FundingRounds, leads & terms
Mirendil Raises $1B Series A and Mercor Raises $750M at $20B Valuation, Topping the Week's Biggest Funding Rounds
Mirendil, an AI lab founded by former Anthropic researchers, raised a $1B Series A less than five months after its seed. Mercor, which sells AI evaluation tools, raised $750M at a $20B valuation. Both rounds led a week of big bets on tools for training and evaluating AI agents.
From the Crunchbase News article of November 13, 2026 and cited sources: Kleiner Perkins and Clearlake Capital led Mirendil's $1B Series A, and Google and Battery also took part. [1] [4] [8] [9]
SOURCES[1] Crunchbase News: The Week's 10 Biggest Funding Rounds news.crunchbase.com[4] FourWeekMBA: Mirendil 5x its seed fourweekmba.com[8] Mirendil on X x.com
AI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Kleiner Perkins and Clearlake Capital led Mirendil's $1B Series A, and Google and Battery also took part. [1] [4] [8] [9]
- Original source
- Crunchbase News article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- November 13, 2026 article published
- Covered
- Nov 14 edition
- Evidence
- 13 sources across 12 sites
- 1Crunchbase News: The Week's 10 Biggest Funding Roundsnews.crunchbase.com
- 2Unite.AI: Mirendil Raises $1Bunite.ai
- 3MLQ: Ex-Anthropic researchers' Mirendil raises $1Bmlq.ai
- 4FourWeekMBA: Mirendil 5x its seedfourweekmba.com
- 5Securities.io: Mirendilsecurities.io
- 6ExplainX: Mirendil-Base claimed vs verifiedexplainx.ai
Reach: not measured yet
Kleiner Perkins and Clearlake Capital led Mirendil's $1B Series A, and Google and Battery also took part. The round came 141 days after a $200M seed that valued the company at $1B. Mirendil did not disclose its new valuation. In September, Bloomberg reported early talks at $5B. [1] [4] [8] [9]
Mercor's $750M Series D at a $20B valuation roughly doubles the $10B it was valued at in October 2025. Nvidia, Salesforce and Zoom joined as new investors. Mercor sells evaluation and post-training infrastructure to enterprises and runs the APEX benchmark series. [7] [11]
Mirendil argues that AI agents can do most of an AI lab's work. It says one researcher directing agents pretrained a coding model in 3.5 months for under $10M, with agents writing over 99% of the code. These figures come from the company and have not been independently replicated. On a bits-per-byte test across seven code repositories Mirendil picked itself, Mirendil-Base scored 0.222. That trails Kimi-K3 at 0.193 but beats GLM-5.3 and DeepSeek-V4-Pro-Base. Bits per byte measures pretraining quality, not performance on tasks like bug fixing. [3] [6]
The deal sizes show investors paying high prices for agent training and evaluation infrastructure, especially from frontier-lab teams. For buyers, the likely implication is that testing models on their own systems becomes a paid product.
Read the full assessment
Much cheaper pretraining remains a hypothesis.
Executive brief
Mirendil, a lab founded by two former Anthropic researchers, raised a $1B Series A 141 days after its $200M seed. That made it the largest round in Crunchbase's Nov. 7–13 list. Its pitch is that AI agents can do most of an AI lab's work. The company says one researcher working with agents pretrained a coding model for under $10M, a claim that has not been independently replicated. Mercor came second with a $750M Series D at a company-announced $20B valuation. Both bets are on the same thing: the infrastructure for training and evaluating AI agents.
What changed and event timeline
Mirendil launches
A $200M seed at a $1B valuation, co-led by Andreessen Horowitz and Kleiner Perkins, with Nvidia joining. The founding team was about 20 researchers from Anthropic, xAI, DeepMind and OpenAI (;).
Mirendil-Base report
Mirendil published a company-run evaluation of its agent-built coding model. Inference provider Baseten said on X that Base-Large scored best on its private codebases ().
Mercor Series D
$750M at a $20B valuation, about double its $10B valuation from October 2025. Nvidia, Salesforce and Zoom joined as new investors ().
Mirendil Series A
$1B led by Kleiner Perkins and Clearlake Capital. Google, Battery, Doerr Capital and others participated (;).
Crunchbase weekly ranking
Mirendil and Mercor top the list. LangChain ($300M at $3B), Torus ($500M) and Mavria also feature ().
Capabilities and access
- It runs hundreds of agents that write code, run experiments for days and provision their own GPUs.
- Mirendil-Base: 100B–500B parameters. Base-Large (about 80% trained) and Base-XL (about 19% trained) are larger siblings.
- Fern: a post-trained model with a context window of up to 10M tokens.
Read the full section
Mirendil (Ultrathink; MarkTechPost)
- Lor: in private beta. It runs hundreds of agents that write code, run experiments for days and provision their own GPUs. It works with Anthropic, OpenAI and open-source models.
- Mirendil-Base: 100B–500B parameters. Base-Large (about 80% trained) and Base-XL (about 19% trained) are larger siblings.
