A research paper introduces PAIR, a method that counterfactually replays agents from identical states to pinpoint harmful context compression events and rewrite compression prompts.
Anthropic models lead all three domains, with Claude Fable 5.1 (Max) first in Code and Work, and Claude Sonnet 5.5 (Max) leading Chat. OpenAI's GPT 6 Astra (Max) placed in the top four across all categories.
Cohere says the WMT26 benchmarks used for North Small Translate were released after the model was created, so it could not be trained to those measurements.
OpenRouter has released Model Router Benchmarks to compare seven model routers side by side across quality, speed, and cost on six benchmarks, covering routers such as Jev Router, Unbiased Pareto, and NVIDIA Switchyard.
Artificial Analysis has open-sourced AA-AgentPerf-Local, which replays recorded agent sessions to measure local inference speed, and published initial 4-bit results for four systems. On models that fit in 32 GB, the GeForce RTX 5090 finished more than 3.5 times faster than the other systems tested.
Diogo Almeida says the datasets released with evals.typesafe.ai are being made easier to work with. He also says the team deprecates every dataset it evaluates on, internally or publicly.
On Ibragim’s my-mini-bench in pi, Sonnet 5.5 and Opus 5.5 High both passed 30 of 30 runs. Sonnet 5.5 took 46 seconds and $0.10 per task, versus 103 seconds and $0.30 for Opus 5.5 High.
Artificial Analysis says Claude Sonnet 5.5 scores 56 on its Intelligence Index at max effort, two points behind Claude Opus 5.5 (max), while using the highest output-token count it has measured and costing about $7.60 per task.
Artificial Analysis says ElevenLabs’ Eleven v4 text-to-speech model, with support for more than 90 languages, ranks first on its Provider Voice arena and Pronunciation Robustness benchmark, and second on Controlled Voice.
Artificial Analysis launched the Cyber Index and Cyber Index Alliance with CollinearAI, IBM, NVIDIA, and Vercel. The index combines three open benchmarks that score how AI agents find and fix vulnerabilities, with Grok 4.7 (xhigh) and MiMo-V2.6-Pro leading at 56.
A paper reports that TypeSafe AI’s Jev, with no extra training, separates alignment failures from good responses at a median AUROC of 0.886. Across 19 benchmarks, one Jev pass cost $0.30, compared with $18.96 for the LLM judges those benchmarks use.
Theo says a $1,000 DeepSWE benchmark of Jev Router performed about as well as GPT-6 Astra on low, while costing slightly more and taking almost five times as long.
Cline says the Pixel Canary stealth model has been released and is free in Cline. On Next.js Agent Evals, Cline says it ties GPT-6 Astra and beats Kimi K3.
elvis says humynlabs built BRIDGE ASR 2.0 to evaluate speech recognition on real two-person conversations lasting 10 to 15 minutes, in 18 Indic languages plus Spanish, Portuguese, and Vietnamese.
CASD has an off-the-shelf coding agent analyze a full set of agent logs and write the findings as rules in one prompt, with no environment access or validation data. DAIR.AI says one pass beat GEPA and SkillOpt on the reported benchmarks at about $1.60 per optimized prompt.
A paper from Microsoft and colleagues reports that the best model answers 59.7% of Taste-Bench questions correctly when choosing the better open direction in long engineering and research tasks. Forks decided by later evidence are much harder, a larger reasoning budget does not raise accuracy, and distilling a teacher’s outcome judgment improves held-out SWE-bench Pro success.
Gemini 3.8 Flash is free in Cline. Cline says it scores 41 on the Artificial Analysis Intelligence Index, runs at 291 tokens per second, and has a 1 million-token context window.
At max effort in Claude Code, Artificial Analysis measures Opus 5.5 at 66, above Opus 5 at 60 and Claude Fable 5.1 at 62. Its cost per task is $13.04, higher than Opus 5’s $10.79 despite lower token prices.
MiMo-V2.6-Pro, Claude Opus 5.5, GPT-6 Luna, and GPT-6 Sol account for 11 new points on Artificial Analysis’s Intelligence Index versus cost-per-task Pareto frontier, across different reasoning efforts. Claude Opus 5.5 (max with fallback) scores 58 at $5.98 per task, the highest result on the index.
Microsoft AI puts that pace at 377 points a year, or 1.6 times the next-fastest climb on the board.