Fastest
Eureka on the web
When you want the answer in the next five minutes.
The agent works the prompt against patents and technical literature, citing every source.
Run your analysis now →https://eureka.patsnap.com/resources/blog/rd-blog/code-generation-with-language-models-patent-landscape/ · Patsnap · data cut-off 2026-07-31 · downloaded from the live page
Patent Landscape · Updated 2026Code generation with language models: patents, leading filers, and where claim space is still open
- 46.2% of all 719 records sit with the five leading assignees, so a newcomer's freedom-to-operate check has to start there, not with the long tail.
- Filings nearly doubled, +95% from 2021 (74) to 2024 (144) — the last year the dataset can call complete given an 18-month publication lag.
- G06N coverage sits at 40.1% against G06F's 88.6%, showing most claims are still anchored in general data processing rather than AI-model-specific architecture.
719
Published Records
46%
Top-5 Share of All Records
+95%
Filing Growth 2021→2024
US
Leading Jurisdiction
Filing growth compares 2021 (74 records) with 2024 (144) — a three-year span. 2024 is the most recent year we treat as complete: publication lags filing by roughly 18 months, so 2025 onwards are still filling in and any growth rate that ends there would understate the field. Top-5 share is the combined record count of the five largest assignees divided by all 719 records in scope (CR5), not by the ranked leaders only.
Published byPatsnap Research··7 min readSourced from Patsnap Eureka
Overview
What this landscape covers
This dataset tracks 719 published patent records at the intersection of code generation models, code completion, and program synthesis, filtered to filings that also reference abstract syntax trees, code embeddings, fine-tuning, unit test generation, static analysis feedback, or inline suggestion. It spans filings from 2015 through the 2026-07-31 data cut-off, capturing the period in which large language models moved from research prototypes into developer-tooling products with claimed technical architecture rather than generic software methods.
The scope is deliberately narrow: it excludes general-purpose LLM patents that never touch code, and it excludes classic static program-synthesis work that never references model-based generation. What remains is the set of filings actually contesting how a model produces, completes, or validates source code.
Filing activity and technology composition, 2015-2026
Source: Patsnap Eureka. Assignee ranking and totals. Derived from a Patsnap search on Code Generation with Language Models covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
Let an AI agent run this analysis on your own technology
Pick a task. Every answer cites the patents behind it.
10,000 free credits to start
The Numbers
Filing trend and technology composition
Two views of the same 719 records: how filing volume moved year over year, and which IPC subclasses the claims actually sit in.
Filing trend, 2017-2026
Filings rose from zero recorded in 2017 to a peak of 152 in 2025, with growth of +95% between 2021 (74) and 2024 (144) — the most recent year the trend can be read as complete. 2025 and 2026 figures will continue to rise as later publications catch up with their filing dates.
IPC subclass composition
G06F (electric digital data processing) appears in 88.6% of the 719 records and G06N (AI-based computing) in 40.1%, confirming most filings are framed as data-processing methods with an AI component rather than AI-architecture claims in isolation. G06Q, H04L, G06K, G06V, G10L and H04M each cover a smaller adjacent slice — commerce systems, transmission, recognition, vision, speech, and telephony use cases layered on top of code generation.
Shares are the percentage of the 719 records in scope. A patent can carry several IPC classes, so the shares add up to more than 100%.
Source: Patsnap Eureka. Filing trend and technology composition. Derived from a Patsnap search on Code Generation with Language Models covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.
Go deeper on Code Generation with Language Models with Eureka
This page is one run against one query. Ask Eureka your own question about code generation with language models and every answer comes back with the patent numbers behind it.
Try EurekaKey Patents
Most-cited records and a representative recent filing
Representative Filing
US12277409B1 — Training a code generation model for low-resource languages
SAS INSTITUTE INC.
The filing (SAS Institute, granted 2025-04-15) claims a pipeline that generates a code synthesis prompt from sampled code synthesis items, has a large language model synthesize raw code segments from that prompt, executes those segments with a code interpreter for the target language, keeps only the segments that execute successfully, and aggregates the validated segments via a second sampling step.The execute-then-filter loop is the operative step: it is what turns raw model output into a validated training or completion corpus, and it is the part most likely to matter in a design-around analysis.
Most-cited records in scope
| # | Publication no. | Patent title | Citations |
|---|---|---|---|
| 1 | US20180365201A1 | System and method for compound data-driven contracts and documentation | 228 |
| 2 | US20200097662A1 | Combined threat score for container images | 158 |
| 3 | US20200097389A1 | Error recovery | 144 |
| 4 | US20200082094A1 | Selectively applying heterogeneous vulnerability scans to layers of container images | 122 |
| 5 | US20200249918A1 | Deep learning enhanced code completion system | 93 |
| 6 | US20240330927A1 | Smart contract generation and validation | 87 |
| 7 | US20190122317A1 | System and method for a computing environment for verifiable execution of data-driven contracts | 79 |
| 8 | US20200357084A1 | System and method for blockchain-based network transitioned by a legal contract | 78 |
| 9 | US20210034335A1 | Multi-lingual line-of-code completion system | 77 |
| 10 | US20220317979A1 | Edit automation using a temporal edit pattern | 74 |
Citation counts favour older filings simply because they have had longer to accumulate citations within this corpus — read them as a signal of influence, not of current technical importance.
