# Generative AI in R&D in 2026: where it measurably speeds up research and where it does not

In 2026 generative AI moved from demos into the R&D pipeline: an AI-discovered drug entered Phase III, chip tools gained agents and an outside evaluator found AI speeding up a frontier lab. The evidence is real but uneven. What works, what does not, and how to measure it.

Research and development is the part of a business that turns money into knowledge, and then knowledge into products. For most of its history its speed was set by people: how many ideas they could think of, how many experiments they could run, how fast they could read the results. In 2026 generative AI started to change that speed in some places, with measured evidence rather than demos.

The clearest signals came from opposite ends of the field. Inside AI labs, an outside evaluator, METR, concluded in November that AI is probably now roughly doubling the pace of one frontier lab's capability progress. In medicine, a drug whose target and molecule were found with generative AI entered a Phase III trial in September, the last stage before a marketing application. In chip design, the two biggest tool vendors shipped AI agents that write and check hardware code. At the same time, careful measurement kept showing that people overestimate how much AI speeds them up.

This article explains where generative AI measurably speeds up R&D, where it does not yet, how to measure gains without fooling yourself, and how an R&D leader should start. For background on the models themselves, see the [generative AI and LLMs topic page](/topics/generative-ai-llms).

## What happened

### METR finds AI roughly doubling Anthropic's research pace (5 November 2026)

METR, a non-profit that evaluates frontier AI systems, published an assessment based on work inside Anthropic from 12 August to 29 October 2026. It estimates Anthropic is improving model capabilities on well-specified tasks at "around twice the rate" of March 2024 to October 2025, with high uncertainty: the true figure could be "closer to 1.3x or 3x". It says "a majority of this acceleration today is likely attributable to Anthropic's extensive use of AI to do R&D", mostly through AI assistance rather than fully autonomous work, and that it is confident the change was not caused by faster growth in training compute. Our detailed analysis, including Anthropic's IPO filing on the same threshold, is in [METR finds AI roughly doubling Anthropic's research pace](/articles/ai-speeding-up-ai-research).

### Synopsys launches AgentEngineer for chip design (reported 28 September 2026)

Synopsys announced AgentEngineer, a set of "domain-specific long-horizon agents" for verification, implementation, analogue design, manufacturing and simulation, built on a new Autopilot platform. Tom's Hardware reported the company's claims of up to 50x faster verification closure (from earlier work with Nvidia, measured against Synopsys's own workflows) and a 30% productivity gain against human experts. General availability is planned for the end of 2026; no customer in production has been named.

### Periodic Labs releases Neon, a science model trained on lab data (17 September 2026)

Periodic Labs, a materials start-up founded by former OpenAI and Google researchers, released Neon, a model of about one trillion parameters post-trained on data from its own laboratories. Reported benchmark results put it ahead of GPT-6 Astra and Claude Fable 5.1 on interpreting X-ray diffraction data, the step that tells a chemist what they actually made. These are vendor-reported results on a benchmark Periodic built.

### Insilico doses the first patient in a Phase III trial of rentosertib (10 September 2026)

Insilico Medicine said the first patients were dosed in GENESIS-IPF-3, a Phase III trial of rentosertib for idiopathic pulmonary fibrosis, a scarring lung disease. The company says both the target, TNIK, and the molecule came from its generative AI platform, and that TNIK had never been linked to fibrosis before. The trial plans about 320 patients in 47 centres in China over 52 weeks. Insilico says approval would take three to four years "under favorable conditions".

### OpenAI says it has built an "automated research intern" (7 September 2026)

OpenAI said it had met a goal set the previous autumn: a system that carries out well-defined research tasks under human direction, including work that would take a skilled researcher several days. Help Net Security reported OpenAI's figures: its research organisation logs 3.1 agent-workdays for every eight hours of human labour, and more than half of successful four-to-eight-hour tasks still needed at least one human intervention. OpenAI's next stated goal is an automated AI researcher by March 2028.

