Meta acquired Conductor, an AI coding startup founded in 2024, Meta confirmed to TechCrunch. Conductor founder and CEO Charlie Holtz said core features keep working until early next year, but the cloud service shuts down November 20 and user cloud data will then be deleted. Meta gave no deal terms.
OpenAI says it will acquire Pierre Computer Company, maker of Code Storage, Trees and Diffs, on undisclosed terms announced November 4. OpenAI said the deal folds Code Storage's Git-based storage features into Codex, which already runs Pierre's code browsing and review tools. No purchase price was disclosed.
Researchers affiliated with XPENG, Peking University and the University of Hong Kong release GroundingPI, an open-source 4B-parameter visual perception model. The team reports it scored 73.68 across 34 perception benchmarks, edging GPT-6 Astra at 71.54, and says simulation tests cut robot demonstration data needs. Weights and code are public.
Anthropic releases FlashProve, an open-source prototype that generates cryptographic proofs that a specific model produced a given response. Anthropic says a single H100 proves Llama 4 Maverick at about 110 tokens per second, with 18.78 MB proofs verified in seconds, roughly 70 to 95 times slower than plain inference on vLLM. Contexts cap at 1,024 tokens.
Vals AI reports that four of seven frontier models kept at least 80% of their Vals Index score at half the cost or less when run at their lowest reasoning setting. Claude Opus 5.5 held 80% of its score at 16% of the cost, researcher Derrick Liang writes. In 22 of 56 model-benchmark pairs, cheaper settings matched leaderboard scores.
Hadley Wickham argues Kimi K3 token prices roughly match the cost of serving them, citing a self-hosted 8-GPU Supermicro server at $550,000 and about $23,000 a month, or $0.25 per million tokens at 75% utilization. Subsidies sit in subscriptions instead: SemiAnalysis estimates a $200 monthly plan delivers $2,000 to $10,000 in tokens.
DeepSeek plans to pre-train its next-generation DeepSeek V5 model on Huawei Ascend chips, STAR Market Daily reports, citing multiple industry sources. Ascend has mostly handled inference until now. The reported plan is unconfirmed, and viability depends on training stability, cluster performance and software maturity.
OpenAI says its models now help engineers design, simulate, and build physical products, and that two staffers used GPT-6 to design a Codex talk device in under a week. A contract manufacturer built prototypes in under a month, though Astra missed production tolerances and assembly details. Model token usage across OpenAI hardware teams grew 23-fold since February.
Baseten adds three products to its inference platform. Grid, a gateway and router with guardrails, opens a public preview Nov. 4; Sandboxes, VM-isolated agent environments, reach public preview Nov. 5, and Loops, serverless training, hits general availability Nov. 6. Baseten says the stack runs at 99.99% uptime.
Fireworks launched Ultra-fast inference in public preview at its Forge conference, claiming up to 1,500 tokens per second on NVIDIA Blackwell GPUs. Nexus Desktop joins the Nexus CLI, with Pro plans starting at $8 per month. Fireworks also previewed task-specific evaluations and Fireworks Sovereign, which runs its inference engine inside customer infrastructure.