Executive summary. Five labs set the pace of 2026's model race. OpenAI shipped GPT-5.4 across ChatGPT, the API and Codex, Anthropic made the newly agentic Claude Sonnet 5 its default model, Mistral and Thinking Machines leaned on efficient hybrid designs, and Meta's Muse Spark marked a new lab's debut.
Agentic capability became the headline feature
The two most widely deployed launches of the year both foregrounded autonomous, tool-using behaviour. OpenAI rolled GPT-5.4 out simultaneously across ChatGPT, its API, where it is served as gpt-5.4, and the Codex coding agent, putting one model behind both consumer chat and developer automation. Anthropic went further on positioning, calling Claude Sonnet 5 its most agentic Sonnet model yet and promoting it to the default for Free and Pro users rather than reserving it for a premium tier. Making an agent-oriented model the no-cost default is a signal in itself, because it shows the labs now expect everyday users, not just engineers, to hand tasks to a model and let it act.
Efficiency and long context, not just raw scale
Away from the flagship chat models, two 2026 releases competed on architecture rather than size. Mistral Small 4 is a 119-billion-parameter hybrid model that keeps only 6.5 billion parameters active per token and pairs that with a 256,000-token context window. Thinking Machines took the same mixture-of-experts logic to a far larger scale with Inkling, a multimodal model carrying 975 billion total parameters but activating just 41 billion, alongside a one-million-token context window. Both designs answer the same production pressure, the cost of running large models at inference time, by separating a model's total knowledge from the compute it spends on any single request.
A new lab, and the open-weights turn
2026 also reshaped who gets to ship a frontier model. Meta introduced Muse Spark as the first model in a new Muse series and the first large language model to come out of Meta Superintelligence Labs, signalling a reorganisation of how the company builds frontier systems. Thinking Machines, meanwhile, made Inkling its first open-weights release, putting a near-trillion-parameter model in the hands of anyone who wants to run or fine-tune it. Between a restructured incumbent and an open-weights challenger, the year widened the set of credible frontier providers beyond the usual short list.
FAQ
Which major AI models launched in 2026?
The most notable 2026 launches were OpenAI's GPT-5.4, Anthropic's Claude Sonnet 5, Mistral Small 4, Meta's Muse Spark, and Thinking Machines' Inkling.
How do Mistral Small 4 and Inkling cut inference cost?
Each uses a mixture-of-experts design that activates only a fraction of its parameters per token: 6.5 billion of 119 billion for Mistral Small 4, and 41 billion of 975 billion for Inkling.
What is notable about Claude Sonnet 5?
Anthropic describes Claude Sonnet 5 as its most agentic Sonnet model to date and made it the default for Free and Pro plans, so most users now reach an agent-capable model without paying extra.
What was the first model from Meta Superintelligence Labs?
Muse Spark, the debut model in Meta's new Muse series, was the first large language model released by Meta Superintelligence Labs.
Sources
OpenAI — openai.com (2026-03-05)
Anthropic — www.anthropic.com (2026-06-30)
Mistral — docs.mistral.ai (2026-03-16)
Meta — about.fb.com (2026-04-08)
Thinking Machines Lab — thinkingmachines.ai (2026-07-15)
