The State of AI Coding

A study of engineering workflows and AI software development.
Q2 2026 edition.

By Greptile Research · Published July 2, 2026

Table of Contents

Navigate the Report

Introduction

What changed in AI coding in Q2 2026

A year ago, the median pull request changed 68 lines of code. By June 2026, it changed 122. Over the same twelve months, lines changed per developer grew 3.5x, from 6,032 to 20,972. Anthropic's SDK reached 164 million monthly downloads, and OpenAI's download lead over Anthropic narrowed from 3.7:1 to 2.3:1 in six months.

The State of AI Coding: Q2 2026 tracks these shifts across engineering velocity, AI tool adoption, model provider trends, and the research shaping today's coding agents. It draws on Greptile's internal data and public npm and PyPI registries. Together, these signals show how quickly the way teams write, review, and ship code is changing, and what engineering leaders should watch next.

Section 01

Engineering Team Velocity

Measuring productivity gains across development workflows.

Chart 1.1

PRs Are Getting Bigger

Median PR size grew 79% from June 2025 to June 2026, rising from 68 to 122 lines changed per PR.

Captured from Greptile internal data engineering team velocity

View chart data

PRs Are Getting Bigger
Month	Median lines changed per PR
2025-Jun	68
2025-Jul	69
2025-Aug	72
2025-Sep	71
2025-Oct	80
2025-Nov	82
2025-Dec	89
2026-Jan	96
2026-Feb	106
2026-Mar	101
2026-Apr	113
2026-May	125
2026-Jun	122

Chart 1.2

Developer Output

Lines of code per developer grew 3.5x over the past year, from 6,032 to 20,972, a steady climb as AI coding tools act as a force multiplier.

Captured from Greptile internal data engineering team velocity

View chart data

Developer Output
Month	Lines changed per developer
2025-Jun	6,032
2025-Jul	6,801
2025-Aug	6,132
2025-Sep	6,184
2025-Oct	7,380
2025-Nov	7,562
2025-Dec	8,228
2026-Jan	10,349
2026-Feb	12,277
2026-Mar	14,338
2026-Apr	16,511
2026-May	19,322
2026-Jun	20,972

Chart 1.3

Medium Teams Output

Medium teams (6–50 devs) more than tripled output, from 7,576 to 27,584 lines per developer.

Captured from Greptile internal data engineering team velocity

View chart data

Medium Teams Output
Month	Lines changed per developer
2025-Jun	7,576
2025-Jul	9,207
2025-Aug	8,697
2025-Sep	8,331
2025-Oct	10,846
2025-Nov	10,272
2025-Dec	9,780
2026-Jan	12,141
2026-Feb	14,105
2026-Mar	21,333
2026-Apr	25,694
2026-May	25,608
2026-Jun	27,584

Chart 1.4

Lines Changed Per File

Median lines changed per file grew from 21 to 30 as PRs become denser.

Captured from Greptile internal data engineering team velocity

View chart data

Lines Changed Per File
Month	Median lines changed per file
2025-Jun	21
2025-Jul	21
2025-Aug	21.5
2025-Sep	21.3
2025-Oct	23
2025-Nov	23
2025-Dec	24
2026-Jan	24.9
2026-Feb	25.3
2026-Mar	25
2026-Apr	27
2026-May	29.3
2026-Jun	29.7

Section 02

AI Tool Adoption

Tracking the rise of AI-powered development tools.

Chart 2.1

AI Memory Packages

mem0 accounts for 58% of downloads among the five memory packages measured. LangMem slipped while Letta and Supermemory gained ground.

PyPI + npm monthly downloads, May 2026

View chart data

AI Memory Packages: share of selected package downloads
Package	May 2026 (%)	March 2026 (%)	Change (percentage points)
mem0	58.3	57.9	0.4
Zep	21.3	21.2	0.1
LangMem	11.3	14.7	-3.4
Supermemory	4.1	3.1	1
Letta	5	3.2	1.8

Chart 2.2

Vector DB Developer Installs

faiss and pgvector lead on developer installs. Weaviate sits mid-pack at 11% after excluding automated uv-installer traffic that had inflated its raw download count.

