556. ResolveAI post-trained SRE-1, a 27B dense model, for root-cause analysis of production incidents. One run matches the frontier model they tested at medium effort for a fifth of the cost; five ensembled tie its max effort for 37% less. No public benchmark.
resolve.ai/blog/introducing-…
555. Gemini 4 Neon is not a new ceiling, it is a cost line: $3/$15 per 1M, 1M context, 95% off cached input. #3 on the Artificial Analysis Intelligence Index at 61 (xhigh), #1 on their Coding Agent Index. Google claims 32% lower cost per task than Carbon at default thinking.
Gemini 4 Neon is rolling out now.
Sitting between Carbon and Hydrogen, it brings frontier intelligence and fast task completion for complex software engineering, knowledge work, and science.
554. RadixArk trained DaisyRose-27B on 1,200 agent tasks for 18 days on 144 H200s: 63,000 GPU-hours, 37% MFU vs 16% for a Megatron-based open trainer. Terminal-Bench 4.1 pass@1 went 5.6% to 12.0%, a doubling, but 88% of tasks still fail. Team-reported, Apache 2.0. huggingface.co/blog/RadixArk…
553. OpenAI says its models found up to 14% PPA improvement on AMD's flagship GPU IP in under 5 months, and cut standard-cell area 13% in Jalapeno's vector unit and 10% in the matrix unit. Vendor-reported, measured against already-optimized baselines.
Working with @AMD, our models identified up to 14% PPA improvement on flagship GPU IP in less than five months—building on designs AMD’s engineers had already spent years optimizing.
The same approach helped improve Jalapeño, our first custom inference chip, with gains verified on real silicon.
New evaluations also show our models getting better at certain chip design tasks.
We’re sharing these results and how we’re working with @Synopsys to bring our model capabilities and their trusted EDA design and validation tools to more chip designers.
openai.com/index/advancing-c…
552. The Gemini API now takes Word, PowerPoint and Excel natively, and you can slice them: page ranges, worksheet tabs, cell ranges, or layers like page image vs extracted text. Only the slice is billed, 280/560/1,120 tokens per page at low/medium/high. ai.google.dev/gemini-api/doc…
551. Rumble-2-Flash is a 0.5B ASR model, Apache 2.0, 40+ languages. On 302 hours of call-centre audio across 14 accents it reports 7.86% WER vs NVIDIA's 30B Nemotron 3.5 Transcribe at 7.78%, so self-hosted transcription stops needing a GPU. Vendor-reported.
550. Luma AI, 이미지·영상 이해와 생성을 통합한 멀티모달 모델 Uni-1 공개. 픽셀 단위 시각 추론, 최대 14장 레퍼런스, 120fps·1080p 실시간 생성과 24fps 오디오 동기(첫 프레임 1초 미만)를 지원하며 Uni-1.1 API도 제공
lumalabs.ai/uni-1
549. Sakana AI, 긴 문맥 문서를 한 번의 forward pass로 LoRA 어댑터에 즉시 내재화하는 Doc-to-LoRA 공개. 하이퍼네트워크가 문서 문맥을 직접 가중치로 변환해, 미지의 텍스트·비전 등 멀티모달 문맥을 추가 그래디언트 파인튜닝 없이 처리
sakana.ai/doc-to-lora/
Cerebras put disaggregated inference into production: AMD Instinct GPUs run prefill, CS-3 wafers run decode. On a shared GPT-OSS-120B endpoint it measured 5x tokens per megawatt vs CS-3-only, with no loss in generation speed.
Vendor-reported. cerebras.ai/blog/cerebras-la…
ARC-AGI-3 has its first 100%: GPT-6.2 Astra. The harness explains as much as the model.
Provider adapter (keeps reasoning state, compacts context): 100%, $4.1K.
Standard (carries only model-selected notes): 70.3%, $24.2K.
On Standard, Claude Fable 5.2 still leads at 81.2%.
GPT-6.2 Astra from @OpenAI on ARC-AGI (Verified):
- ARC-AGI-3: 100.0% (SOTA), $4.1K (provider adapter), 70.3%, $24.2K (standard)
- ARC-AGI-2: 95.8% (SOTA), $0.50/task
- ARC-AGI-1: 98.5%, $0.17/task
GPT-6.2 Astra is the first model to achieve a 100% score on ARC-AGI-3.
Microsoft: MAI Code 2 gives 96% of Claude Opus 5.5 at 5% of the cost. Read the footnote: 3 of 4 benchmarks ran on Microsoft's own harness, costs are a 50% launch discount, and Microsoft says it's "explicitly not yet tuned" for long-horizon agentic work. microsoft.ai/news/introducin…
Cheaper, not just better: GPT-6.2 Astra beats GPT-6 Astra by 5.2 points on Agents' Last Exam at under a third the cost per task, and Claude Fable 5.2 by 5.5 at 5%. Token prices drop 20% and cached input 60%. Less advertised: both 6.2 models omit blockers more often.
Some of you have been asking very politely. Some of you have just been asking.
GPT-6.2 Astra and Sol are now available.
Super efficient and built to perform where it counts—the real world.
openai.com/index/introducing…
Runta's harness eval, 27,280 runs on 5 open models: 64.5% of list-price inference cost went to runs that did not pass, most ending in a confident answer rather than a timeout. At a fixed model, the harness moved cost per task 1.8-5x.
runta.com/blog/introducing-f…
OVHcloud's first LLM: 7B total, 1B active over 256 experts per layer, hybrid Differential-TPA + Mamba-3. No eval scores, just fit: 8 GB GPU, 160 tok/s on an RTX A4500, 32k context. Trained from scratch on 3 EU public supercomputers, ~EUR 7M of compute hours, Apache 2.0.
OVH launches Olala, its first family of language models. First of the family, Olala mini 7A1B is a compact model with 7 billion parameters, 1 billion of them activated, trained from scratch on European public supercomputers. It relies on an innovative architecture, runs even without a GPU, and will be released as open source.
A model has become a strategic building block. We chose to build it rather than depend on it. This model proves that European companies can innovate and train agentic LLMs from scratch.
Next step: Olala v1.1 by the end of the year with improved performance, then Olala v2.0, which will exceed 100 billion parameters.
More info ➡️ lnkd.in/gyreQ8nz
Come and discover Olala mini 7A1B and the details of our cloud and AI innovations at the OVHcloud AI Summit, on 19 November.
#OVHcloudAISummit
Quietly big: Hugging Face swapped its own router for OpenRouter's on Inference Providers. Same token and API, but a failed request now retries on the next provider. Your own provider API keys stop working.
Open-weight is now 71.7% of OpenRouter tokens.
huggingface.co/blog/openrout…
Sarvam's Token Factory serves Sarvam 105B, GLM 5.3, DeepSeek V4.1 Flash, Kimi K3.1 Pro and Gemma 4 31B from B200s in India on one API.
Their own table is the tell: the 2.1x first-token win over stock vLLM+Dynamo is prefill, not decode. Per-user output 192 vs 190 tok/s.
Today, we are launching Sarvam Token Factory for everyone.
We host our own models, and the full range of best-in-class open-weight models, in India. We match leading inference providers on performance and reliability.
For enterprises, this means more models to choose from, and an AI stack they can own without lock-ins.