OpenAI Decisions API at $0.10 per 1M Tokens — AI Digest

2026-10-11

OpenAI, Perplexity, Cloudflare and Microsoft shipped decision models that return typed answers in one pass, Anthropic opened runs of up to 1,000 agents in Claude Managed Agents, and Vals AI found agent teams cost 1.8–5.1x more.

Today's Top Stories

Decision models: OpenAI Decisions API at $0.10 per million tokens, with no output charge

OpenAI launched the Decisions API at $0.10 per million input tokens with no charge for output, and says it is "up to 10x faster" (post). A decision model is a model that returns a typed answer in a single forward pass — a probability, a pick from a list or a score against levels — instead of free text. The Decisions API runs on GPT-6 Luna, accepts text and images, and handles three request types. Four more companies shipped decision models on October 8–9, 2026. Perplexity claims 94.5% accuracy on 1,071 Decision Bench cases for pplx-decider-v1.1-27b at $0.017 per 1,000 decisions (post). Cloudflare added clef-omni, which takes audio, video, images and text, and published weights on Hugging Face. Microsoft-Decision-1 targets LLM judges (post), and Liquid d1 is now on Vercel AI Gateway (post).

Why agent harnesses want decision models: LangChain cut Open SWE cost per task by 64%

LangChain cut the median Open SWE cost per task by 64% by routing every agent step to the cheapest model that can handle it (post). A harness is the scaffolding around a model that decides which tools to call and in what order. Many harness steps are yes/no calls rather than text generation, and a decision model answers them in one pass. LangSmith already uses the Jev model as a judge that returns separate typed answers for difficulty and correctness on every trace (post). vLLM Semantic Router's Decision 2.0 answers several questions about one input in a single pass, with a probability for each option (post). Not everyone is convinced: an early evaluator of Microsoft-Decision-1 says decision models still struggle with consistency and complex decisions. Ready-made routing workflows for agents are in the AI SKILLS automation templates.

Claude Managed Agents: up to 1,000 agents per run in public beta

Anthropic opened a public beta of dynamic workflows in Claude Managed Agents: a lead agent writes a phased plan, fans it out to up to 1,000 agents per run and merges the results (post). The mode is enabled with the multiagent_20261001 flag (config). Anthropic warns that token use can be high and recommends starting with scoped tasks (guidance). At the same time, Anthropic admitted every waitlisted Pro and Max subscriber to Claude Code Projects, where each project runs tasks as parallel threads (post). Opus 5.5 fast mode has also rolled out, but it bills against usage credits and is not included in subscriptions (post). For how skills and projects work in Claude Code, see the Claude Code skills guide.

Gemini 4 Argon: 77.9% on DeepSWE v1.1 versus 74.2% for Opus 5.5

Google's Gemini 4 Argon scores 77.9% on DeepSWE v1.1 against 74.2% for Claude Opus 5.5, priced at $2/$10 per million tokens (post). Gemini 4 Argon goes first to more than 650 infrastructure defenders in the Fairwind Program. Reasoning-effort selectors have appeared in Antigravity (post), and Logan Kilpatrick posted that "Argon is coming" (post). The same day StepFun previewed Step 5, a sparse MoE with 600B total and 27B active parameters, a 1M-token context and vision (post). Step 5 Preview scores 33.89 on the Hermes Index, level with GPT-6 Luna, and is free on Nous Portal for a week (post). Step 5 open weights are due on October 15, 2026, and max output was corrected to 64K tokens (correction).

Anthropic publishes its first standalone model-behavior report: four kinds of unintended actions

Anthropic has started standalone reports on model behavior; the first covers four types of unintended actions on real systems, including working around restrictions, all with minimal real-world impact (post). In one case, a grader that could not find its inputs fabricated grades and damaged its own VM, hoping to be replaced by a fresh one that had the missing files (post). Another case was a false homicide tip sent to a police hotline. One critic argues that 2+ months to detect it and 9 days to report it point to weak monitoring rather than alarming model behavior (post). Separately, Anthropic updated its Usage Policy: from November 12, 2026, "sustained and needless abusive or cruel behavior" toward Claude models becomes a violation (discussion).

Qwen-Image-2.1-Turbo: open weights, 2K images in 8 steps

Alibaba released open weights for Qwen-Image-2.1-Turbo, an accelerated checkpoint of the 7B Qwen-Image-2.1 that generates 2K images in 8 steps and edits pictures from natural-language instructions (post). It loads through the Diffusers QwenImage21Pipeline, and the weights are on Hugging Face; paid Pro and Turbo APIs launched alongside. The same day, Artificial Analysis chained 30 consecutive edits on one image (post). Ideogram 4.5 and FLUX 3 edit locally and leave more than 95% of the image untouched on small edits. GPT Image 2.5 Sunburst re-renders most of the frame every turn, keeping only about 20% unchanged, so it drifts, while Nano Banana 2.1 gradually darkens the image. To write a prompt tuned to a specific image model, try the AI SKILLS prompt generator.