- Fern: a post-trained model with a context window of up to 10M tokens.
- Weights and API access are not yet available.
Mercor (Business Wire via Yahoo)
- Sells evaluation and post-training infrastructure to enterprises.
- Runs the APEX public benchmark series.
- Uses simulated enterprise-app environments from Deeptune, which it acquired in July 2026.
Technical analysis for researchers and developers
- Mirendil's method claim: agents wrote more than 99% of the training code. Training used only public data.
- What was measured: bits per byte (a tokenizer-independent measure of how well a model predicts text) across seven public code repositories. Mirendil-Base scored 0.222.
- Why that matters: bits per byte is a pretraining signal. It says nothing about SWE-style bug fixing or terminal tasks.
Read the full section
- Mirendil's method claim: agents wrote more than 99% of the training code. Training used only public data. The human researcher directed strategy and checked results (MLQ).
- What was measured: bits per byte (a tokenizer-independent measure of how well a model predicts text) across seven public code repositories. Mirendil-Base scored 0.222. Kimi-K3 scored better at 0.193. Mirendil-Base beat GLM-5.3 (0.242), DeepSeek-V4-Pro-Base (0.243) and Nemotron-550B-Base (0.247) (ExplainX).
- Why that matters: bits per byte is a pretraining signal. It says nothing about SWE-style bug fixing or terminal tasks.
- Reproducibility: Mirendil chose the repositories itself and has not published a cost breakdown. Without weights, outsiders cannot reproduce the results.
Claims and evidence
None of the funding figures has been confirmed by a regulatory filing.
Read the full section
| Claim | Status |
| Mirendil $1B; Kleiner Perkins and Clearlake lead | Company-announced (X); matched by Crunchbase |
| Mirendil valuation | Not disclosed. A September Bloomberg report described preliminary talks at $5B (Ultrathink) |
| Model trained for under $10M in 3.5 months | Self-reported; not replicated (MLQ) |
| Baseten's lowest-bits-per-byte result | No scores published; Baseten has a commercial relationship with Mirendil (ExplainX) |
| Mercor $750M at $20B | Company-announced (press release); repeated by SiliconANGLE. No lead investor named |
| Mercor annualized revenue above $2B (June) | Self-reported by co-CEO Brendan Foody (SiliconANGLE) |
| Mercor network of 5M+ experts | Self-reported |
None of the funding figures has been confirmed by a regulatory filing.
Context and prior work
- Mirendil's CEO, Behnam Neyshabur, co-led Google DeepMind's Blueshift team. CTO Harsh Mehta built the first version of Anthropic's internal AI research platform (Ultrathink).
- Mercor's growth tracks frontier labs' demand for expert data to train models after pretraining.
- In the same week, LangChain raised $300M led by CapitalG to "run, improve and govern AI agents" (Crunchbase News).
Read the full section
- Mirendil's CEO, Behnam Neyshabur, co-led Google DeepMind's Blueshift team. CTO Harsh Mehta built the first version of Anthropic's internal AI research platform (Ultrathink).
- Mercor's growth tracks frontier labs' demand for expert data to train models after pretraining. Its valuation went from $10B to $20B in about 13 months (SiliconANGLE).
- In the same week, LangChain raised $300M led by CapitalG to "run, improve and govern AI agents" (Crunchbase News).
Limitations, safety and contested findings
- Every Mirendil performance figure is run by the company. Neyshabur has promised independent evaluations later (Digg summary).
- Kimi-K3 outperforms Mirendil-Base on Mirendil's own leaderboard.
- Mercor's $750M includes an employee share tender.
Read the full section
- Every Mirendil performance figure is run by the company. Neyshabur has promised independent evaluations later (Digg summary).
- Kimi-K3 outperforms Mirendil-Base on Mirendil's own leaderboard.
- Mercor's $750M includes an employee share tender. The split between new capital and secondary sales was not disclosed, so the full amount should not be read as new money (Securities.io; SiliconANGLE).
- No independent reporting reviewed here corroborates the deal terms beyond what the companies announced.
Business and practitioner implications
- Evaluation is now a product. Buyers should expect to pay for evaluations on their own data.
- Research automation: if Mirendil's cost claim holds up, the cost of pretraining a model could fall sharply.
- Investor appetite: in one week, investors put $1B into a five-month-old company. That points to aggressive pricing for agent-infrastructure teams with frontier-lab backgrounds.
Read the full section
- Evaluation is now a product. Mercor co-CEO Adarsh Hiremath argues that models can score well on public benchmarks yet fail on a specific company's systems and policies (SiliconANGLE). Buyers should expect to pay for evaluations on their own data.
- Research automation: if Mirendil's cost claim holds up, the cost of pretraining a model could fall sharply. Treat it as unproven until independent evaluations or weights appear.
- Investor appetite: in one week, investors put $1B into a five-month-old company. That points to aggressive pricing for agent-infrastructure teams with frontier-lab backgrounds.