New Models
SWE-2 Flash
Cognition debuts SWE-2 Flash, post-trained on Inkling-Small
A fast, cheap coding model post-trained on Thinking Machines' open Inkling-Small (12B active). It scores 24.0% on Cognition's own FrontierCode 2.0, against 21 for GPT-6 Luna and 27 for Haiku 5.5. Cognition says it costs one fifth of SWE-2 and runs about 3x faster, and it is free in Devin Desktop and CLI through November 30 for Pro, Max and Teams users. Not open weights.
24.0% FrontierCode 2.01/5 the price of SWE-2~3x faster than SWE-2
New ModelsOpen weights
Kimi K3.1 Pro open weights
Moonshot releases the Kimi K3.1 Pro weights
A 2.8T parameter MoE with 104B active, 1M token context and native image input, about 1.56 TB in MXFP4. Moonshot reports 65.7 on Terminal-Bench 4.0, just under Claude Opus 5.5 at 66.4, but its own table shows 31.5 on Program Bench against 65.0 for Opus 5.5. Open weights, not open source in the usual sense: large products must show the Kimi name, and large API resellers need a separate agreement. $4 / $20 per million tokens from Moonshot.
2.8T / 104B total / active parameters65.7 Terminal-Bench 4.0 (Moonshot-reported)$4 / $20 per million input / output tokens
New Models
GPT-6 Cyber
OpenAI launches GPT-6 Cyber for Daybreak Red
OpenAI's cybersecurity model, for the premium tier of its Daybreak program. It scores 48.5% on ExploitGym, up from 32.1% for GPT-5.6 Cyber (GPT-6 Astra scores 42.4%), found 81 new high-severity Firefox bugs, and avoided exploiting out-of-scope services 99.93% of the time without being prompted to. OpenAI says it plans to dramatically expand Daybreak. It broke live in the middle of the TL;DR.
48.5% ExploitGym81 new high-severity Firefox bugs99.93% in scope without being prompted
New Models
Images 2.6
OpenAI rolls out Images 2.6 with GIFs and two API models, Flare and Sunburst
Two new API image models, Flare (fast) and Sunburst (quality, #1 on DesignArena's Image Arena), at the same price as Images 2.5, and OpenAI says 2.6 is 50% faster. In ChatGPT, 2.6 can use saved references and make GIFs, since generation is now fast enough for several frames while you wait. It also adds SVG export.
50% faster than Images 2.5, per OpenAI
New ModelsOpen weights
Pika H3 Flex
Pika open sources H3 Flex, the top open weights model on DesignArena's text-to-video board
Pika released the weights of H3 Flex, which it says is the top open weights model on DesignArena's text-to-video board.
New ModelsOpen weights
Ripple-65B Preview
Ricursive open sources Ripple-65B Preview, a dense 65B model pretrained from scratch
A dense 65B model pretrained from scratch on 12 trillion tokens, with up to 1M context, under Apache 2.0. Ricursive reports 80.5% on TerminalBench 2.1, just ahead of Reflection Beam (501B) at 80.1%, with 1.4M RL rollouts (about 70x fewer than Beam), no bought post-training data and a $32 million total project cost. All scores are self-reported, and its own table shows it behind on DeepSWE and SWE-bench Verified. The base model is released too.
80.5% TerminalBench 2.1 (self-reported)12T pretraining tokens$32M total project cost
New Models
Hy4
Tencent launches Hy4 with vision and computer use
Tencent's first Hy model with vision and computer use, with 1M token context. Tencent says it leads open models on its own finance eval and promises weights later. For now it is on OpenRouter at $0.417 in and $1.251 out per million tokens with a launch discount.
1M token context$0.417 / $1.251 per million input / output tokens on OpenRouter
