OpenRouter's own 32-day benchmark table puts DeepSeek V4 Flash 0731 at 76.8% TAU-Bench on one host and 46.4% on another. After an operator's measurements hit the top of Hacker News, OpenRouter said a QoS tier is coming.
ARC Prize ran GPT-6 Astra on ARC-AGI-3 under two harnesses and got 62.7% and 99.9%. Same model, same games. The only difference was whether reasoning state survived between calls. The leaderboard will now carry both numbers.
Nvidia Nemotron-3-Ultra-CC scored 535.4 at IOI 2026, above the 498.27 of the best human contestant. The paper breaks the score down, and most of it came from a test-time loop that resamples 200 candidates rather than from post-training.
Armature ran Claude Code, Codex, and Cursor through 16,893 sessions where each agent had to pick a third-party service and write the integration. The three agreed 42% of the time, and switching the repo language changed the winning email provider.
MBZUAI-backed IFM released six K2 Horizon models under Apache 2.0 on September 3. The 7B beats Qwen3.5-9B on SWE-bench Verified at 70.6%, but the training-code repos and dataset links are still closed.
OpenAI shipped GPT-6 Astra on September 3 at $10 and $50 per million tokens, 2.5x GPT-5.6 Sol. On coding work it burns a third of the tokens Sol needed, and the independent Intelligence Index puts both at 61.
Google shipped Gemini 3.8 Flash on September 2 at the same $0.75 per million input tokens as 3.7 Flash. DeepSWE climbs from 65.3% to 73.7%, but general agent work lands at 19.1% against Claude Opus 5 at 51.8%.
The anonymous ox-alpha model that appeared free on August 20 and topped OpenRouter usage within six days is GLM-5.3-Flash. Zhipu released the 320B-A18B weights under MIT the same day, and the free window is now closed.
Cerebras unveiled CS-4 on August 18, packing three wafers into one rack. It claims 4,400 tokens per second per user on gpt-oss-120B and up to 30x faster than GPUs, without naming the GPU product it measured against.
Anthropic shipped Claude Opus 5 on July 24 at the same $5/$25 per million tokens as Opus 4.8, claiming more than double the coding score of the previous generation and near-Fable 5 results at half the cost. Every number is Anthropic-measured.
MAI-Code-1-Flash is rolling into GitHub Copilot for VS Code, putting Microsoft’s own 5B coding model inside Copilot’s model picker and Auto routing path.
MiniMax M3 combines 1M context, multimodality, and coding-agent benchmarks, but its weights and technical report are still pending verification.