Frontier model brief · Issue 01
A signal-first view of the newest models from the labs shaping AI—what launched, what changed, and why it matters.
Verified release stream
Only first-party announcements. No composite scores, rumours or leaderboard theatre.
Google’s newest Flash model, tuned for token-efficient coding, knowledge work and multimodal tasks.
Pushes a fast model into long-horizon agents and computer-use workflows without Pro-tier pricing.
A fast frontier model built around coding, agentic execution and professional knowledge work.
Pairs deep software-engineering ability with an advertised serving speed of 80 tokens per second.
A three-tier family: Sol for frontier work, Terra for balance and Luna for fast, efficient tasks.
Brings stronger multi-agent work and programmatic tool coordination across ChatGPT, Codex and the API.
Meta’s multimodal reasoning model for agentic tasks, computer use, coding and rich media understanding.
Marks Meta’s return to frontier APIs with native multi-agent orchestration and active context management.
An embodied navigation model that follows language instructions using a single ordinary RGB camera.
Shows a compact model moving beyond screens into autonomous navigation across real environments.
Anthropic’s fifth-generation Sonnet model for autonomous agents, coding, tool use and knowledge work.
Brings performance close to Opus 4.8 at a lower price, with stronger planning and long-running execution.
An open flagship model designed to hold context and execute over long software-engineering horizons.
Combines a stable million-token window with multiple reasoning-effort levels and openly released weights.
Cohere’s first agentic coding model, built to run locally with a small active compute footprint.
Makes sovereign coding agents practical: 30B total parameters, but only 3B active at inference time.
Anthropic’s fifth-generation frontier model for sustained coding, science, vision and demanding knowledge work.
Introduces a capability tier above Opus for complex, days-long and asynchronous tasks.
A natively multimodal frontier model for coding and agents, built on MiniMax Sparse Attention.
Combines open weights, a million-token context window and long-horizon agentic execution in one model.
A new API generation offered as V4-Pro and V4-Flash, with both thinking and non-thinking modes.
Keeps compatibility with both OpenAI- and Anthropic-style interfaces, lowering the cost of switching.
Reading the radar
“Latest” means the newest named model announcement we could verify from each lab’s own newsroom or documentation. Release dates reflect the original announcement, not later availability updates.
Claims are deliberately attributed to the releasing company. Benchmark results are not normalized across labs, so this index does not rank models or declare a winner.