### Google makes AlphaEvolve generally available (July 2026)

Google opened AlphaEvolve, its code-evolving agent, to cloud customers. It uses Gemini models to propose changes to a program, scores each version with an evaluation function the customer writes, and keeps the best. InfoQ reported customer claims including Klarna doubling machine learning training throughput after exploring about 6,000 candidate programs, and FM Logistic cutting warehouse picking routes by 10.4%. All figures come from Google or its customers.

### An AI-designed vaccine passes its first human safety trial (5 June 2026)

The University of Cambridge and its spin-out DIOSynVax reported that pEVAC-PS, a vaccine whose antigen was designed with machine learning to cover a whole family of coronaviruses, was safe in a Phase I study of 39 volunteers. It produced immune responses against SARS-CoV-2, SARS and related bat viruses.

## Where generative AI actually speeds up R&D

R&D is a loop: decide what to try, build it, test it, learn, repeat. Generative AI helps most where the "test" step is fast, cheap and automatic. A good analogy is a spell-checker. It is extremely useful because every suggestion can be checked instantly against a dictionary. A tool that suggests a better plot for your novel is harder to trust, because there is no dictionary for good plots.

### AI research inside AI labs

AI labs are the extreme case, because much of their R&D is software and much of it can be scored automatically. METR's assessment suggests AI is already speeding up capability progress at one lab, and OpenAI describes agents logging more work hours than its staff. Both stress that today's gain comes mostly from assistance: humans still choose priorities, judge results and decide what ships.

The labs treat this as a safety matter too. If better AI makes research faster and faster research makes better AI, the loop can compound. Each major lab writes a threshold for it in advance, like a smoke alarm installed before anything burns. Google DeepMind updated its version in April 2026:

| Lab framework | What it tracks | Threshold idea |
|---|---|---|
| Anthropic Responsible Scaling Policy | AI-driven acceleration of AI progress | Compressing two years of 2018 to 2024 progress into one year |
| OpenAI Preparedness Framework | "AI Self-improvement" | High: a capable research engineer assistant for every researcher; Critical: recursive self-improvement |
| Google DeepMind Frontier Safety Framework 3.1 | Machine learning R&D | Acceleration: AI progress substantially faster than historical rates; Automation: fully automating an AI research team |

The detail of each threshold, and how METR measured Anthropic against its own, is covered in our [METR article](/articles/ai-speeding-up-ai-research).

### Drug discovery and protein design

Here generative models propose things: new disease targets, small molecules, antibody sequences, vaccine antigens. The 2026 milestones are real but early. Rentosertib reached Phase III after a 12-week Phase IIa study published in Nature Medicine in 2025, in which patients on the highest dose gained an average of 98.4 mL of lung capacity while those on placebo lost 20.3 mL, as compiled by IntuitionLabs. Generate:Biomedicines announced in December 2025 two Phase III trials, SOLAIRIA-1 and SOLAIRIA-2, of GB-0895, an AI-engineered asthma antibody, with about 1,600 patients. Cambridge's vaccine shows the same approach for antigens.

Where it does not yet speed things up is the clinic. An ASCO abstract counted 117 AI-enabled drug assets from 63 companies in interventional trials as of December 2025; 60 had completed Phase I and only 8 had completed Phase II. IntuitionLabs reported that no AI-discovered drug had received full FDA approval as of mid-2026. AI can shorten the design phase, but a 52-week trial still takes 52 weeks. Patients, not models, set the clock.

### Materials and chemistry

Materials R&D is a search through a huge space of possible compounds, followed by slow lab synthesis and testing. Generative models can propose candidates quickly; the bottleneck is making them and proving what was made. Periodic Labs' Neon targets that second step, interpreting X-ray diffraction results. Notice the data: Periodic trains on records from its own high-throughput labs, so the advantage sits with whoever owns the experimental loop.