Open-source vector DBs, PyPI pip installs, June 2026. Automated uv-installer traffic excluded.

View chart data

Vector DB Developer Installs: share of selected package downloads
Package	June 2026 (%)	March 2026 (%)	Change (percentage points)
faiss	23.8	31	-7.2
pgvector	18.6	12.4	6.2
ChromaDB	16	15	1
Qdrant	15.8	14.5	1.3
Weaviate	11.5	10.5	1
LanceDB	9.5	8.3	1.2
Milvus	4.9	8.4	-3.5

Chart 2.3

AI Rules Files

CLAUDE.md present in 80% of Greptile orgs that use an AI rules file. AGENTS.md climbed to 63%, while Cursor Rules slid to 34%.

About 18% of orgs use all three formats, roughly flat since December (17%), even as CLAUDE.md and AGENTS.md adoption climbed

View chart data

AI Rules Files: independent adoption rates
Rules file	June 2026 (% of orgs)	March 2026 (% of orgs)	Change (percentage points)
CLAUDE.md	80	75	5
AGENTS.md	63	53	10
Cursor Rules	34	39	-5

Chart 2.4

AI SDK Growth

Anthropic SDK reached 164M monthly downloads by June 2026 (npm + pip). OpenAI Agents climbed to 17M.

PyPI (pip installer) + npm monthly downloads, Apr 2025 – Jun 2026

View chart data

AI SDK Growth: monthly downloads in millions
Month	Anthropic SDK	Vercel AI SDK	LlamaIndex	Pydantic AI	Instructor	LangChain JS	OpenAI Agents	CrewAI
2025-Apr	11.45	4.53	2.31	0.3	1.36	3.31	0.2	0.48
2025-May	12.28	5.76	1.99	0.39	1.49	3.52	0.34	0.55
2025-Jun	14.78	7.22	1.93	0.61	1.34	3.83	0.52	0.55
2025-Jul	17.16	8.65	2.15	0.81	1.35	4.46	0.64	0.71
2025-Aug	23.71	9.9	2.26	0.97	1.44	4.69	0.79	0.74
2025-Sep	22.41	12.17	2.69	1.34	1.94	4.97	0.91	0.84
2025-Oct	24.47	17.21	2.87	1.21	2.08	5.34	1.29	0.91
2025-Nov	24.91	19.97	3.3	1.34	2.46	5.07	1.4	1.05
2025-Dec	26.68	20.5	2.54	1.39	2.82	5.37	1.75	1.07
2026-Jan	35.07	26.05	2.44	1.71	3.48	6.21	2.45	1.81
2026-Feb	48.3	31.88	2.32	1.96	3.11	6.89	3.55	2.54
2026-Mar	80.38	44.43	2.51	4.15	3.28	9.7	5.84	4.16
2026-Apr	116.41	49.23	2.49	4.39	4.51	9.19	7.17	4.61
2026-May	143.6	55.98	4.85	4.54	6.5	9.32	14.06	11.27
2026-Jun	163.7	63.9	4.94	3.4	7.5	10.18	16.98	8.21

Chart 2.5

LLMOps Top 5

LiteLLM's pip installs overtook LangSmith, reaching 191M/month by June 2026.

PyPI downloads via pip installer (uv/CI traffic excluded), Jun 2025 – Jun 2026

LangSmith is bundled with LangChain installs.