Each row carries its publication number; clicking a row searches Eureka by that number.
Source: Patsnap Eureka. Citation counts and representative records. Derived from a Patsnap search on Code Generation with Language Models covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
Run it yourself
Put your own technology through the same analysis
Where to run it
For builders
MCP server & REST API
When it has to run inside your own pipeline.
Patent search, landscape analysis and assignee resolution as MCP tools. Drop them into any agent framework, or call REST directly.
Browse MCP servers →Insights
What the numbers mean for a filing decision
Three patterns in the data that change where a new filing or freedom-to-operate search should start.
Concentration
46.2% of 719 records
held by the top 5 assignees
The top of the field is dense, not the whole field
Just over 46% of all records in scope sit with five assignees, and the top 10 combined reach 58.0%. That leaves a long tail of single- or low-filing entrants working in the remaining space — useful territory for a narrower claim, provided it clears the leaders' core architecture first.
Based on the full 719-record assignee ranking.
Momentum
+95%, 2021→2024
filing growth over three years
Growth is real but concentrated in complete years only
Filings grew from 74 in 2021 to 144 in 2024, a rate the dataset can state with confidence because both years are far enough past the roughly 18-month publication lag. 2025's provisional 152 and 2026's partial 32 should not be read as a slowdown or a peak — they are still filling in.
2024 is the last year treated as complete.
Composition
88.6% vs 40.1%
G06F share vs G06N share
Most claims are framed as data processing, not AI architecture
G06F covers 88.6% of the 719 records while G06N, the AI-model-specific subclass, covers 40.1%. The gap suggests many filers are still claiming code generation as a data-processing method with a model inside it, leaving room for claims that target the model architecture or training loop directly.
Shares are of all 719 records; a record can carry multiple classes.
Signals
5 citations, 228 max
top cited record
Citation leaders skew toward infrastructure and error handling, not generation itself
The most-cited records in this set concern contract documentation, container-image threat scoring, and error recovery rather than code generation proper — a reminder that citation volume tracks age and adjacency as much as it tracks relevance to the core technology.
See the most-cited table for the full list.
Eureka AI Agent
Looking for what nobody has claimed yet?
Eureka can read the same corpus for gaps instead of for coverage: under-claimed branches adjacent to code generation with language models, with the prior art for and against each one.
Source: Patsnap Eureka. Co-assignee relationships and derived observations. Derived from a Patsnap search on Code Generation with Language Models covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
Players
Who is filing, and where the room to move sits
The ranked leaders hold the core architecture claims; momentum among smaller filers points to where activity is still shifting.
Leader
209 records
filed by the top assignee
One filer sits well ahead of the field
The leading assignee's 209 records dwarf fifth place at 20 and tenth place at 14, a gap large enough that any FTO review in this space has to clear that single portfolio before looking anywhere else.
Counted in records across the 719-record scope.
Momentum, up
+100% YoY
Forethought Tech Inc, latest year
Smaller filers are moving fastest, not the leaders
Forethought Tech Inc doubled its latest-year filings and Huawei Cloud Computing Technologies rose 50%, while several larger, earlier-established filers (US Bank National Association, U S Bancorp, For Good AI) posted sharp year-on-year declines. Momentum and portfolio size are not the same signal here.
Latest-year YoY change, recent-year momentum data.
Momentum, down
-82% YoY
US Bank National Association, latest year
Some early movers are pulling back
US Bank National Association's filings fell 82% year-on-year and U S Bancorp fell 76%, suggesting the financial-services angle on code generation (likely internal tooling or compliance-adjacent applications) has cooled after an earlier filing push.
Recent-year momentum by assignee.
Recent-year filing momentum by assignee
| Assignee | Recent year | YoY |
|---|---|---|
| US BANK NATIONAL ASSOCIATION | 5 | -82% |
| U S BANCORP NAT ASSOC | 4 | -76% |
| FOR GOOD AI INC | 4 | -64% |
| Huawei Cloud Computing Technologies Co., Ltd. | 3 | +50% |
| FORETHOUGHT TECH INC | 2 | +100% |
| Microsoft Technology Licensing, LLC | 1 | -67% |
| International Business Machines Corporation | 0 | — |
| Oracle International Corporation | 0 | -100% |
Source: Patsnap Eureka. Assignee-level momentum. Derived from a Patsnap search on Code Generation with Language Models covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
What's Next
Where to take this from here
The landscape points to next steps depending on whether the reader is filing, defending, or scouting.
