A question picked from the 17 records added this week. We had read the move to open weights as narrowing on the license side, but this week two records moved the other way.
It is the first model from a new US lab. Company-reported scores were SWEBench Verified 81.6%, GPQA Diamond 90.6%, AIME 2026 97.7% and Terminal Bench v2.1 80.1% (on BF16), and it ships in three versions, BF16, FP8 and NVFP4, with the NVFP4 checkpoint at about 280GB
The authors reported that in the setting where the constraints have to be found by the model, even the best configuration averaged only 0.56 points and solved just 14 of the 95 tasks. Even in the 23 tasks where every necessary fact was obtained from the stakeholder, only 10 scored 0.8 or above, and they concluded that finding the unstated constraints is the bottleneck
In a company-reported sample (Sep 21–27, 2026; about 250,000 US conversations), politics, policy and society topics made up 5.75%, elections and voting 0.51%, conversations asking who to vote for 0.01%, and election-integrity policy violations 0.002%. In its own evaluation the political balance score for Opus 5.5, Sonnet 5.5 and Haiku 5.5 was 99%, but an outside reviewer pointed to a tendency to add caveats that were not asked for and to the risk of missing an operation broken into steps that each look harmless
What ties the events together — my own reading.
A new lab released a model with 501 billion parameters in total under the loosest license there is, and a small forecasting model opened up in much the same way. Right up to last month, the higher a model ranked the more it turned to a license of its own, but this week the cases on the other side piled up. That said, every performance figure released is company-reported, and there are no outside results redoing them yet, so it is early to call a trend from one week.
In a new evaluation modelled on the work of actually handing results over to a customer, even the best setup finished only 14 of 95 tasks. Where they got stuck was not ability but the part where they had to dig out, on their own, the conditions nobody had told them. Another study out the same week said that bundling several agents together raises the score but pushes the cost up three or four times, and as much as seven.
Laptops with unified memory raised to 128GB have started selling, and one forecasting model runs on an ordinary processor alone, with no graphics card. A bundle also came out that handles chat, judging and dictation in one place while leaving memory as it was. Apart from the race on performance, moves to change where things get used gathered in the same week.
Take the week's quiz (10 questions)
Records and notes, together or apart.
A conversation that asks someone who has actually run a continuously running agent what the prerequisites are. The conclusion lands not on the model or the prompt but on two things. One is a single dedicated machine that
A hands-on run that follows all the way through the work of nailing down, as files, a structure that removes the need to ask for a fix every time. The structure has six parts: a standards document setting down tone and d
A note that explains an agent as several chat windows with different roles strung together. If you ask for planning, design and development all in one window, context that has nothing to do with the job at hand gets mixe
ASUS, Dell, HP, Lenovo, Microsoft and others are taking part in a staggered rollout. The launch put AI run on the device front and center alongside creative work and gaming, but compute performance (TOPS) was not stated
Company-reported, it covers 34 language and script combinations with 2.601 trillion tokens in native scripts plus 224.6 billion in Latin script, and includes 438.4 billion restored tokens kept only where the quality scor
It is the first model from a new US lab. Company-reported scores were SWEBench Verified 81.6%, GPQA Diamond 90.6%, AIME 2026 97.7% and Terminal Bench v2.1 80.1% (on BF16), and it ships in three versions, BF16, FP8 and NV
A record in which someone who pushed work automation forward on their own for a year, with no outside help, lays the process out in three stages. The first stage is collecting: piling five streams — email, calls, meeting
The authors reported that in the setting where the constraints have to be found by the model, even the best configuration averaged only 0.56 points and solved just 14 of the 95 tasks. Even in the 23 tasks where every nec
It put protecting teachers' educational judgment forward as a premise, and, timed to Teachers' Day, also explained how to use ChatGPT's study mode (Study) in class. No figures on the number of users or on effectiveness w
AISLE said that in its first year out of stealth it was assigned more than 500 CVEs across more than 100 codebases, and that it found all 12 vulnerabilities in OpenSSL's January security release and 18 of the 24 in curl'
It starts from the point that looking only at the endpoints lets preference changes from early and late training cancel each other out and hide the parts where the teacher changed consistently. The authors reported that
Company-reported, the 209M model ranks No. 1 in the pretrained zero-shot categories of GIFT-Eval, TIME and fev-bench, and the 28M model beat Chronos-2, which is about four times larger. On an 8-thread server CPU the 5M m
Because these are activated LoRAs, the KV cache built by the base model can be reused for the judgment that follows, and IBM said GPU memory with all three components added stays close to that of the base model alone. It
The authors reported that, trained on 12.3 million question-answer pairs on top of Qwen3.5-4B, it ranked No. 1 among the 8 models compared, with an overall 65.64%, and was No. 1 on 7 of 12 benchmarks. They reported it wa
It shipped two versions, a 16-step distilled one and a 50-step base one, together with FP8 and NVFP4 quantized weights and a DreamBooth LoRA training script. Only a speed figure was given — the distilled version makes a
The authors reported that on OSWorld-v2 the average partial score of the 8 models evaluated rose by 19.6 points and that steps per task fell by 53.0% on average compared with running screen control alone. They released t
A conversation that explains the change in the cell structure of memory chips at a non-specialist's level and weighs what it means. A cell is made of one container that holds electricity and one switch that opens and clo
A how-to that builds a research agent which runs on its own at a set time each week, using nothing but a scheduling feature. It likens building one to hiring a person and splits it into four stages — deciding what to ask
A hands-on run that attaches a reverse-engineering toolkit — one that takes a finished program apart and recovers its structure — to a coding agent. On a request the toolkit runs the specialist analyzers on your behalf a
The signing was timed to Singapore International Cyber Week, and technical exchange with local safety and assurance bodies on evaluation, testing standards and incident response is planned. Company-reported, Singapore ra
In a company-reported sample (Sep 21–27, 2026; about 250,000 US conversations), politics, policy and society topics made up 5.75%, elections and voting 0.51%, conversations asking who to vote for 0.01%, and election-inte
Company-reported figures: an Artificial Analysis speech-to-speech index of 86.9 (GPT-Live-1 83.2, Gemini 3.8 Live 82.6), Full Duplex Bench 97%, and response latency under 1 second. xAI said it is used in Tesla vehicles a
Company-reported, a single Sol agent at 52.1% rose to 61.8% once it worked as a team, and a team of a Sol lead with 16 Luna planners reached 61.8% for $155, about 30% of the cost of an all-Sol team that delegated every s
The authors said the vision track was built from 820 videos and 2,460 questions and the track that also uses sound from 240 videos and 720 questions, all annotated by people and checked several times, and that 50 models