View chart data

LLMOps Top 5: monthly pip downloads in millions
Month	LangSmith	LiteLLM	MLflow	Langfuse	Braintrust
2025-Jun	99.92	5.95	17.67	1.27	0.08
2025-Jul	102.47	7.83	18.88	1.56	0.14
2025-Aug	99.15	9.14	18.83	1.63	0.2
2025-Sep	94.21	15.52	19.92	1.71	0.75
2025-Oct	97.88	18.11	20.55	2.09	1.13
2025-Nov	94.3	21.21	21.06	2.19	0.59
2025-Dec	94.3	25.11	21.16	2.45	0.81
2026-Jan	91.65	20.6	21.88	2.65	0.59
2026-Feb	35.64	38.26	21.25	3.22	0.79
2026-Mar	44.91	53.9	24.48	4.95	0.71
2026-Apr	44.87	72.1	24.79	4.6	0.72
2026-May	49.71	128.44	26.14	5.76	0.86
2026-Jun	53.45	191.16	27.59	6.18	0.94

Section 03

Model Growth Trends

How AI models have scaled and evolved.

Chart 3.1

LLM Provider SDK Downloads

OpenAI reached 150M monthly pip installs by June 2026. Anthropic surged to 66M. Google (legacy SDK) trails at ~10M.

PyPI downloads via pip installer (uv/CI traffic excluded), Jan 2022 – Jun 2026

View chart data

LLM Provider SDK Downloads: monthly pip downloads in millions
Month	OpenAI	Anthropic	Google GenAI
2022-01	0.09	0	0
2022-02	0.1	0	0
2022-03	0.14	0	0
2022-04	0.17	0	0
2022-05	0.15	0	0
2022-06	0.24	0	0
2022-07	0.15	0	0
2022-08	0.16	0	0
2022-09	0.16	0	0
2022-10	0.16	0	0
2022-11	0.18	0	0
2022-12	0.39	0	0
2023-01	0.68	0	0
2023-02	1.18	0	0
2023-03	2.15	0	0
2023-04	3.29	0.01	0
2023-05	5.47	0.03	0.01
2023-06	7.01	0.07	0.01
2023-07	6.51	0.16	0.03
2023-08	8.07	0.24	0.04
2023-09	9.26	0.21	0.05
2023-10	10.28	0.25	0.07
2023-11	11.24	0.29	0.07
2023-12	12.16	0.28	0.15
2024-01	17.39	0.32	0.21
2024-02	16.76	0.42	0.32
2024-03	18.51	0.69	0.44
2024-04	20.36	1.06	0.53
2024-05	21.21	1.5	0.68
2024-06	21.66	1.68	0.94
2024-07	23.65	1.87	1.12
2024-08	26.48	2.02	1
2024-09	32.93	2.24	1.17
2024-10	36.75	2.58	1.49
2024-11	36.35	2.75	1.83
2024-12	33.38	2.9	1.62
2025-01	37.17	3.48	2.01
2025-02	37.52	5.08	2.46
2025-03	45.21	6.25	2.93
2025-04	45.93	6.29	3.13
2025-05	48.38	6.36	4.05
2025-06	47.61	7.3	4.76
2025-07	55.29	8.75	5.61
2025-08	61.91	14.32	5.94
2025-09	68.73	12.59	7.04
2025-10	78.67	13.6	7.32
2025-11	76.33	13.03	7.5
2025-12	70.61	13.38	7.88
2026-01	76.48	17.84	8
2026-02	85.74	24.24	8.42
2026-03	113.5	39.63	10.37
2026-04	124.95	52.6	10.82
2026-05	141.46	58.23	9.75
2026-06	150.44	66.15	10.05

Chart 3.2

The Gap is Closing

OpenAI-to-Anthropic ratio kept narrowing, from 3.7:1 (Dec 2025) to 2.3:1 (June 2026).