### Chip design and engineering software

Chip design already runs on software tools and huge automated test suites, so it suits agents well. Cadence put its ChipStack AI Super Agent into early access in February 2026 for writing designs and testbenches, planning tests and debugging, and Synopsys followed with AgentEngineer. The vendor claims are large, but they are vendor claims, and the jobs where agents shine are verification and other steps with clear pass or fail signals. The same logic explains AlphaEvolve's results: it works only where someone can write a scoring function, and InfoQ notes it will exploit anything the scorer does not measure.

## How to measure R&D productivity gains honestly

A faster dishwasher does not make dinner arrive sooner if the oven is the slow part. R&D is the same. Speeding up one task matters only in proportion to how much of the total cycle that task takes, and only if the next step can absorb the extra output. That is why "AI made me 10x faster at coding" and "our programme moved 1.2x faster" can both be true.

### Lessons from METR's developer studies

Almost every R&D team now writes code with help from agents such as [Claude Code](/tools/claude-code), and this is where the best controlled evidence exists. In July 2025 METR published a randomised controlled trial with 16 experienced open-source developers working on 246 real issues from their own repositories. Each issue was randomly assigned to allow or forbid AI tools, mainly Cursor Pro with Claude 3.5 and 3.7 Sonnet. With AI allowed, tasks took 19% longer. Before the study the developers expected a 24% speed-up, and afterwards they still believed AI had sped them up by 20%. METR later gave the confidence interval for the slowdown as +2% to +39%.

The context matters. These were experts on large codebases they knew well, using early-2025 tools. The February 2026 follow-up, with 57 developers and more than 800 tasks, found raw estimates of 18% speed-up for returning developers and 4% for new ones, with confidence intervals that include zero. METR thinks developers probably are faster with AI in 2026 than in 2025, but warned that 30% to 50% of developers said they skipped some tasks they did not want to do without AI, which makes the comparison unreliable. The lesson for an R&D leader is not "AI slows people down". It is that **felt speed-up is a poor measure, and even good experiments get harder once people depend on the tool.**

### Four metrics that survive scrutiny

- **Cycle time.** The elapsed time from question to answer, or from idea to validated result. Measure the whole loop, not the step AI touched.
- **Hit rate.** The share of designs, candidates or experiments that pass the next real gate: a wet-lab assay, a silicon test, a review.
- **Cost per validated result.** Include compute and model spend, lab time and the expert hours spent checking AI output. OpenAI's report of median researchers using more than $600 a day of inference at API prices shows this cost is not trivial.
- **Quality and rework.** Defects found later, retracted results, failed replications. Speed bought with rework is borrowed, not earned.

### Use controlled comparisons where you can

The gold standard is the one METR used: randomly assign comparable work to "AI allowed" and "AI not allowed", then compare. Where that is impossible, use a frozen baseline period, a matched comparison team, or a staggered rollout where some groups start later. Write down beforehand what you will measure and when, and record what else changed (headcount, budget, scope), because attributing a gain is a [causal inference](/tools/causal-inference) problem. Basic [statistics](/tools/statistics), such as confidence intervals, stop you celebrating a lucky month. And publish the method with the number: as [a score is only as good as its report](/articles/a-score-is-only-as-good-as-its-report) argues, a figure without its method cannot be trusted.

## Why some R&D speeds up and some does not

Put the evidence together and a simple rule appears. Generative AI accelerates R&D in proportion to three things:

1. **How cheaply its output can be checked.** Code with tests, chip designs with simulators and kernels with benchmarks are checked in seconds. Drug efficacy is checked in years.
2. **How much of the cycle the AI-assisted step occupies.** Faster molecule design saves months in a decade-long drug programme. Faster verification matters more in chips, because verification is often the largest share of a chip project's engineering effort.
3. **Who owns the feedback data.** Labs that generate their own experimental data, like Periodic or Insilico, can retrain on results. Teams that only prompt a general model cannot.

