AI Technology RadarAI Technology Radar

Model News H1 2026

knowledge
Adopt

The most relevant foundation-model releases of the first half of 2026.

Trends - what moved

  • Extreme release cadence, smaller jumps. OpenAI went GPT-5.2→5.5 and Anthropic Opus 4.6→4.8 plus Sonnet 5 within six months, with fast deprecation of old versions. For engineering teams this shifts the effort from "which model" to model pinning, regression evals and planned migration paths (see Model Discovery & Selection).
  • From chatbot to agent teams. Agent Teams (Opus 4.6), Dynamic Workflows (Opus 4.8) and software-operation capabilities (GPT-5.5) shift the focus to orchestrated knowledge work. 1M-token context became standard across vendors through more efficient attention mechanisms.
  • Open weights reached the frontier - and roles reversed. DeepSeek V4, Qwen 3.5 and Mistral Medium 3.5 deliver near-frontier quality at fractional cost (DeepSeek V4: ~80% SWE-bench Verified as MIT open weights). Meanwhile Meta of all companies pivoted to closed source with Muse Spark and pushed Llama 5 to 2027.
  • The mid-tier got frontier-grade. Sonnet 5 and Gemini 3.5 Flash deliver near-Opus/near-Pro performance at workhorse prices - for most agent workloads the mid-tier is now the rational default.

Releases by player

  • Anthropic: Claude Opus 4.6 (Feb, 1M context beta, Agent Teams), Sonnet 4.6 (Feb), Opus 4.7 (Apr), Opus 4.8 (May, Dynamic Workflows), Sonnet 5 (Jun)
  • OpenAI: GPT-5.4 (Mar, consolidated frontier lineup), GPT-5.5 (Apr, agentic work across documents/spreadsheets/software, 1M context)
  • Google: Gemini 3.1 Pro (Feb, ~doubled reasoning, unchanged pricing), Gemini 3.5 Flash (May, beats 3.1 Pro on coding/agents at 4x speed); Gemini 3.5 Pro still pending as of July - blog.google
  • Meta: Muse Spark (Apr) - first proprietary, API-only frontier model; strategic departure from the open-weight Llama course

Notable open-weight releases

  • DeepSeek V4 Preview (DeepSeek, China) - V4-Pro (1.6T/49B active) and V4-Flash, MIT license, 1M context, ~80% SWE-bench Verified at a fraction of frontier prices; final release mid-July 2026
  • Qwen 3.5 (Alibaba, China) - 397B-A17B, natively multimodal, 201 languages; Medium/Small series followed
  • Mistral Small 4 / Medium 3.5 (Mistral, France) - Small 4: multimodal instruct/reasoning/coding hybrid with 256k context; Medium 3.5: 128B dense, ~78% SWE-bench Verified, self-hostable on 4 GPUs