# Anthropic authors report progress on provable inference for large language models [Canonical reading](https://escapemeridian.com/readings/anthropic-authors-report-progress-on-provable-inference-for-3a9e7ff4) Anthropic’s authors present FlashProve, an open-source research prototype intended to let users check that a model response came from specific committed model weights, without seeing those weights. They argue that this could help users, auditors and regulators detect model substitution and verify claims about which model served a request. They also describe possible uses in monitoring AI development and enforcing model approval, while noting that proofs of inference do not show that hardware is not also being used for training. The authors report that FlashProve generates proofs for Llama 4 Maverick at about 110 tokens per second on a single H100 GPU. Each proof is about 19 MB, and checking it takes seconds on a CPU; the proving process takes roughly 70–95 times the GPU time of ordinary inference in their comparisons. The release includes code and Lean proofs, but the authors identify gaps between the verified protocol and executable code. The system proves an integer model under greedy decoding, supports contexts up to 1,024 tokens, and does not hide the model architecture. How well can the approach’s proof speed and verification guarantees extend to longer contexts and deployed inference systems? Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source Anthropic · Article Original source date: Nov 2, 2026 (2026-11-02T00:00:00.000Z). Date basis: Date stated by the publisher. Added to Escape Meridian: 2026-11-11T03:08:54.639Z. First observed: 2026-11-03T06:40:08.778Z. Retrieved: 2026-11-03T14:47:31.658Z. - [Progress on provable inference for large language models](https://www.anthropic.com/research/provable-inference) ## Limits The authors describe FlashProve as a research prototype and say its proof covers an integer version of the model under greedy decoding, not necessarily outputs from floating-point inference or sampled decoding. The authors do not guarantee that the Lean protocol, paper and fast executable code match one another; they also say the formal statements require manual auditing. The system supports contexts up to 1,024 tokens and does not hide the model architecture. The reported performance is measured on Llama 4 Maverick under the stated test setup; the post describes intermediate progress, not a completed research agenda. ## Open question How well can FlashProve’s proof speed and verification guarantees extend to longer contexts and deployed inference systems? ## Source-stated The authors describe FlashProve as a research prototype for proving that responses were generated by committed model weights, while keeping the weights hidden. > The goal of provable inference (PI) is to enable a user with limited computational resources and without any information about the weights to verify that a response came from the requested model. - [Supporting source](https://www.anthropic.com/research/provable-inference) ## Source-stated The authors report proof generation at about 110 tokens per second for Llama 4 Maverick on one H100 GPU, with proofs around 19 MB and a reported GPU-time overhead of roughly 70–95 times ordinary inference. > The online phase, on a batch of 24 requests, takes the prover about 223 seconds, which is 110 tokens per second, and each proof is 18.78 MB. - [Supporting source](https://www.anthropic.com/research/provable-inference) ## Source-stated The authors say formal verification does not establish that the released executable code matches the formally defined protocol, and that the system proves an integer model under greedy decoding. > The paper and the fast executable code are derived from the Lean protocol, but we do not guarantee that all three (Lean, paper, code) match each other. - [Supporting source](https://www.anthropic.com/research/provable-inference) ## Source-stated The authors identify the current 1,024-token context limit as a technical limitation. > The system currently proves inference over contexts of up to 1,024 tokens. - [Supporting source](https://www.anthropic.com/research/provable-inference) ## Source and contributor context Anthropic · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media Media declared by the original publisher; availability checked separately from the editorial reading. - [Progress on provable inference for large language models · Anthropic](https://www.anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F2eba335a28661a8253187053fa36add230af728b-2000x1125.jpg&w=1920&q=75) ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Model weights The learned numbers that make up a trained model; publishing them lets others run it. - [Definition reference](https://developers.google.com/machine-learning/glossary#weight) ## Model tokens Small chunks of text, often parts of words, that a language model reads and writes; usage and prices are often counted in tokens. - [Definition reference](https://developers.google.com/machine-learning/glossary#token) ## NVIDIA H100 An NVIDIA graphics processing unit (GPU) built on the company's Hopper