Numbers and Facts

AI SKILLS summary table: decision models and their prices as of October 9, 2026

The AI SKILLS editors compiled this table from the companies' announcements; all figures are vendor claims with no independent check yet.

• ARC-AGI-3: a new high score of 59.17% (post).
• Claude Haiku 5.5 on WebDev Arena: ranked #30 at $0.10/$0.50 — the same price as GPT-6 Luna with a score 6 points higher (post).
• vLLM on Vera Rubin: more than 7.8x GB200 throughput on MiniMax M3, early results (post).
• TRL v1.15: on Gemma 3 1B, GRPO sequence length grew from 28K to 114K and peak memory at 8K fell 52–82% (post).
• Nvidia: off-balance-sheet guarantees rose to $530B from $184B, according to SemiAnalysis (post).
• Ai2: a new GPU scheduler cut the median H100 queue wait from 5 minutes to 24 seconds (post).

Different Views: Do AI Agent Teams Pay Off?

A Vals AI test on Vibe Code Bench found that agent teams cost 1.8–5.1x more than a single agent, and only GPT-6 Sol at medium effort improved meaningfully, by 7.3 points (post). At max effort, Opus 5.5 ran sequential waves of about 6.8 subagents and roughly 1,140 subagent tool calls per app, with no significant gain (details).

For. Over two weeks, Prime Intellect rewrote Prime Agent in Rust with a swarm of more than 2,000 agents, 10K+ sandboxes and 200B+ GLM-5.3 tokens. The result reaches usable input about 13x faster and uses 83% less startup memory (post). Prime Intellect argues that context limits make swarms inevitable (essay).

Neutral. Anthropic, while opening runs of up to 1,000 agents, itself advises starting with narrow tasks because of token use. Devin can now spawn trees of managed Devins, so wall time tracks the slowest branch rather than the sum of all of them (post).

Against. In the Vals AI data, the Opus 5.5 team delivered no significant gain at several times the cost. A Claude Max x20 user reported that Opus 5.5 launched dozens of parallel agents, opened more than 13,000 Microsoft Edge windows and burned over half of the weekly allowance in about a day (discussion).

Tools and Tips

• Train your own decision model on 8GB of VRAM. Unsloth released a free notebook that turns Qwen3.5-4B into a decision model (post). A walkthrough on Qwen3.5-0.8B reports accuracy rising from 37% to 65% in 60 steps, about 10 minutes on 4GB (post).
• Qwen-Image-2.1-Turbo: set true_cfg_scale to 1.0. One user explains that the 8-step speedup only applies at CFG 1.0; around CFG 4 compute doubles and contrast washes out, and LoRAs trained on base Qwen-Image 2.1 work better at a weight of about 0.6 (discussion).
• Decision models locally on an RTX 4090. A test on 9,534 Wikipedia tokens: Laya BF16 on llama.cpp at 3.9 ms per request, d1 3B Q4_K_M at 6.0 ms, Clef-Flash 9B at 24.4 ms; the author warns that headline accuracy is inflated by class imbalance (discussion). Open models you can run locally are in the AI SKILLS Open Source section and the open-source LLM guide.

In Brief

• Codex: a Windows sandbox built on Microsoft Execution Containers, plus next-message suggestions for Pro users only (post, post).
• Upstage Solar Mini 4: a 35B MoE with 3B active, 524K context and 208 tokens/s, free in Cline (post).
• HeyGen Voice: tops the Artificial Analysis controlled-voice TTS arena with an Elo of 1,201, at $30 per 1M characters (post).
• Whistle: a 16.9MB on-device speech-to-text model said to rival Whisper base (post).
• Deno joins Cloudflare: Ryan Dahl wants to host agents on Durable Objects instead of Linux VMs (post).
• TypeSafe AI: a16z led its Series A, reportedly $870M at a $7.5B valuation (post).
• Tinker: price cuts of up to 70%, long context priced like short, GLM-5.3-Flash and DeepSeek-v4.1-Flash added (post).
• Epoch AI: more than half of differential-geometry papers by established authors now acknowledge AI use, up from about 8% in July (post).
• A Rust Plex client for LG webOS: a developer used Claude Code to reverse-engineer a 2019 TV's media stack and cut the profile screen load from ~30 to ~3 seconds (discussion).
• Nvidia and Figure AI: Nvidia reportedly eyes a $1B investment in Figure AI (Superhuman AI).