Automated evaluation is the real enabler. AlphaEvolve, chip agents and lab assistants all depend on [evals](/tools/evals) that score work without a human. Where a machine judge is used, such as an [LLM-as-judge](/tools/llm-as-judge), it must itself be checked against expert labels before anyone trusts its scores.

## What it means in practice

**Design is getting cheap; validation is the new bottleneck.** In drugs, materials and chips, models now generate more candidates than teams can test. Budgets will shift towards assays, simulators, test benches and reviewers.

**Proprietary experimental data is a moat.** Anyone can use a general model; only you have your experimental results, including failures.

**Vendor numbers need translating.** "Up to 50x" describes one step. Ask what end-to-end cycle time and hit rate did.

**Safety and governance follow the acceleration.** The labs whose R&D speeds up most are writing thresholds for it. Industrial R&D needs a smaller version: rules for what AI may do without review.

**The human role moves up a level.** In every 2026 example, people set priorities and judge results.

## What to do now

1. **Map your R&D loop first.** Write down each step from idea to validated result with its typical time and cost. Target the slowest step that has a cheap, automatic check.
2. **Start with one well-instrumented pilot.** Pick a contained area such as test generation, literature triage or candidate screening, with a clear success metric and an owner.
3. **Freeze a baseline before you start.** Record cycle time, hit rate and cost per validated result for a pre-AI period, and do not move it after seeing results. See [snapshots that keep numbers meaningful](/articles/snapshots-keep-numbers-meaningful).
4. **Run a controlled comparison.** Randomise comparable tasks or stagger the rollout across teams for a few weeks. Treat self-reported speed-ups as a supplement only.
5. **Invest in evaluation as much as generation.** Build the test suites, simulators and expert-labelled sets that let you trust AI output. Without them, more output means more risk.
6. **Protect IP and confidential data.** Use enterprise contracts that exclude training on your data, keep sensitive work on approved or self-hosted models, and block unapproved tools; the risks are described in [LLMjacking and shadow AI](/articles/llmjacking-and-shadow-ai-data-leaks). Record which ideas and designs were AI-assisted and who contributed what, because patent law in major jurisdictions, including the US and the UK, requires a human inventor.
7. **Capture your own data loop.** Log every experiment, including failures, in a structured form a model can learn from later.
8. **Define review tiers.** Decide which AI outputs can be used directly, which need one reviewer and which need formal sign-off, and keep a [human in the loop](/tools/human-in-the-loop) for anything that reaches patients, customers or silicon.
9. **Put it under governance.** Name an accountable owner, track AI spend per project, and review results quarterly as part of your [AI governance](/tools/ai-governance) process.

## Risks and open questions

**Most headline numbers are self-reported.** Synopsys, Periodic, Google's AlphaEvolve customers and OpenAI all report their own figures. METR's Anthropic assessment is independent, but much of its key evidence was redacted.

**Clinical proof is still missing.** No drug discovered with generative AI has completed Phase III. A failure in GENESIS-IPF-3 or the GB-0895 trials would not disprove AI design, but it would remind everyone that design speed is not success.

**Measurement is getting harder.** METR's 2026 update shows that once people rely on AI, they will not work without it, which undermines controlled trials. Organisations may lose the ability to say what AI is actually worth.

**Safety incidents are real.** Help Net Security reported that OpenAI shut down a training container service in July 2026 after agents compromised its research infrastructure. Agents with access to R&D systems need the same controls as any powerful user.

**Watch next:** whether Anthropic designates a model as past its AI R&D threshold, the first AI-designed drug Phase III readouts, whether Synopsys names production customers after general availability, Isomorphic Labs' first clinical trial, and new controlled studies of AI in R&D teams outside software.

## Frequently asked questions

### What is generative AI in R&D?

It is the use of models that produce new content, such as text, code, molecules, protein sequences or designs, inside the research and development loop. Typical uses are drafting code and experiments, proposing candidate compounds or designs, analysing results and summarising literature.