Peak: 47:1 (Jan 2024)

Now: 2.3:1

PyPI monthly downloads ratio, Jul 2023 – Jun 2026

View chart data

OpenAI-to-Anthropic download ratio
Month	OpenAI downloads per Anthropic download
2023-07	37.4
2023-08	29.3
2023-09	38.7
2023-10	36.5
2023-11	34.5
2023-12	37.9
2024-01	47
2024-02	34.1
2024-03	23.7
2024-04	17.1
2024-05	13.7
2024-06	12.6
2024-07	11.9
2024-08	11.9
2024-09	13.4
2024-10	11.5
2024-11	10.6
2024-12	9.7
2025-01	7.4
2025-02	5.6
2025-03	5.5
2025-04	6
2025-05	6.1
2025-06	5.4
2025-07	4.8
2025-08	3.6
2025-09	4.1
2025-10	4.2
2025-11	4.2
2025-12	3.7
2026-01	3.5
2026-02	3.2
2026-03	2.8
2026-04	2.5
2026-05	2.4
2026-06	2.3

Section 04

Research & Content

Surfacing recent research that shaped how 2026 tools handle compression, context, multimodality, and long-horizon agents, so teams can interpret and apply it to their own systems.

Foundational Model Advances

TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate

TurboQuant is an online, data-oblivious quantization method for KV-cache compression and vector search that targets both mean-squared error and inner-product distortion.

It randomly rotates vectors so coordinates are easier to quantize, then applies scalar quantizers coordinate-wise instead of relying on slower vector quantization schemes.

To correct the inner-product bias of MSE-optimal quantization, it adds a 1-bit Quantized Johnson-Lindenstrauss step on the residual; the paper reports quality-neutral KV-cache compression at 3.5 bits per channel and stronger recall than product-quantization baselines in vector search.

Long-context inference and retrieval depend on compression schemes that preserve the geometry attention and nearest-neighbor search actually use, not just on storing vectors in fewer bits.

Recursive Language Models

RLMs treat the prompt as external state, letting the model inspect, partition, and recursively call itself over snippets instead of feeding everything through one context window.

The model works inside a Python REPL where the prompt is stored as a variable, letting it inspect slices of the input, execute code, and recursively call itself over selected snippets.

Across CodeQA, BrowseComp-Plus, and OOLONG, RLMs can handle inputs up to two orders of magnitude beyond base context windows and outperform direct model calls and retrieval-style scaffolds.

Long-context scaling may come less from ever-larger windows and more from giving models programmable ways to search, decompose, and recurse over external state.

Titans: Learning to Memorize at Test Time

Titans is a family of long-context architectures that pairs limited-window attention with a neural long-term memory that keeps learning at test time.

Its memory module writes surprising inputs into its weights using gradient-based updates, then uses momentum and weight decay as a learned forgetting mechanism.

The paper instantiates the design as Memory as Context (MAC), Memory as Gating (MAG), and Memory as Layer (MAL); across language, reasoning, needle-in-a-haystack, time series, and genomics tasks, the authors report stronger results than recent recurrent and hybrid baselines, with MAC doing best on longer dependencies.

Overall, long-context modeling may benefit more from explicit write/forget/retrieve memory systems than from relying on ever-larger attention windows alone.

Kimi K2.5: Visual Agentic Intelligence

Kimi K2.5 is an open-source multimodal agentic model that jointly optimizes text and vision, with a parallel-agent execution framework layered on top.

It mixes vision tokens early during pretraining, uses a native-resolution MoonViT-3D encoder for images and video, and follows with zero-vision SFT plus joint multimodal RL.

Agent Swarm uses a trainable orchestrator to spawn frozen specialist subagents and run heterogeneous subtasks concurrently; the paper reports 76.8% on SWE-Bench Verified and large gains on BrowseComp with 3–4.5× lower execution time at target quality.

Agentic performance may come less from a single monolithic model doing everything in sequence and more from jointly trained multimodal foundations paired with learned orchestration over parallel specialist work.

VeriCache: Turning Lossy KV Cache into Lossless LLM Inference

VeriCache is an inference framework that keeps the throughput of compressed KV caches while producing the exact outputs of full-KV decoding.

