The day in AI

Thursday 1 October 2026

←→

Amazon and Cloudflare released open-weight decision models built on Qwen to rival Jev

The day in brief

Amazon Web Services and Cloudflare each released open-weight decision models on 1 October, small models that return a scored choice from a fixed set of answers.

Amazon Web Services (AWS) released Strands Decider 2B, built on the Qwen3.5-2B language model with text generation removed. It picks one answer from a closed set and gives a calibrated confidence for that pick. It follows the design of Jev, the decision model from TypeSafe. It comes out through Strands Labs, a project started by Marc Brooker, an AWS distinguished engineer. AWS says it's meant for agent workflow steps that don't need a full LLM, at lower latency and cost, and it runs locally.

SQ Magazine reports a median decision latency of about 115 milliseconds on a single Nvidia RTX 3090. The same report places it third of 33 models in its size class on JevBench, which scores accuracy and calibration together.

Cloudflare released Clef and Clef-flash on the same day under Apache 2.0. They are the first models trained by its Workers AI team. They answer yes or no, multiple choice and score questions with probabilities. Clef is post-trained from Qwen3.8-27B and Clef-flash from Qwen3.5-9B. Each reads the prompt in one prefill pass and then makes the decision in a single non-autoregressive step, so it doesn't generate an answer token by token. Both accept images, have a 64k context window and work with the Jev API.

Cloudflare reports a median latency of 209.3 milliseconds for Clef and 38.8 milliseconds for Clef-flash across 43 benchmarks. On the same tests it puts Jev at 524.1 milliseconds. Clef costs $0.24 per million tokens on Workers AI, and The Register puts Jev at $0.042 per million tokens. Cloudflare also announced a reinforcement learning (RL) fine-tuning product for Clef.

Black Forest Labs released FLUX 3 Image as a closed API. It takes scene layouts as JSON bounding boxes with element ids, and edits leave pixels outside the box numerically unchanged. Black Forest Labs reports that 67.8 to 89.7% of pixels stayed identical across edits. Output is native at about 4K and it accepts up to ten reference images. It starts at about $0.0205 per 768 pixel image, with an introductory 50% discount until 8 October. An open-weight version is said to be expected within weeks.

The decision model numbers have no independent check yet. Amazon's figures come from a press report and Cloudflare's come from its own testing, so the two have not been compared on the same benchmark.

Also in the news

  • Synopsys and OpenAI announced a multi-year partnership on 30 September to build GPT-Synopsys, a model that operates Synopsys chip design tools, with shared revenue, and some details of how it runs are only partly confirmed.
  • OpenAI added a Try on button to clothing listings in ChatGPT on 1 October, which uses ChatGPT Images 2.5 to show the user wearing the item from a selfie.
  • OpenAI launched dots at DevDay 2026 on 29 September, proactive assistants powered by GPT-6 Astra with their own cloud computer, for Pro and Enterprise users.
  • OpenAI says it parted ways with three safety researchers on 1 October after an investigation into mishandled confidential information.
  • Anthropic released Claude Code mods, small TypeScript functions shipped in plugins that change Claude Code's behaviour and run without a sandbox.
  • Microsoft released MAI-Transcribe-2-Streaming, which it says gives a first transcript in about 100 milliseconds, plus the MAI-Voice-2.1 and Voice-2.1-Flash speech models.
  • Suno opened Suno Speech to all users as a beta, generating spoken voice and original background music as one track.
  • DeepSeek released Ascend versions of its DeepEP and DeepGEMM libraries for Huawei's Ascend 950 chips on 29 September.
  • Nvidia announced the Open Agent Safety Platform on 28 September, pairing the OpenShell runtime with a hardware watchdog called Sentry, according to Tom's Hardware.
  • OpenAI reportedly disrupted a data extraction campaign that it links to Moonshot AI, with related activity reaching more than 15,000 users.
  • Anthropic reportedly said GLM-5.3-Flash built a working exploit chain for $20.40.
  • AMD is reported to be buying World Labs for $8.2 billion.
  • ElevenLabs reportedly doubled its valuation to $22 billion in a $300 million employee share sale.
  • The FTC has reportedly opened an investigation into OpenAI, Anthropic and other AI companies over consumer risks from their products.
  • Armadin, the startup founded by Mandiant founder Kevin Mandia, reportedly raised $255.5 million at a valuation above $2.5 billion.

Everything from this day

Launches and products

1 Oct 2026AmazonStrands Decider 2B Amazon Web Services released Strands Decider 2B, an open source 2B decision model that picks from fixed options and scores its confidence, and runs locally.1 Oct 2026GoogleProject Suncatcher prototype satellite Google launched a prototype satellite with four TPUs into low Earth orbit with Planet, the first in-orbit test of its space-based AI computing research project.1 Oct 2026Black Forest LabsFLUX 3 Image FLUX 3 Image adds JSON bounding-box layouts and edits that leave pixels outside the box numerically unchanged, native ~4K output, up to ten references.1 Oct 2026OpenAIChatGPT virtual try-on OpenAI added a Try on button to clothing and accessory listings in ChatGPT, which generates images of the user wearing the item from an uploaded selfie.1 Oct 2026CloudflareClef and Clef-flash Cloudflare released Clef and Clef-flash, open-weight decision models under Apache 2.0 that return probabilities for typed answers, hosted on Workers AI and Hugging Face.1 Oct 2026TavusGriffin Tavus unveiled Griffin, a Human Interaction Model for real-time video conversation. In the company's test, 48% of participants thought Griffin-Lite was a real person.1 Oct 2026Cantinaapex-flash-1 Cantina released apex-flash-1, an MIT-licensed open-weights security model fine-tuned from GLM-5.3-Flash that solved 40 of 60 held-out vulnerability tasks.1 Oct 2026AnthropicClaude Code mods Claude Code mods are small TypeScript functions that rewrite prompts, add UI or replace built-in features, shipped inside plugins.1 Oct 2026SunoSuno Speech (beta) Suno Speech generates voice and music together as one track, e.g. a dramatic reading or meditation with original background music.1 Oct 2026MicrosoftMAI-Transcribe-2-Streaming, MAI-Voice-2.1 and Voice-2.1-Flash Streaming transcription ranks #1 on Artificial Analysis with first hypothesis in ~100 ms; Voice-2.1-Flash has 150 ms end-to-end latency.

People and teams

Three unnamed OpenAI safety researchersOpenAI → unknown. OpenAI says it parted ways with three safety researchers after an investigation found they mishandled sensitive company information outside established procedures. The Wall Street Journal reported they allegedly shared confidential information with a third-party AI safety organization, which OpenAI did not confirm or detail.
Tomek Korbak, Mikita Balesni, and Jasmine WangOpenAI → unknown. OpenAI terminated three safety researchers after an internal investigation found they mishandled sensitive company information outside established procedures. The Wall Street Journal reported that the researchers allegedly shared confidential material with a third-party AI safety organization.