Run a freedom-to-operate check against the leader's portfolio
With one assignee holding 209 of 719 records, any new filing in core generation or completion claims needs to map against that portfolio specifically, not just the top-10 aggregate.
Explore assignee portfolios in Eureka →Draft claims into the thinner IPC branches
The 40.1% G06N share against 88.6% G06F suggests model-architecture-specific claims, rather than generic data-processing framing, still have room.
Map white space in Eureka →Track momentum, not just portfolio size
Assignees with small but growing filing counts are a better early signal of where the field is moving than the largest static portfolios.
Watch assignee momentum in Eureka →Source: Patsnap Eureka. Forward-looking reading of the same dataset. Derived from a Patsnap search on Code Generation with Language Models covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
FAQ
Common questions about this landscape
This dataset identifies 719 published patent records matching code generation models, code completion, and program synthesis language combined with technical terms like abstract syntax tree, fine-tuning, and static analysis feedback. The true worldwide total is somewhat larger because publication lags filing by roughly 18 months, so 2025 and 2026 filings are still arriving in public databases. Treat 719 as the current visible scope rather than a final count.
Filing activity is highly concentrated: the top five assignees together hold 46.2% of all 719 records in scope, and the top ten hold 58.0%. The single leading assignee alone accounts for 209 records, far ahead of fifth place at 20 and tenth place at 14. That said, momentum data shows some smaller filers growing faster year-on-year than the largest incumbents, so portfolio size and current filing velocity tell different stories.
Yes, through the last year the data can treat as complete: filings grew from 74 in 2021 to 144 in 2024, a 95% increase over that three-year span. Figures for 2025 (152) and 2026 (32, partial) look lower or flat only because publication lag means recent filings have not all surfaced yet — they should not be read as a slowdown.
G06F (electric digital data processing) appears in 88.6% of the 719 records and G06N (AI-based computing) in 40.1%, meaning most filings are framed primarily as data-processing methods rather than pure AI-model architecture claims. Smaller but notable slices touch G06Q (business/commerce systems, 9.3%), H04L (digital transmission, 4.0%), and G06K, G06V, G10L, and H04M for recognition, vision, speech, and telephony overlaps respectively. A record can carry multiple classes, so these percentages do not sum to 100%.
The gap between the 88.6% G06F share and the 40.1% G06N share suggests model-architecture-specific and training-loop claims are less crowded than general data-processing framing of code generation. Sub-areas such as unit test generation from partial context, static-analysis feedback loops for model retraining, and low-resource-language adaptation show thinner direct coverage relative to core completion claims. Any filing into these branches still needs a clearance check against the leading assignee's 209-record portfolio first.
US12277409B1, assigned to SAS Institute and granted 2025-04-15, claims a pipeline that generates a code synthesis prompt, has a large language model produce raw code segments from it, executes those segments to filter for successful ones, and aggregates the validated output via a second sampling step. It is most directly relevant to anyone building an execute-then-filter validation loop for model-generated code, particularly for low-resource languages. It does not obviously block code generation approaches that skip execution-based filtering entirely or that validate through static analysis instead of interpreter execution.
Answers are grounded in the same dataset. Derived from a Patsnap search on Code Generation with Language Models covering 2015–2026, data cut-off 2026-07-31. Counts reflect published records only and shift as new filings publish.Run this in Eureka MCP
Research Code Generation with Language Models in depth with Eureka
Go past this page: query the whole code generation with language models corpus yourself, in your own scope.
Every answer comes back with patent numbers you can open.
Disclaimer. This page is generated from Patsnap Eureka data drawn from a limited snapshot of global patent and scientific-literature records, and is provided for general information and reference only.
Patent data carries inherent limitations: recent filings (typically the most recent 18–24 months) are under-counted due to standard publication lag; counts may be reported at either a patent-family or a patent-record basis and are not always directly comparable; classification, applicant-name, and citation data may contain errors, duplicates, or omissions; and the underlying search query defines and constrains the scope shown. As a result, the analysis may be incomplete or inaccurate and may not reflect the full technology landscape.
Nothing on this page constitutes an exhaustive prior-art, novelty, freedom-to-operate, or validity search, nor does it constitute legal, financial, investment, or professional advice, and it should not be relied upon as such. Any patent, commercial, or strategic decision should be verified independently and reviewed with qualified patent, legal, and domain professionals. Patsnap makes no warranties, express or implied, as to the accuracy, completeness, or fitness for any particular purpose of the information presented.
Machine translation. Assignee and organisation names originally recorded in Chinese, Japanese or Korean have been rendered into English by an AI translation step so that the tables stay readable. These renderings are best-effort and may not match a company's registered English name; the original name is what the underlying patent record carries, and it is what any Eureka query launched from this page uses.