architecture and designed to accelerate AI workloads, including large language models. - [Definition reference](https://www.nvidia.com/en-us/data-center/h100/) ## CPU The central processing unit: a general-purpose processor that carries out a computer's instructions. - [Definition reference](https://aws.amazon.com/what-is/cpu/) ## Graphics processing unit A chip that performs many calculations at once; the main hardware used to train and run AI models. - [Definition reference](https://en.wikipedia.org/wiki/Graphics_processing_unit) ## Anthropic Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems. - [Explore this context](/entities/anthropic) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Lean Lean is a programming language and proof assistant used to express mathematical statements and check formal proofs. - [Explore this context](/entities/lean) - [Reference](https://lean-lang.org/) - [Reference](https://lean-lang.org/theorem_proving_in_lean4/Axioms-and-Computation/) ## Editorial provenance Version 1. Publication ID: b696c0f4-5de2-4374-8ecb-a2967efc6635. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-11T03:08:54.639Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 6f598a2705df6e8c8050ebb24c548d1e6f910aee586cc96b8b47ab6368c36c35. Source text hash: e962b9e51fbac571a286aadfd1a91360aadd8f06499b3df5051ca3a534169565. ## Revision and source-check history ### Version 1 — 2026-11-11T03:08:54.639Z Anthropic’s authors present FlashProve, an open-source research prototype intended to let users check that a model response came from specific committed model weights, without seeing those weights. They argue that this could help users, auditors and regulators detect model substitution and verify claims about which model served a request. They also describe possible uses in monitoring AI development and enforcing model approval, while noting that proofs of inference do not show that hardware is not also being used for training. The authors report that FlashProve generates proofs for Llama 4 Maverick at about 110 tokens per second on a single H100 GPU. Each proof is about 19 MB, and checking it takes seconds on a CPU; the proving process takes roughly 70–95 times the GPU time of ordinary inference in their comparisons. The release includes code and Lean proofs, but the authors identify gaps between the verified protocol and executable code. The system proves an integer model under greedy decoding, supports contexts up to 1,024 tokens, and does not hide the model architecture. How well can the approach’s proof speed and verification guarantees extend to longer contexts and deployed inference systems? ## Original source Anthropic · Article Original source date: Nov 2, 2026 (2026-11-02T00:00:00.000Z). Date basis: Date stated by the publisher. Added to Escape Meridian: 2026-11-11T03:08:54.639Z. First observed: 2026-11-03T06:40:08.778Z. Retrieved: 2026-11-03T14:47:31.658Z. - [Progress on provable inference for large language models](https://www.anthropic.com/research/provable-inference) ## Limits The authors describe FlashProve as a research prototype and say its proof covers an integer version of the model under greedy decoding, not necessarily outputs from floating-point inference or sampled decoding. The authors do not guarantee that the Lean protocol, paper and fast executable code match one another; they also say the formal statements require manual auditing. The system supports contexts up to 1,024 tokens and does not hide the model architecture. The reported performance is measured on Llama 4 Maverick under the stated test setup; the post describes intermediate progress, not a completed research agenda. ## Open question How well can FlashProve’s proof speed and verification guarantees extend to longer contexts and deployed inference systems? ## Source-stated The authors describe FlashProve as a research prototype for proving that responses were generated by committed model weights, while keeping the weights hidden. > The goal of provable inference (PI) is to enable a user with limited computational resources and without any information about the weights to verify that a response came from the requested model. - [Supporting source](https://www.anthropic.com/research/provable-inference) ## Source-stated The authors report proof generation at about 110 tokens per second for Llama 4 Maverick on one H100 GPU, with proofs around 19 MB and a reported GPU-time overhead of roughly 70–95 times ordinary inference. > The online phase, on a batch of 24 requests, takes the prover about 223 seconds, which is 110 tokens per second, and each proof is 18.78 MB. - [Supporting source](https://www.anthropic.com/research/provable-inference) ## Source-stated The authors say formal verification does not establish that the released executable code matches the formally defined protocol, and that the system proves an integer model under greedy decoding. > The paper and the fast executable code are derived from the Lean protocol, but we do not guarantee that all three (Lean, paper, code) match each other. - [Supporting source](https://www.anthropic.com/research/provable-inference) ## Source-stated The authors identify the current 1,024-token