Heyaiwiki.
Models Best Guides Glossary News Tools About
English

DeepSeek Large Language Models

DeepSeek V4-Pro

The open-weight value champion — near-frontier coding (SWE-V ~85%) under an MIT license with a 1M-token context at a fraction of closed-model prices.

Visit DeepSeek All LLMs models

DeepSeek V4-Pro is the open-weight value champion of 2026. A 1.6T-total / 49B-active mixture-of-experts, it posts roughly 85% on SWE-bench Verified — within a few points of the closed frontier — under a permissive MIT license and at a fraction of the price.

Its DeepSeek Sparse Attention (DSA) is what makes a 1M-token context affordable, using a small fraction of single-token FLOPs versus the previous generation at full context. The companion Flash model has since been replaced by DeepSeek V4.1-Flash, which DeepSeek reports beats V4-Pro itself on most agentic rows at $0.30 / $1.20 per million tokens at peak.

The honest caveats: it trails the absolute top on the hardest benchmarks, and some enterprise legal teams remain cautious about geopolitical and export-compliance questions despite the open license.

Update, October 2026

  • GA release, August 13, 2026. The V4-Pro-0813 build rolled out on the app, web and API, with much stronger agent results — DeepSWE v1.1 62.7, Terminal-Bench 2.1 87.9, CyberGym 83.3, Toolathlon Verified 74.1 and HLE 42.7 (60.0 with tools) — plus native OpenAI Responses API support for Codex and three thinking-effort levels (low, high, max).
  • New pricing, from August 16. Peak/off-peak billing replaced the flat rate: $1.32 input (cache miss) and $3.96 output per million tokens at peak, and $0.66 / $1.98 off-peak (half price). Cache-hit input is $0.044 at peak and $0.022 off-peak. Concurrency is limited to 500.
  • V4.1-Flash overtakes it on most agent rows. DeepSeek first said V4-Pro requests would route to the new Flash model on September 14, then decided to keep serving V4-Pro with billing unchanged “in response to user demand”, while a V4.1-Pro is still to come.

Best for

  • High-volume coding and agents on a budget
  • Self-hosted deployments needing 1M context
  • Cost-sensitive RAG pipelines

Pros & cons

Strengths

  • Open weights (MIT) with near-frontier SWE-bench Verified (~85%)
  • 1M context made cheap via DeepSeek Sparse Attention
  • 5–10× lower price than the closed frontier

Limitations

  • A few points behind the very top on the hardest tasks
  • Some enterprise teams flag geopolitical/export-compliance uncertainty

Sources

More LLMs models

Anthropic

Claude Fable 5

Anthropic's first Mythos-class model for general use — state-of-the-art on nearly every benchmark, with safety classifiers that fall back to Opus 4.8 on sensitive topics. Suspended by US export controls June 12–30, back globally since July 1.

$10 / $50 per 1M tokens (input / output)
Anthropic

Claude Fable 5.1

Anthropic's Sept 1, 2026 Mythos-class update — ~25% cheaper than Fable 5 on typical workloads via 75% lower cache reads ($0.25), 60% fewer cyber false positives, Terminal-Bench-Science 52.6% vs 24.7%. Same model as the restricted Mythos 5.1.

$10 / $50 per 1M tokens (input / output)
Anthropic

Claude Haiku 5.5

Anthropic's Oct 7, 2026 small model — from $0.10 / $0.50 per 1M tokens (about 75% cheaper than Haiku 4.5), first Haiku with adjustable effort, 1M context, OSWorld 2.1 offline 72.4%.

$0.10 / $0.50 per 1M tokens for prompts up to 100K (input / output); $0.50 / $2.50 above 100K

Last updated: 2026-10-12. Specs and pricing change fast, so verify on the vendor's site before relying on them.