-based by 2030. For engineering leaders, the budget question changes from "how many seats" to "how many accepted changes at what cost". Usage pricing rewards teams that know their volume. It punishes teams that discover it on the invoice.
The value question changes too. Saved minutes in an IDE are not yet business value. As the ETHORITY evidence synthesis Saved hours are not AI value argues, a time gain becomes value only when the released capacity goes into work that changes an outcome you can attribute.
What to do Monday
- Measure two numbers before any tool decision. Record lead time for changes and change failure rate for one team over four weeks. Without that baseline, no supplier claim can be checked.
- Compare three tools on cost per accepted change. Convert seats, credits and tokens to one unit at your own volume. Include review time for AI-written code, which the 66% "almost right" finding says will not be zero.
- Read the training and data terms for the plan you will actually buy. Free and trial tiers can carry different terms. Ask for the EU region in writing if client code is involved.
- Name the model under every tool. Record model, provider and announced retirement date, and set a review when a model changes.
- Set a stop condition. Agree in advance what result ends the pilot. If lead time and failure rate do not improve at your volume, stop and keep the evidence.
The Good AI Prism keeps the registry cards for this area current, with every decision fact sourced and dated. For a written comparison of two or three coding offerings, the AI Decision Brief offer puts the deciding facts side by side.
Sources and further reading
- Stack Overflow Developer Survey 2025: AI ↗ · Jul 2025
- METR: Measuring the impact of early-2025 AI on experienced open-source developer productivity ↗ · 10 Jul 2025
- Google: How much does AI impact development speed? An enterprise-based randomized controlled trial (arXiv 2410.12944) ↗ · 2024–2025
- McKinsey: The state of AI in 2025 (1,993 respondents, 105 countries) ↗ · 5 Nov 2025
- McKinsey: The state of AI, August 2026 edition (1,719 respondents, 97 countries) ↗ · 25 Aug 2026
- Gartner: Worldwide AI spending to grow 49.5% in 2026 ↗ · 16 Sep 2026
- a16z: enterprise AI survey (via Good AI Prism AI Models brief) ↗ · Jan 2026
- Good AI Prism Global Market Study 2026/27 (ETHORITY) → · 22 Oct 2026
- Good AI Prism Registry cards, provider pages read 20 Oct 2026 (ETHORITY) → · 20 Oct 2026
- OpenAI: API deprecations (GPT-4o and GPT-4.1 snapshots) ↗ · read 25 Oct 2026
Put the thinking to work
AI Opportunity Case →
A defensible proceed, test, redesign, defer or stop decision. The fee is independent of which outcome the evidence supports.
A written comparison of up to three AI offerings · €195 net →
Start with one question and up to three offerings. Request a sourced comparison, provider questions and a cost sketch using your assumptions.
YOUR NEXT STEP
Discuss your question with Sten Franke.
Bring your business question or the figures you already know. ETHORITY investigates within the agreed scope; together with Sten Franke you agree the next piece of work.