The Lab · 20 current model entries
Names change faster than people can learn them. The Lab keeps official positions, independent observations and Human Bit tests separate, then lets you compare only the fields relevant to your job.
Coverage rule
We include a model when it is currently accessible, materially different from another tier and relevant to a real user decision. The catalogue order is not a rank.
OpenAI · Anthropic · Google · xAI · Moonshot AI · DeepSeekFilter the bench
Start with the work surface or provider you can actually use. Select up to three entries to compare their stated fit, limits, evidence coverage and review dates side by side.
A fast ChatGPT model for routine questions, drafting, rewriting and low-consequence transformations. It is no longer the everyday default; GPT-5.6 Luna now fills that role for Free and Go users.
OpenAI’s main reasoning model for complex knowledge work, research, coding and tasks that need more deliberate analysis.
The highest-capability GPT-5.6 option for difficult and longer-running work where a standard reasoning pass is not enough.
OpenAI’s new flagship for research, coding, computer-use and long-running agentic work. It began rolling out on 3 September 2026, first to a limited set of organisations and then to ChatGPT Plus, Pro, Business and Enterprise, the OpenAI API and AWS.
Anthropic’s current Sonnet balance for everyday professional work, coding and agentic tasks, with adaptive thinking by default.
Anthropic’s current model for complex agentic coding and enterprise work, and the model it points Opus 4.8 users at. Thinking is on by default, so the effort setting, not a thinking switch, is how you control depth and cost.
A high-capability Claude model for complex agentic coding and enterprise work, now superseded by Claude Opus 5 but still available on every platform that carried it.
Anthropic's prior most capable widely released model for demanding reasoning and long-horizon agentic work, now superseded by Fable 5.1 but still listed as a legacy model with unchanged pricing and limits.
Anthropic's current model for demanding reasoning and long-horizon agentic work, at the same headline price as Fable 5 but with prompt-cache reads cut to a quarter of the cost, and stronger long-running coding, research and document work.
Google’s earlier fast Flash model, still stable and deployed broadly across consumer, developer and enterprise products, but no longer the current Flash: Google now points new work at 3.6, 3.7 or 3.8 Flash.
A preview model for precise multi-step reasoning, software engineering and tool use across text, images, video, audio and PDFs.
Google’s lower-latency, lower-cost model for high-volume extraction and lightweight multimodal jobs.
xAI’s previous flagship for coding, agentic tasks and knowledge work. It remains available, but xAI released Grok 4.6 on 12 August 2026 and now recommends that as its most capable model; The Human Bit's lab now carries Grok 4.6 as its own entry.
xAI's current flagship, recommended for coding and chat, building on Grok 4.5 with more emphasis on long-running agents and interactive or visual work.
Kimi's open-weight multimodal flagship for long-horizon coding, knowledge work and tool-using tasks, with a one-million-token context window.
Kimi’s dedicated coding model for long-context, tool-using software work, with a separate high-speed endpoint for the same model.
Google's fast prior-generation model for agentic coding, code generation and spatial reasoning, now sharing the Flash tier with Gemini 3.7 Flash, which Google names as the latest and most capable Flash model.
Google's prior fastest Flash model for complex agentic tasks at scale, now sharing the Flash tier with Gemini 3.8 Flash, which Google names as the latest and most capable Flash model.
Google's current fastest Flash model for complex agentic tasks at scale, positioned as the successor to 3.7 Flash with stronger long-horizon software engineering and multi-step reasoning, while currently sharing the same published rate.
DeepSeek's V4 API generation, offered as a cheap high-volume Flash model and a costlier, more capable Pro model, both with a million-token context window.
Each entry needs an official record, a practical decision, an evidence state and a review owner. Missing independent evidence remains visible rather than being filled with a guess.