Decision Models Crowd In as Cloudflare, Amazon and Perplexity Open Up

Friday, 2 October 2026 · Compiled 18:45 IST

Cloudflare, Amazon and Perplexity all released open decision models within a day. Claude Code mods, Copilot desktop control and new Senate liability proposals add to a crowded news cycle.

1. Cloudflare releases open Clef decision models and an RL fine-tuning product

Reporting Cloudflare released Clef and Clef-flash, decision models hosted on Workers AI and compatible with the Jev API. It says Clef leads the Jev Decision Index and is open-sourcing both models on Hugging Face under Apache 2.0. It also introduced a reinforcement-learning product for customers to fine-tune Clef for their own decisions.

Analysis The first big infrastructure vendor to pair open weights with a tuning service in this new category. The index is Jev's own benchmark and the ranking is Cloudflare's report of it, so it needs independent replication.

Sources: Cloudflare · 1 Oct 2026 · Primary release

2. Amazon open-sources Strands Decider 2B for agent workflows

Reporting AWS's Strands Agents team released Strands Decider 2B, a small open-source decision model for local development and fast experimentation. The team says such models pick among fixed options, return reliability scores and run at very low latency, but are weaker than reasoning models on complex problems and cannot write text. TechCrunch described it as Amazon's answer to Jev.

Analysis A second hyperscaler-backed entrant in a week, aimed at routing and approval steps inside agent loops. The team itself states the limits: it is a component for decisions, not a general model.

Sources: Strands Agents / TechCrunch · 1 Oct 2026 · Primary announcement · Independent report

3. Perplexity publishes pplx-decider-v1-27b, a decision model built on Qwen

Reporting Perplexity posted pplx-decider-v1-27b on Hugging Face, fine-tuned from Qwen3.8-27B with a calibrated-probability API. Its model card reports an 85.71% overall score across 11 benchmarks versus 84.51% for Jev and 74.76% for the base Qwen model, measured through Perplexity's API. Jev scores higher on several individual tasks, including WinoGrande and BBH.

Analysis Perplexity's evidence shows a narrow overall edge and mixed task results, so the claim is not a clear win. It needs about 49 GiB of weight memory, much larger than Amazon's 2B model. License terms were not checked.

Sources: Perplexity model card · 1–2 Oct 2026 · Model card and benchmark table

4. Claude Code adds mods, TypeScript functions that change how the tool works

Reporting Anthropic introduced mods for Claude Code in the CLI and desktop app. A mod can rewrite prompts, block or retry tool calls, approve permission requests, redact secrets from output and change the interface. Mods ship inside plugins. Anthropic warns they run with the same machine access as Claude Code and are not sandboxed.

Analysis A move from fixed hooks to deeper extension, which helps teams shape agent behavior and also widens the supply-chain risk of plugin sharing. Permission-approving mods deserve the same scrutiny as any code that runs locally.

Sources: Anthropic · 1 Oct 2026 · Primary announcement

5. GitHub Copilot can now operate desktop apps in public preview

Reporting GitHub released computer use for Copilot CLI and the Copilot app on macOS and Windows. Copilot can read accessible app content, click controls, type, scroll and drag across applications, including software with no API or MCP integration. It asks approval before controlling an app and organizations can disable the feature.

Analysis Computer use is moving from model demos into mainstream developer tooling, extending coding agents to legacy GUI workflows. Preview status and approval prompts matter, because a GUI agent can take wrong actions in real applications.

Sources: GitHub Changelog · 1 Oct 2026 · Primary changelog

6. Lovable apps can publish into Microsoft tenants through Copilot Managed Runtime

Reporting Lovable launched a public-preview integration that publishes apps built in Lovable into a company's Microsoft tenant, using work logins and existing IT controls. Cloud Wars reports Microsoft introduced Copilot Managed Runtime on 25 September to host AI-generated apps under enterprise governance.

Analysis A practical answer to where vibe-coded apps run once they matter to a company. It is a preview release, and hosting inside a tenant does not make generated code secure or correct.

Sources: Lovable / Cloud Wars · 30 Sep–1 Oct 2026 · Lovable update · Independent coverage

7. Hawley and Murphy plan an AI Agent Accountability Act as Thune eyes codifying safeguards

Reporting Axios reports Senators Josh Hawley and Chris Murphy plan a bill to make companies criminally and civilly liable when an AI agent hacks a system. Separately, Majority Leader John Thune said he is discussing a legislative framework to codify some safety protections after the White House accord. Administration officials say existing law can address AI harms.

Analysis Congress is testing whether to turn voluntary commitments into law. The bill is planned, not introduced, and Thune described conversations rather than a bill, so passage is uncertain.

Sources: Axios · 30 Sep–1 Oct 2026 · Liability bill report · Thune comments

8. OpenAI reportedly seeks $30 billion as SoftBank completes its $30 billion commitment

Reporting Bloomberg and The Information, via The Next Web, report OpenAI is in early talks to raise at least $30 billion at about a $1.4 trillion valuation before the new money, as a bridge before an IPO expected in 2027. No term sheet has been signed. SoftBank said it executed a final $10 billion tranche on 1 October, bringing its cumulative investment to $64.6 billion and about 13% ownership.

Analysis The SoftBank payment is confirmed by the investor; the new round is only reported talks and terms could change. Together they show frontier labs depending on very large private capital while IPO plans move.

Sources: SoftBank / The Next Web · 1 Oct 2026 · SoftBank filing · Round report

9. Tencent reportedly leases about 100,000 AI chips from Oracle in Southeast Asia

Reporting TrendForce, citing Reuters and the Financial Times, reports Tencent signed a five-year, roughly $7 billion lease for about 100,000 advanced AI chips across Oracle data centers in Southeast Asia, with about 30% paid upfront. It notes US rules do not currently bar Chinese firms from leasing capacity overseas.

Analysis It shows how export limits on buying chips are being met through leasing. The deal is reported, not confirmed by the companies here, and rules could change.

Sources: TrendForce / Reuters / FT · 1 Oct 2026 · Report

10. Microsoft AI launches MAI-Transcribe-2-Streaming and MAI-Voice-2.1 models

Reporting Microsoft AI released MAI-Transcribe-2-Streaming, a real-time transcription model for 60 languages with first partial results in just over 100 milliseconds, plus MAI-Voice-2.1 and a faster Flash version. Microsoft says the transcription model ranks first for accuracy on Artificial Analysis and sits on its accuracy-versus-latency frontier.

Analysis Another in-house Microsoft model line aimed at voice agents. The ranking is Microsoft's account of a third-party leaderboard and rankings can change quickly.

Sources: Microsoft AI · 1 Oct 2026 · Primary announcement