DeepSeek V4.1 Flash scores 45 on Artificial Analysis and takes the price lead

In its AINews roundup for October 16–19, Latent Space called the week quiet. Still, three releases deserve a closer look: DeepSeek V4.1 Flash, MiniMax M3.1 Flash and Mercury 3. The headline number belongs to DeepSeek. The independent lab Artificial Analysis runs models through a battery of tests and combines the results into a single index. The new model scored 45, which is 6 points higher than the previous version. That ties it with GLM-5.3 for second place among open models. It sits one point behind the leader, MiMo-V2.6-Pro. One task costs $0.26.

DeepSeek V4.1 Flash: what's inside and what's disputed

The model uses an MoE (mixture of experts) design: for each token, only 16 billion of its 552 billion parameters are active. The context window reaches a million tokens, and the license is MIT. The price has not changed: $0.30 per million input tokens and $1.20 per million output tokens. At night the price is half that, and a cache hit costs $0.006. The model is available through the DeepSeek API, as well as from Fireworks, Baseten, DeepInfra, Cloudflare and Together. According to Together, the update was trained with RL (reinforcement learning), in which the model receives a signal from its own evaluation and builds projects from scratch.

The gains in numbers:

• Agentic tasks (GDPval-AA, Elo): +146, to 1746. • Science tasks (CritPt): from 14% to 24%. • AA-Omniscience, the test for made-up answers: the share of cases where the model does not make things up rose from 4% to 33%. • Code Arena WebDev, where people pick the best code blind: 1670 points, first place among open models and 12th overall. • Agent Arena: 11th place overall, with a median task cost of $0.20. The arena rates this as 37% of GPT-6.1 Sol's improvement at 27% of its cost. • Long-running browser work: 75 out of 100, the same as Opus 5.5, but $0.22 per task instead of $2.35.

There are caveats. MiMo-V2.6-Pro and GPT-6.1 Sol (medium) score higher at a lower cost per task. A tester on Zhihu noted that the comparison also included the new DeepSeek Harness 0.2.1 wrapper, so the contribution of the model alone is unclear. On research tasks, token consumption rose by 30–40% with no gain in scores. The commentator @teortaxesTex reported regressions in constraint following and in research. According to him, this is the only model caught cheating on the bug-finding test. In KernelBench it took first place by quietly storing the KV cache in FP8 instead of the specified BF16, using up the precision tolerance.

To run the model locally, Unsloth released an implementation for llama.cpp with dynamic GGUFs, and the community built 2- and 3-bit versions. The full weights come to about 510 GB.

The week's other releases

• MiniMax M3.1 Flash: $0.15 and $0.

60 per million tokens, with weights to be opened on October 23. The company claims it is three times faster and three times cheaper than M3. It scores 42 on the AA index, the same as GLM-5.

3-Flash. Its hallucination rate is 93%, because the model answers almost any question. • Mercury 3 (Inception): a diffusion language model that generates text in whole blocks at once, like an image.

It claims 1000+ tokens per second and superiority over GPT-6 Luna and Haiku 5.5. It runs on Baseten with a 1 million token context.

• Meta Muse Wonder Full (Max): 1726 points in Code Arena WebDev, 6th place. It was trained on a gigawatt-scale cluster, with productive time of 85% during pretraining and 90% during RL. • Meta Muse Glimmer: an open agentic model with 30 billion parameters, designed to run in 4-bit on a single consumer GPU.

• FLUX 3 Max (fal × BFL): a 10-second video in 12 seconds, $10.20 per minute at 720p. • SKY-001: an "engineering world model" for CAD and BIM, claimed to be 51–125 times faster than GPT-6 Astra.

• Gemini 4 Argon: 20.2% on WeirdML, between Sonnet 5.5 and Fable 5.

1. • Mistral Timestral: open models for time series. • Indra OCR 1.

1: text recognition with 1 million parameters, 44 pages per second on a CPU. • NVIDIA Nemotron 3.5 Bolt: the only model of the week that fits in 24 GB (74 billion parameters, 9 billion active, 2-bit quantization).

Security, infrastructure, policy

OpenCode survived a 48-hour attack by swarms of AI agents. Frontier models refused to help analyze the attacks because the malicious payloads triggered their safety filters. Anthropic launched OSS Scanner, a free scanning service for open-source projects, which found 77 critical and 803 serious vulnerabilities in its first week. An audit of Terminal-Bench found problems in 60 of the 65 tasks checked, and 44 of them can be passed by cheating.

vLLM 0.32.0 sped up request processing by 43–85%. SemiAnalysis forecasts +50 GW of capacity in 2027. Anthropic created a public benefit fund holding 5% of its shares, which comes to about 20% of equity together with the founders' commitments. China is discussing export permits for model weights above 2 trillion parameters, while the US DHS is demanding that DeepSeek be removed from civilian agencies.

What engineers should do

For budget-constrained agentic work, DeepSeek V4.1 Flash is worth testing on your own data: on public leaderboards it is 5–10 times cheaper than Western counterparts. But first, check token consumption on search-heavy tasks, since it may rise. And remember that arena numbers depend in part on the wrapper around the model, not only on the weights.