GPT-56T 861 —
MUSE-SPK 837 +0.2%
GPT-56SC 790 -4.6%
GLM-5 781 -0.4%
CL-OP55X 780 -5.1%
GROK-46H 780 -5.1%
QWEN-38X 748 -9.2%
GPT-6A 743 -9.4%
KIMI-K3X 742 —
CL-FAB5H 698 -6.1%
CL-OP5H 675 -6.2%
GEM-38FH 672 -0.7%
CL-OP5X 670 -5.5%
CL-OP55H 668 —
CL-OP46H 657 -5.9%
CL-OP47H 648 -6.1%
GPT-56S 618 -0.6%
GEM-37FH 610 -7.2%
GEM-36FH 593 —
CL-OP48H 588 —
CL-OP47 581 -0.2%
GEM-35FH 580 —
GPT-55H 541 -7%
INKL 531 —
GEM-31P 512 -0.2%
CL-OP46 498 +0.4%
GEM-3P 498 -0.2%
CL-OP48 492 +0.4%
GPT-52 464 —
GPT-55 423 —
GPT-56T 861 —
MUSE-SPK 837 +0.2%
GPT-56SC 790 -4.6%
GLM-5 781 -0.4%
CL-OP55X 780 -5.1%
GROK-46H 780 -5.1%
QWEN-38X 748 -9.2%
GPT-6A 743 -9.4%
KIMI-K3X 742 —
CL-FAB5H 698 -6.1%
CL-OP5H 675 -6.2%
GEM-38FH 672 -0.7%
CL-OP5X 670 -5.5%
CL-OP55H 668 —
CL-OP46H 657 -5.9%
CL-OP47H 648 -6.1%
GPT-56S 618 -0.6%
GEM-37FH 610 -7.2%
GEM-36FH 593 —
CL-OP48H 588 —
CL-OP47 581 -0.2%
GEM-35FH 580 —
GPT-55H 541 -7%
INKL 531 —
GEM-31P 512 -0.2%
CL-OP46 498 +0.4%
GEM-3P 498 -0.2%
CL-OP48 492 +0.4%
GPT-52 464 —
GPT-55 423 —

Live Feed

5mo ago release

Qwen3.6-27B: 55.6 GB Beats the 807 GB Model It Replaced on Every Coding Benchmark

Alibaba's dense 27B model outscores the 397B Qwen3.5-397B-A17B on SWE-bench Verified (77.2% vs 76.2%), runs quantized in 16.8 GB on a single consumer GPU, and matches GPT-5.4 on the same benchmark. Apache 2.0, no commercial restrictions.

5mo ago research

GPT-5.4 Pro Closes a 1968 Erdős Problem With a Method Mathematicians Overlooked for 90 Years

An amateur with a ChatGPT Pro subscription solved Erdős problem #1196 in 80 minutes using a single prompt. Stanford mathematician Jared Lichtman called it the first AI proof at the level of Erdős's Book. Terence Tao extended it into new theory within 24 hours.

5mo ago research

Meta Paper: Test-Time Scaling for Coding Agents Is a Representation Problem, Not a Compute Problem

A Meta research paper shows that compressing prior agent attempts into structured summaries — then using those summaries to guide new attempts — lifts Claude 4.5 Opus from 70.9% to 77.6% on SWE-Bench Verified and 46.9% to 59.1% on Terminal-Bench v2.0. The ceiling on long-horizon coding agents was not compute. It was memory.

5mo ago research

Anthropic Ran a Secret Agent Marketplace Where AI Closed Real Deals for Real Money

Dubbed Project Deal internally, Anthropic built a classified marketplace where Claude-powered agents acted as both buyers and sellers, negotiating and completing transactions on actual goods with actual currency. No human approved the final trade.

5mo ago release

OpenAI Open-Sources Privacy Filter: 97.4% F1 on PII Masking, 50M Active Params, Runs Offline

OpenAI released Privacy Filter on April 22 under Apache 2.0 — a 1.5B-parameter PII detection model that processes up to 128,000 tokens in a single pass, runs fully locally, and hits 97.4% F1 on the PII-Masking-300k benchmark.

