News

OpenAI switched off GPT-4, o1, o3-mini and more OpenAI

OpenAI's API stopped serving older models on 23 Oct, including GPT-3.5 Turbo, GPT-4, GPT-4 Turbo, o1, o1-pro, o3-mini, o4-mini, GPT-4.1 nano and gpt-image-1. Calls to them now fail.

Source: OpenAI's deprecations page, read 25 Oct

MiniMax M3.1 Flash weights are out MiniMax

MiniMax published the weights for its cheap agent model. Free for non-commercial use; businesses must show a 'Built with' notice, and firms over $20M a year in revenue need written permission.

Best for

Cheap coding and office agents you can host yourself

You give up

Custom license with limits on business use; very large to self-host

Source: MiniMax's post, TechNode, read 25 Oct

Kimi K3.1 Pro Moonshot AI

Moonshot's new top model, cheaper than Kimi K3 at $2 in and $10 out per million tokens, with limited free use on the web. Open weights are promised by 4 Nov.

Best for

Long coding agents and office tasks at a lower price than frontier models

You give up

Weights are only a promise until 4 Nov; most test results are Kimi's own

Source: Moonshot's post, Kimi price page, read 25 Oct

OpenAI Decisions API is generally available OpenAI

OpenAI's endpoint for yes/no, choice and score answers left beta. It runs on GPT-6 Luna and now takes up to 400 questions per request, up from 200.

Best for

High-volume routing, scoring and classification steps inside agents

You give up

Tied to GPT-6 Luna; gives decisions only, never free text

Source: OpenAI API changelog, read 25 Oct

Canva Craft Canva

Canva's first image model, for every Canva user in Dream Lab. It plans layout and text before drawing. Quality claims are Canva's own so far.

Best for

Canva users making posters, slides and social posts with text in the image

You give up

Only inside Canva or its API; quality claims are Canva's own so far

Source: Canva's announcement, TechCrunch, read 25 Oct

1Password for Agents 1Password

Lets AI agents sign in for you after you approve on your own device, without the agent seeing the password. Most named agents, such as Claude Duet, are listed as coming soon.

Source: 1Password's announcement, TechCrunch, read 25 Oct

OlmoEarth v1.3 Ai2

Open satellite embeddings for the whole planet. Useful for map and crop research, but outside the everyday jobs we cover.

Source: Ai2's OlmoEarth page, SiliconANGLE, read 25 Oct

Claude API version 2026-10-21 breaks old code Anthropic

If you build on the Claude API: the new API version and SDK majors reject temperature, top_p and top_k, and send keys as a Bearer header. Read the migration guide before upgrading.

Source: Anthropic's migration guide, Claude release notes, read 25 Oct

Nemotron 3.5 Embed 8B NVIDIA

Open embedding model, free for commercial use, that searches text and page images such as tables and scans together. A 1B version and a reranker came out alongside it.

Best for

Search over scanned documents, tables and charts in your own retrieval setup

You give up

Needs a large GPU; quality claims are NVIDIA's own so far

Source: NVIDIA's post, Model card, read 25 Oct

Flint Recursive

Free open coding model (Apache-2.0) that runs on a laptop. Speed and test figures are the maker's own; Wired's hands-on was positive.

Best for

Local coding agents on a 48GB Mac or GPU with no API fees

You give up

Needs smaller and clearer tasks than paid assistants

Source: Recursive's post, TechCrunch, read 22 Oct

mammoth-1 and elephant-1 Baseten

New open "decision models". No named license yet, and the rankings come from Baseten without saying who ran them. We look again when an outside board lists them.

Source: Baseten's post, read 22 Oct

Kumo Relational NVIDIA

Real and open, but a prediction tool for company databases, outside the jobs we cover. The only test results are the authors' own.

Source: Hugging Face blog, read 22 Oct

Voxtral 3 Mistral AI

Mistral's new speech family: transcription, text to speech and a voice agent kit from one EU-hosted vendor.

Best for

Voice agents that want listening and speaking from one EU vendor

You give up

Weights and self-host license not public

Source: Mistral's post, The Decoder, read 21 Oct

Eleven Realtime v1 ElevenLabs

ElevenLabs' first speech to speech models for live voice agents, in 90+ languages. $0.08 a minute plus model tokens.

Best for

Live voice agents that need many languages and voices

You give up

Per-minute fee plus tokens adds up

Source: ElevenLabs' post, ElevenLabs' price page, read 21 Oct

Ming-3.0 TTS Ant Group

Open text to speech under MIT, with voice cloning and tags for laughs and sighs. Free for commercial use.

Best for

Self-hosted voice with cloning, free for commercial use

You give up

Ranks below Fish Audio Drama 3; no memory figure

Source: Model card, TechNode, read 21 Oct

DeepInfer DeepSeek

Big news, but it runs one DeepSeek model on two chip types. vLLM still does more for most teams.

Source: GitHub, Bloomberg, read 21 Oct

TorchTPU Google

The code went public, but it was announced in April and only helps teams already on Google Cloud TPUs.

Source: GitHub, read 21 Oct

MiniMax M3.1 Flash MiniMax

A cheap fast model ($0.15 in, $0.60 out per million tokens) with a free version on OpenRouter. Not on a public board yet, so no score.

Best for

High-volume simple tasks on a tight budget

You give up

VentureBeat reports a high rate of made-up answers

Source: MiniMax's post, VentureBeat, read 20 Oct

Mercury 3 Inception

Cheap and fast on paper ($0.20 in, $0.70 out per million tokens, half price at launch), but every quality claim is Inception's own. We look again when an outside board lists it.

Source: Inception's post, read 20 Oct

FLUX 3 Max fal

Fast text and image to video built on FLUX 3 Video, at a launch discount until 7 Nov. Speed claims are fal's own, and only one major outlet covered it.

Source: fal's page, VentureBeat, read 20 Oct

BAND desktop BAND

Launched 15 Oct with no price, no access details and no outside tests, so there is nothing yet to judge it on.

Source: BAND docs, read 20 Oct

Gemini 4 Argon opens up Google

Google's top model is now on the paid Gemini API and in Google AI Plus from $4.99 a month. It leads Arena's text votes; votes are still early.

Best for

Everyday chat and writing near the top of Arena votes, from $4.99 a month

You give up

Behind Claude and GPT for web app coding; API price doubles on 1 Dec

Source: Google's plans page, Gemini API pricing, read 20 Oct