Small models ate the middle
The capability floor rose faster than the ceiling. Most production inference no longer touches a frontier model.
#models
20 pieces tagged models,
from January 21, 2025 to February 23, 2026.
The open weights year
Twelve months that took open models from interesting to unavoidable, and where the gap actually sits.
Claude Opus 4.5 and the compaction problem
A frontier release with a large price cut, plus effort controls and context compaction as a first-class feature.
Gemini 3 arrives with an IDE attached
Google ships a frontier model and Antigravity, an agent-first development environment. The bundling is the strategy.
GPT-5.1 and the return of the model picker
Instant and Thinking as named modes, adaptive reasoning, and personality controls. The router lesson got learned.
Haiku 4.5 and the collapsing cost of good-enough
A small model at frontier-adjacent coding performance, priced for volume. The economics of agent fleets just changed.
Claude Sonnet 4.5 and the agent that runs for thirty hours
A model tuned for long-horizon autonomous work, plus checkpoints and context editing in the SDK.
GPT-5 lands with a router and a backlash
One model that decides how hard to think, an abrupt deprecation of everything else, and a lesson about attachment.
OpenAI ships open weights for the first time since GPT-2
gpt-oss-120b and gpt-oss-20b under Apache 2.0. A strategic reversal, six years late, and genuinely useful.
Kimi K2 is a trillion-parameter open weights release
Moonshot ships a 1T-parameter MoE with 32B active, tuned for agentic tool use, with weights you can download.
Grok 4 and the benchmark that ate the discourse
xAI claims frontier results with heavy test-time compute. The number that matters is the one nobody quotes.
Gemini 2.5 goes generally available with a thinking dial
Pro and Flash hit GA, Flash-Lite arrives, and every tier exposes a thinking budget.
everything else tagged #models 8
- The Illusion of Thinking, and the argument about what reasoning isnews
- Claude 4 and the agent that runs for hoursnews
- Qwen3 ships a whole family under Apache 2.0news
- o3 and o4-mini put tools inside the reasoning loopnews
- Gemini 2.5 Pro is Google's best model and it showsnews
- GPT-4.5 is the end of an era, politelynews
- Grok 3 and the case for buying your way to the frontiernews
- DeepSeek R1 puts a reasoning model under an MIT licensenews