5mo ago policy

OpenAI Puts $25,000 on Finding a Bio Jailbreak in GPT-5.5 — 5 Questions, Codex Desktop Only

OpenAI launched a targeted Bio Bug Bounty for GPT-5.5 the same day the model shipped, offering $25,000 for a universal jailbreak that clears five biosafety questions. The model is classified as High biological capability under the Preparedness Framework.

5mo ago release

OpenAI Launches Free ChatGPT for Clinicians — GPT-5.4 Scores 59.0 on HealthBench Pro, 35% Above the Physician Baseline

OpenAI is giving verified US clinicians free access to ChatGPT with HIPAA support, clinical search, and citation-backed responses. Alongside the launch, it released HealthBench Professional — an open benchmark developed with the AMA and NLM — where GPT-5.4 scores 59.0 against a physician baseline of 43.7.

5mo ago research

Google DeepMind's Vision Banana Beats SAM 3 and Depth Anything V3 — With Zero Specialist Training

A new Google DeepMind paper proves image generation training produces generalist vision capability. Vision Banana, built by instruction-tuning Nano Banana Pro on a small fraction of vision data, outperforms SAM 3 on three segmentation benchmarks and Depth Anything V3 on metric depth estimation — all in zero-shot transfer.

5mo ago research

Google Adds Marvell as Chip Design Partner for AI Inference — Marvell +6%, Broadcom -2%

Google is in advanced talks with Marvell Technology to co-develop two new AI chips: a memory processing unit and an inference-optimized TPU. The move diversifies Google's custom silicon supply chain beyond Broadcom and triggered an immediate semiconductor market reaction.

5mo ago research

61GW of AI Data Center Capacity Is Phantom: Sightline Finds Only 40% of Announced Projects Will Survive

Sightline Climate tracked 710 data center projects representing 102.3GW of announced capacity through 2030. Only 40.8GW qualifies as derisked. Grid interconnection queues, local opposition, and permit timelines are blocking projects at scale. The constraint has shifted from chips to concrete and public permission.

5mo ago research

DeepSeek V4's Architecture Paper: 90% Less KV Cache at 1M Tokens, a New Residual Path, and Muon at Scale

DeepSeek's V4 technical report details three architectural changes that cut 1M-token compute to 27% and KV cache to 10% of V3.2. Hybrid compressed-block attention, a stabilised residual path, and the Muon optimizer running at trillion-parameter scale form a new efficiency recipe for long-context open models.

5mo ago funding

Google Commits $40B to Anthropic at $350B Valuation — With Another $30B Tied to Milestones

Alphabet is investing $10B in Anthropic immediately, with up to $30B more contingent on performance targets. Combined with Amazon's $25B commitment, Anthropic now has $65B in pledged backing from two of the three largest cloud providers.

5mo ago funding

Tesla Buries a $2B AI Hardware Acquisition in Its Q1 10-Q — Target Company Unnamed

A single sentence in Note 14 of Tesla's Q1 2026 filing discloses an agreement to acquire an unnamed AI hardware company for up to $2 billion in stock and equity awards, with $1.8B contingent on milestones. The all-stock structure on a company with $44.7B in cash points to an earnout-heavy acqui-hire.

5mo ago model

xAI Ships Grok Voice Think Fast 1.0 — Flagship Voice Agent Posts 70% Autonomous Resolution at Starlink

xAI's new voice agent model reasons in the background, handles complex multi-turn conversations, and resolves 70% of Starlink customer support calls without human escalation. It is available via API from April 23.

5mo ago research

OpenAI Open-Sources Chain-of-Thought Monitorability Evals — and Back-Fills Results for GPT-5.4 Through o3

OpenAI has released datasets and reference code from its chain-of-thought monitorability research, introducing a new metric called g-mean2 and back-filling results across four frontier models. The goal: make CoT monitoring a standardized, externally verifiable safety layer.