context limit as a technical limitation. > The system currently proves inference over contexts of up to 1,024 tokens. - [Supporting source](https://www.anthropic.com/research/provable-inference) ## Source and contributor context Anthropic · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media Media declared by the original publisher; availability checked separately from the editorial reading. - [Progress on provable inference for large language models · Anthropic](https://www.anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F2eba335a28661a8253187053fa36add230af728b-2000x1125.jpg&w=1920&q=75) ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Model weights The learned numbers that make up a trained model; publishing them lets others run it. - [Definition reference](https://developers.google.com/machine-learning/glossary#weight) ## Model tokens Small chunks of text, often parts of words, that a language model reads and writes; usage and prices are often counted in tokens. - [Definition reference](https://developers.google.com/machine-learning/glossary#token) ## NVIDIA H100 An NVIDIA graphics processing unit (GPU) built on the company's Hopper architecture and designed to accelerate AI workloads, including large language models. - [Definition reference](https://www.nvidia.com/en-us/data-center/h100/) ## CPU The central processing unit: a general-purpose processor that carries out a computer's instructions. - [Definition reference](https://aws.amazon.com/what-is/cpu/) ## Graphics processing unit A chip that performs many calculations at once; the main hardware used to train and run AI models. - [Definition reference](https://en.wikipedia.org/wiki/Graphics_processing_unit) ## Anthropic Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems. - [Explore this context](/entities/anthropic) - [Reference](https://www.anthropic.com/news) - [Reference](https://www.anthropic.com/research/team/economics) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Lean Lean is a programming language and proof assistant used to express mathematical statements and check formal proofs. - [Explore this context](/entities/lean) - [Reference](https://lean-lang.org/) - [Reference](https://lean-lang.org/theorem_proving_in_lean4/Axioms-and-Computation/) ## Editorial provenance Version 1. Publication ID: b696c0f4-5de2-4374-8ecb-a2967efc6635. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-11T03:08:54.639Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 6f598a2705df6e8c8050ebb24c548d1e6f910aee586cc96b8b47ab6368c36c35. Source text hash: e962b9e51fbac571a286aadfd1a91360aadd8f06499b3df5051ca3a534169565. ### uncertain — 2026-11-11T03:08:54.639Z — anthropic/claude-opus-5.5 Publication: b696c0f4-5de2-4374-8ecb-a2967efc6635 attribution — uncertain: The post is published by Anthropic, but the byline is mixed: several authors are affiliated with EPFL, MIT, StarkWare, RESI and Succinct, or are Anthropic fellows. The title and summary say 'Anthropic authors' and 'Anthropic's authors', which overstates the affiliation. Smallest repair: 'Anthropic researchers and academic collaborators report...' or 'An Anthropic research post reports...'. > Alessandro Chiesa (EPFL, StarkWare), Miranda Christ (MIT, RESI) summary — uncertain: The reading leaves out the post's lead framing that Claude produced the zkSNARK improvements and found novel ways to apply cryptographic tools. That is a central idea for this collection, and restoring one attributed sentence would fix it. Several terms that matter for understanding are also unexplained for a general reader: 'committed model weights' (a hidden fingerprint of the weights), 'Lean proofs' (machine-checked formal proofs), 'integer model under greedy decoding', and 'H100 GPU'. The question is substantive and follows from the evidence. > Claude has produced improvements on zkSNARKs for inference, building a system that is substantially faster than prior open-source systems. date — pass: The originalPublished date of 2026-11-02, taken from the publisher page, matches the feed date and comes before first observation. No earlier submission date is indicated. support — pass: The source supports each reported figure: about 110 tokens/s on one H100, proofs of about 19 MB, verification in seconds on a CPU, and roughly 70–95x the GPU time of vLLM. The listed limitations are also stated in the source: the integer model under greedy decoding, the 1,024-token context, architecture not hidden, and no guarantee that the Lean protocol, paper and code match. The source also states the caveat that proofs do not rule out training on the side. The described uses are attributed to the authors as possibilities, not presented as demonstrated benefits. > it does not establish that the hardware is not, e.g., training on the side media — pass: The selected media is the verified direct link to the original Anthropic research post. It is the most direct way for a reader to access the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/3a9e7ff4ac6901f078cdb57c1f23e13426aebd16198e32e9a30d1b91c8de3851/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/anthropic-authors-report-progress-on-provable-inference-for-3a9e7ff4.json)