SearchBook a call

THE BULLETIN / EDITION OF

Model reviews.

What each current AI model is good at, where it falls short, and which jobs it suits. Every fact links to its source.

14 reviews

Anthropic

Claude Opus 5.5

The strongest all-round model in independent tests this month, if you can pay for its tokens and long runs.

Good at

  • Agentic coding and web development
  • Long, multi-step knowledge work
  • Scientific and maths reasoning

Bad at

  • Cost per task
  • Speed and output length
  • No audio input or image output

OpenAI

GPT-6.1 Sol

Strong value among frontier models on early numbers, but wait for more independent results before switching critical work.

Good at

  • Low cost per task at high quality
  • Coding and computer use
  • Very long context

Bad at

  • Very little independent data yet
  • Surcharge on prompts over 272K tokens
  • No audio, realtime or fine-tuning

Google

Gemini 3.8 Flash

A fast, cheap multimodal model, a clear step below the leaders on hard agentic work.

Good at

  • Speed
  • Video, audio and PDF input
  • Science questions and reasoning puzzles

Bad at

  • Long agentic tasks versus the leaders
  • Price doubles in January 2027
  • No audio or image output

xAI

Grok 4.7

A solid second-tier coding model at a low output price, not a leader on any independent board we read.

Good at

  • Low output price for its tier
  • Coding in Cursor and Grok Build
  • Mid-length context of 500K tokens

Bad at

  • Trails leaders on independent indexes
  • No Batch API
  • Fast variant not on the public API

Meta

Muse Spark 1.3

A low-priced closed model with solid independent scores, a young developer platform and no open weights yet.

Good at

  • Quality for the price
  • Output speed
  • Image and video understanding

Bad at

  • Young API and tooling
  • No published parameter count or detailed benchmark table
  • Behind the leaders on maths

Mistral AI

Mistral Medium 3.5

Choose it for control and European hosting, not for top-end quality.

Good at

  • Open weights you can host yourself
  • European vendor and hosting
  • Function calling and structured output

Bad at

  • Low independent index scores
  • Priced high for its measured quality
  • Licence exception for large companies

DeepSeek

DeepSeek-V4.1-Flash

Very low prices and open weights, if you can accept Chinese data storage on the API.

Good at

  • Very low price per task
  • Output speed
  • Open weights under MIT

Bad at

  • Factual recall versus DeepSeek-V4-Pro
  • Hard agentic work versus leaders
  • API data stored in China

Alibaba Cloud (Qwen)

Qwen3.8-Max

Strong for images and web front ends; for agents and value, its open siblings often do better.

Good at

  • Image understanding
  • Web development
  • A broad open-weight family

Bad at

  • Slow output
  • High cost per task
  • Agentic benchmarks

Moonshot AI

Kimi K3

A top open-weight model for coding and research, but too big to host cheaply and costly through the API.

Good at

  • Web development and coding
  • Text, image and video input
  • Long context of 1M tokens

Bad at

  • Output price close to closed frontier models
  • Slow output and long run times
  • Very large to self-host

Google

Gemma 4 (E2B, E4B, 12B Unified, 26B A4B, 31B)

The default open family for on-device and single-GPU work, not a substitute for a frontier model.

Good at

  • Running on phones, laptops and one GPU
  • Permissive Apache 2.0 licence
  • Image input on every size, audio on small sizes

Bad at

  • Far below frontier models on independent indexes
  • Old training data (January 2025)
  • No audio on the two largest sizes

OpenAI

gpt-oss-120b

A dependable, fast, permissively licensed local model that newer open models now beat on quality.

Good at

  • Fits on one 80GB GPU
  • Fast output
  • Apache 2.0 licence

Bad at

  • Knowledge cutoff of June 2024
  • Text only
  • Low independent index scores

IBM

Granite 4.2 (3B, 8B, 30B)

A sensible, well-licensed small model for local agents and tools, with thin independent evidence so far.

Good at

  • Small sizes that run locally
  • Switchable reasoning (full, low, off)
  • Tool calling for enterprise agents

Bad at

  • No Indian languages in the supported list
  • Little independent evaluation
  • Card warns of inaccurate or biased output

Sarvam AI

Sarvam 30B and Sarvam 105B

The first open model family to try for Indian languages, pending independent evaluation.

Good at

  • All 22 scheduled Indian languages
  • Small active size for cheap serving (30B)
  • Apache 2.0 licence

Bad at

  • No independent evaluation we could find
  • Short context for the 30B
  • Modest agentic coding scores

Xiaomi

MiMo-V2.6-Pro

The top open-weight model on Artificial Analysis this month, if your users work in English or Chinese.

Good at

  • Top open-weight score on Artificial Analysis
  • Very low cost per task
  • Text, image, video and audio input

Bad at

  • English and Chinese only
  • Slow output
  • One week old, little track record