### Does generative AI really speed up R&D?

In some places, yes, with evidence. METR estimates AI roughly doubled one frontier lab's capability progress, and chip and code tools show large gains on steps with automatic checks. Where validation is slow, as in clinical trials, the overall cycle has barely changed.

### What did METR's 2025 developer study find?

In a randomised trial, 16 experienced open-source developers took 19% longer on real tasks when AI tools were allowed, while believing they had been about 20% faster. A 2026 follow-up suggested tools have improved but found the experiment design itself was breaking down.

### Has any AI-designed drug been approved?

Not as of mid-2026, according to industry trackers. Insilico's rentosertib entered Phase III in September 2026 and Generate:Biomedicines' GB-0895 is in Phase III, so approvals, if they come, are years away.

### How should an R&D leader start with generative AI?

Map the R&D loop, choose one step with a cheap automatic check, freeze a baseline, and run a controlled pilot. Measure cycle time, hit rate and cost per validated result, and set rules for data, IP and human review before scaling.

## Summary

Generative AI became a measurable force in R&D in 2026, but unevenly. METR found AI roughly doubling Anthropic's pace of capability progress, mostly through assistance. Insilico's rentosertib reached Phase III, chip-tool vendors shipped agents, and Periodic Labs trained a science model on its own lab data. The common thread is that AI speeds up the steps that can be checked cheaply and automatically, while slow validation, such as clinical trials and lab synthesis, sets the overall pace. R&D leaders should map their loop, pilot where checks are cheap, measure cycle time, hit rate and cost per validated result against a frozen baseline, and invest as much in evaluation, data and governance as in generation.

## Sources

- [METR via AI Alignment Forum: Independent assessment of AI R&D acceleration at Anthropic](https://alignmentforum.org/posts/N5AmYkRks5CqAxvBg/metr-independent-assessment-of-ai-r-and-d-acceleration-at)
- [METR: Measuring the impact of early-2025 AI on experienced open-source developer productivity](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/)
- [METR: Uplift study update, February 2026](https://metr.org/blog/2026-02-24-uplift-update/)
- [Help Net Security: OpenAI just hit a milestone on the road to self-improving AI](https://www.helpnetsecurity.com/2026/09/07/openai-research-automation-intern/)
- [PPC Land: Google raises security to level 2 for 3 types of dangerous AI capability](https://ppc.land/google-raises-security-to-level-2-for-3-types-of-dangerous-ai-capability/)
- [News-Medical: Generative AI-driven drug rentosertib enters Phase III trial for idiopathic pulmonary fibrosis](https://www.news-medical.net/news/20260910/Generative-AI-driven-drug-Rentosertib-enters-Phase-III-trial-for-idiopathic-pulmonary-fibrosis.aspx)
- [Eigent: Periodic Neon, a lab-trained AI scientist](https://www.eigent.ai/blog/periodic-neon-lab-trained-ai-scientist)
- [ScienceDaily: AI-designed vaccine clears first human safety trial](https://sciencedaily.com/releases/2026/06/260605023357.htm)
- [Tom's Hardware: Synopsys debuts Autopilot platform for developing chips autonomously using AI](https://tomshardware.com/tech-industry/semiconductors/synopsys-debuts-autopilot-platform-for-developing-chips-autonomously-using-ai-new-agentengineer-platform-is-poised-for-general-availability-by-the-end-of-2026)
- [Digital Engineering 24/7: Cadence unleashes ChipStack AI Super Agent](https://www.digitalengineering247.com/article/cadence-unleashes-chipstack-ai-super-agent)
- [InfoQ: AlphaEvolve generally available](https://infoq.com/news/2026/07/alphaevolve-generally-available)
- [IntuitionLabs: AI-discovered drugs in clinical trials, 2026](https://intuitionlabs.ai/articles/ai-discovered-drugs-clinical-trials-2026)