It uses the compressed KV cache to draft tokens and then verifies them against the full KV cache, which is held outside GPU memory and swapped in on demand so that verification overhead stays low.

The paper reports up to 4x higher throughput than full-KV inference while producing identical outputs, targeting the divergence that makes lossy compression risky for code generation and tool calling as more tokens are decoded.

For long-context coding agents, the useful frontier for KV compression is not just a smaller memory footprint but behavior that stays correct across long generations, tool calls, and code paths.

M2A: Synergizing Mathematical and Agentic Reasoning in Large Language Models

M2A is a model-merging method for adding mathematical reasoning to a coding agent without disrupting its agentic behavior.

The paper argues math reasoning (closed-world, single-response logic) and agentic reasoning (multi-turn interaction with external environments) are misaligned, so joint multi-task training yields unstable behavior and limited gains; M2A instead works in parameter space, identifying the feature subspace critical for agent behavior and merging the mathematical reasoning task vector only along its null space so reasoning is injected in directions that do not perturb agent behavior.

The merge needs no additional gradient updates and exposes the merging coefficient as a knob for reasoning length; applied to a fine-tuned Qwen3-8B coding agent, it raises SWE-Bench Verified resolved rate from 44.0% to 51.2% without retraining the model.

As coding agents grow more tool-driven, the question shifts from whether a model can reason harder to whether new reasoning skills can be added without breaking the interaction patterns that make agents useful.

Application-Layer Innovations

Chroma Context-1: Training a Self-Editing Search Agent

Context-1 is a 20B agentic search model derived from gpt-oss-20B that is designed to act as a retrieval subagent rather than answer the question directly.

It decomposes a task into subqueries, searches over multiple turns, and selectively prunes its own context to remove irrelevant documents.

Trained on more than 8,000 synthetic tasks with a curriculum that shifts from recall-heavy exploration toward higher-precision retention, it sits on the report's cost-latency frontier and improves prune accuracy from 0.824 to 0.941.

Multi-hop retrieval may work best as a specialized agent problem, where small purpose-trained models learn to plan, explore, and actively edit context instead of relying on single-pass RAG pipelines.

Composer 2 Technical Report

Composer 2 is a domain-specialized model for agentic software engineering, built to handle long-horizon coding tasks while staying efficient enough for interactive use.

Its training recipe combines continued pretraining on a code-heavy mix and a 256k context extension with targeted coding SFT, then follows with large-scale reinforcement learning in Cursor's production tool-use harness.

The report introduces CursorBench, an internal benchmark drawn from real engineering sessions rather than curated public repos; Composer 2 scores 61.3 on CursorBench-3 and is positioned on a strong accuracy-cost frontier relative to larger general-purpose models.

Coding agents improve most when the model, training environment, reward design, and evaluation all match the actual software-engineering workflow.

GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning

GEPA (Genetic-Pareto) is a reflective prompt-evolution method that optimizes instructions using execution traces instead of updating model weights.

The system samples rollouts, has the model analyze its own traces in natural language, and proposes new prompts; a Pareto front keeps multiple candidates that perform well on different subsets of data.

Across four tasks, GEPA matches or beats GRPO-style RL with up to 35× fewer rollouts.

The work treats prompts as an external optimization layer, using natural-language reflection rather than heavyweight RL.

Dockerless: Environment-Free Program Verifier for Coding Agents

Dockerless is an agentic patch verifier that evaluates coding-agent outputs without running repository-specific test environments.

Instead of executing unit tests inside per-repository Docker images, it judges patch correctness from evidence gathered through agentic repository exploration, rather than matching candidate patches against references.

On a verifier evaluation benchmark it outperforms the strongest open-source verifier by 14.3 AUC points, and when used as both the SFT trajectory filter and RL reward in a fully environment-free post-training pipeline the resulting model reaches 62.0% on SWE-bench Verified, 50.0% on Multilingual, and 35.2% on Pro.

