The Threshold Report - report for 2026-09-08

Science+10Sep 4

Anthropic says Claude completed Lean proof of Fermat's Last Theorem

On Sept. 4, Anthropic said a Claude Code-based multi-agent system produced an end-to-end formalization of Fermat's Last Theorem in Lean, which checks proofs by computer, over 11 days. Machine checking can make long mathematical arguments easier to verify and help researchers manage a growing volume of AI-assisted proofs. The agents used the Prove2Me collaboration system and about six billion output tokens from an internal research model roughly comparable to Claude Fable 5.1. Anthropic says the system generated 13 million lines of Lean and 30,300 intermediate theorems, with 29,500 used in the final proof.

SOURCES · anthropic.com

Agents+5Sep 7

Image: thenextweb.com

Astra's 99.9% ARC score depends on OpenAI's harness

Yesterday, The Next Web reported that GPT-6 Astra's ARC-AGI-3 score differed sharply between two evaluation harnesses, meaning the software that manages a model's context and tools during a test. OpenAI's Provider Adapter preserves opaque reasoning state and handles context differently, making the surrounding setup part of the measured system. ARC Prize recorded 62.7% for $26,098 with its standard harness and 99.9% for $18,817 with the adapter. Using the adapter with no reasoning effort, Astra still scored 96.7%, above its maximum-reasoning result under the standard harness.

SOURCES · thenextweb.com

Robots+5Sep 7

Image: electrek.co

XPeng starts automated production line for IRON humanoid

XPeng said yesterday that it had moved IRON from research prototypes onto an automotive-style production line in Guangzhou. Manufacturing the robot through an industrial line tests whether XPeng can repeatedly build the design under production quality systems. More than 80% of the line's core processes are automated, while IRON uses three in-house Turing AI chips and is targeting mass production by the end of 2026.

SOURCES · electrek.co

Agents+2Text+2Sep 8

Image: The Threshold Report/GPT Image 2

Accelerator Agents releases Gemini tools for moving models to TPUs

Accelerator Agents published a repository of Gemini-assisted TPU development tools today. Developers adapting models for TPUs now have a reusable starting point for migration work that usually requires specialized accelerator knowledge. MaxCode converts PyTorch code and model layers to JAX and MaxText, while MaxKernel drafts, ports, profiles, and tests TPU code written with Pallas. The tools use Gemini for agent reasoning and include workflows for CUDA-to-Pallas conversion and generating test setups.

SOURCES · github.com

Agents+2Sep 7

Image: github.com

LLM-as-a-Verifier cuts token use and adds self-check tests

Version 0.2.0 of the public LLM-as-a-Verifier framework arrived yesterday with a prefix-cache optimization that reuses earlier input during evaluations with many task-run histories. The project says the change reduces uncached input tokens by a factor of about 3.4. Agent builders can use the package to rank candidate runs, track progress during a task, or stop attempts with low prospects before spending more inference budget. In Terminal-Bench 2.1 tests, repeatedly selecting a model's own runs exceeded its reported pass-at-one score, meaning its success rate on a single attempt.

SOURCES · github.com

Compute & Infra+2Sep 7

Ironwood undercuts Nvidia chips in third-party cost estimate

SemiAnalysis published what it calls the first third-party InferenceX results for Google's TPUv7 Ironwood yesterday, running Qwen3.5 397B with FP8, an eight-bit number format. If software support outside Google matures, operators serving supported open models could gain a lower-cost option for some inference work. At a lower-interactivity setting, the tests put Ironwood's throughput 5% above both Nvidia chips. At 100 tokens per second per user, SemiAnalysis estimated $0.181 per million total tokens for Ironwood, compared with $0.222 for B200 and $0.276 for B300.

SOURCES · inferencex.semianalysis.com

Robots+2Text+2Sep 8

Image: github.com

Qwen releases open driving model with planning and 3D perception

Today, Qwen publicly released Qwen-Drive-1.0, a downloadable 4-billion-parameter vision-language model that connects driving-scene understanding with trajectory planning. Researchers can inspect and run one open system for both parts of the driving task. The release includes code and demo scenes for visual question answering, 3D perception, and five-second driving trajectories. Qwen says its planner, optimized using reward scores, reached 90.7 on the PDMS measure in NAVSIM v1.1.

SOURCES · github.com

Robots+2Sep 4

Astra places blocks reliably, struggles with puzzle insertion

Robocurve reported on Sept. 4 that GPT-6 Astra put a block into a bowl in 19 of 20 robot-arm trials, compared with 8 of 20 for Claude Fable 5.1. The result suggests stronger general-purpose models can improve straightforward vision-guided manipulation while leaving precise physical work unreliable. Astra took 2.5 minutes and an estimated $0.94 per bowl run. On puzzle-piece insertion, both models completed only 2 of 20 attempts.

SOURCES · openai.robocurve.org

Anthropic says Claude completed Lean proof of Fermat's Last Theorem

Astra's 99.9% ARC score depends on OpenAI's harness

XPeng starts automated production line for IRON humanoid

Accelerator Agents releases Gemini tools for moving models to TPUs

LLM-as-a-Verifier cuts token use and adds self-check tests

Ironwood undercuts Nvidia chips in third-party cost estimate

Qwen releases open driving model with planning and 3D perception

Astra places blocks reliably, struggles with puzzle insertion

AI Investments

Security

Wiz finds Copilot-linked flaw that could expose Snowflake Jira

Other News

Named testing methods fail to improve coding agents reliably

OpenAI reports 20% drop in AI serving costs

Upcoming

OpenAI contracts future Malaysian data-center capacity

GLP starts construction on Foshan data-center campus