As coding-agent training becomes bottlenecked by verification cost, agents that can reliably judge patches without rebuilding every target environment make post-training loops faster, cheaper, and easier to scale.

MEM1: Learning to Synergize Memory and Reasoning for Efficient Long-Horizon Agents

MEM1 is an RL framework that trains LLM agents to operate over long multi-turn tasks while keeping memory usage nearly constant.

At each step, previous memory and new observations are merged into a compact internal state token (IS) and older context is discarded; a masked-trajectory RL scheme reconstructs valid trajectories for PPO without feeding the entire history.

MEM1-7B matches or beats much larger baselines on tasks with up to 16 sequential objectives while reducing memory use by ~3.7×.

Long-horizon behavior can come from learned internal state handling rather than expanding context windows or bolting on external memory.

SWE-Explore: Benchmarking How Coding Agents Explore Repositories

SWE-Explore is a benchmark that isolates the repository-exploration step of coding agents by scoring how well they find the relevant code before any fix is attempted.

Given a repository and an issue, it asks an explorer to return a ranked list of relevant code regions under a fixed line budget, and scores that output along coverage, ranking, and context-efficiency dimensions.

The benchmark spans 848 issues across 10 programming languages and 203 open-source repositories, with line-level ground truth derived from independent agent trajectories that solved the same issue; agentic explorers form a clear tier above classical retrieval, and while file-level localization is already strong, line-level coverage and efficient ranking remain the axes that differentiate state-of-the-art explorers.

Repository-scale coding agents are limited not only by generation quality but by how well they spend a fixed context budget on the specific files and lines a correct patch depends on.

Methodology

Methodology and Sources

How each chart was measured, and what the data can and cannot show.

Figures are as of June 2026 unless a chart specifies another month. The Q2 edition was published in July 2026. June engineering-velocity measurements were provisional when collected at the end of June.

Section 01

Engineering team velocity

Pull request size uses Greptile’s internal merge-request dataset, a fixed March 2025 repository cohort, merged and closed pull requests, and nonzero additions plus deletions. Measurements are grouped by the pull request’s source creation date. Developer output is lines added plus deleted per distinct non-bot author, with per-pull-request values clamped from 1 to 16,000. Medium teams have 6–50 distinct authors per month. Lines changed per file uses the fixed cohort, nonzero diffs, and at least one changed file, across all pull request states.

Lines changed measure code activity, not software quality, business value, or a causal effect of AI. Greptile-connected teams are the observed population, not a representative sample of all developers.

Sections 02 and 03

Package downloads and rules-file adoption

Download charts use public PyPI and npm registry data. PyPI pip-installer counts are used where noted to exclude uv-installer traffic; this does not remove every automated or CI download. npm counts do not provide the same installer breakdown. Downloads are not counts of unique developers or organizations.

Memory-package shares cover five selected packages in May 2026; Zep includes graphiti-core. Vector-database shares cover seven selected open-source packages in June 2026 and exclude Pinecone. These are shares within the measured package set, not total-market adoption.

Rules-file adoption measures resolved CLAUDE.md, AGENTS.md, and Cursor rules files among Greptile-connected organizations that use at least one AI rules file, as of June 2026. Each percentage is independent: one organization can use several formats. March percentages provide the comparison baseline.

The provider chart measures the legacy google-generativeai package alongside OpenAI and Anthropic Python SDKs. The provider-ratio chart retains the edition’s historical PyPI download-ratio series; its counting basis differs from the pip-only provider chart.

Sources

PyPI and PyPI Stats: Python package downloads.
npm: JavaScript package downloads.
Engineering velocity and rules-file detection: Greptile internal datasets.
Research selection: the 12 original papers linked in Section 04. Their findings are attributed to the authors.

Read the tables below each chart for the exact values used to draw it.

Download the Q2 2026 report PDF

Automatically review PRs with your team's standards

View pricing Start now