State of Frontier LLMs — Syllabus

This is a living catalog of frontier large language model releases. It's for practitioners, builders, and decision-makers who already use LLMs and need to stay current: what just shipped, how it actually performs, and whether it's worth switching to. You don't need an ML research background — the foundations module teaches the benchmark literacy you need to read any release critically. By the end you'll read a model launch the way an analyst does, not the way a press release wants you to.

What you'll be able to do

After the foundations module you'll be able to decode a model release on sight: what the headline numbers mean (Artificial Analysis Intelligence Index, GPQA Diamond, SWE-bench, GDPval-AA, agentic and coding evals), how to weigh context window, latency, and per-token pricing against each other, and what "open weights" actually buys you versus a closed API. You'll spot the common traps — benchmark contamination, cherry-picked evals, ratios flattered by low baselines, "preview" passed off as generally available — and place any new model in both its lab's lineage and the broader competitive frontier.

Three skills the 2026 frontier has made essential, and which the model lessons drill throughout:

• Cost per task, not cost per token. A lab can raise your bill without changing its price list — Gemini 3.8 Flash held its exact per-token rate while cost-per-task rose ~40%, because the model reasons longer and reasoning tokens bill as output. Conversely Grok 4.6's cost advantage far exceeds its sticker discount because it finishes tasks in fewer turns.
• Reading the effort curve. "Is it a reasoning model?" is obsolete. Frontier models now expose a reasoning-effort dial spanning ~4× in cost, and the curve often flattens at the top — GPT-6 Astra scores identically at xhigh and max. Picking the effort level now matters more than picking the tier.
• Verifying licences per checkpoint. Openness is a per-model, per-version property, never a lab property. Four labs tightened flagship licences in a single quarter while keeping permissive terms on smaller models.

Each lab module then turns that skill on real releases. You'll be able to explain, for any model in the catalog, what changed from its predecessor, who it's for, how it compares to its peers, and where it falls short. For the open-weights models you'll also see a runnable implementation of a core idea from the lab's own technical report, so the architecture isn't just a name.

Prerequisites

Practical familiarity with using an LLM (you've prompted ChatGPT, Claude, Gemini, or a local model and have a rough sense of tokens and context). No machine-learning math, no training experience, and no other LearnOS course is required. The single hard prerequisite inside this course is the foundations lesson "How to read a model release" — every model lesson assumes it.

How this course is different — a living catalog

Most courses teach a fixed body of knowledge. This one is a feed. Every notable release earns a lesson, written to a single fixed template so the catalog stays scannable, and a weekly auto-update routine proposes new entries as labs ship. The orientation and foundations modules are stable; the lab modules grow indefinitely. That means the roadmap below is a snapshot — expect new lessons (and eventually new lab modules) to appear over time.

How fast? Between this syllabus's last two reviews, six new flagships entered the catalog and three of them — Claude Fable 5.1, Gemini 3.8 Flash, and GPT-6 Astra — shipped within seventy-two hours of each other on September 1–3, 2026. Two lab-level facts also changed: xAI ceased to exist as a standalone company (Grok now ships under SpaceXAI), and Meta completed its pivot away from open weights at the frontier.

The roadmap

Foundations unlocks everything. The lab modules are independent of one another — once you can read a release, jump to whichever lab you care about, in any order. Within a lab, lessons run oldest to newest so each reads as a lineage.

Time commitment

The stable core — this syllabus plus the two foundations lessons — is about an hour and is the only required reading. After that the catalog is reference-style: each model lesson is a focused 15–25 minutes, and you read the ones relevant to you. A reasonable cadence is the foundations module in one sitting, then one lab module per week, which also matches how often the catalog updates.

How to study in LearnOS

Treat the model lessons as active reference, not passive reading. The strongest evidence in learning science is that retrieving information yourself beats re-reading it — the testing effect (Roediger & Karpicke 2006). So: after reading a release, use chat-with-lesson to quiz yourself ("how does this model's coding score compare to its predecessor, and why?"), turn the load-bearing facts into flashcards for spaced review, and for the open-weights lessons actually run and tweak the code block rather than just reading it. Use ⌘K search to compare across labs ("which models ship a 1M-token context window?", "which flagships changed licence this quarter?").

The capstone

Your capstone is a one-page model-selection brief for a real decision you face: pick a use case (a coding agent, a high-volume classification job, an on-prem deployment), then use the catalog to justify a model choice — citing the specific benchmarks, the cost-per-task math (not just per-token rates), the context-window and licensing constraints, the reasoning-effort level you'd run it at, and the credible alternatives you rejected. The deliverable is the analyst skill the whole course builds toward, applied once to something you care about.

What this course omits

This is a course about releases and how to read them, not a model-training course: it does not teach you to pretrain, fine-tune, or RLHF a model, and it is not a prompt-engineering course. It tracks frontier and notable open-weights text/multimodal LLMs — it does not attempt to catalog every fine-tune, image/video/audio generation model, or embedding model. And because the frontier moves weekly, any specific benchmark number is a snapshot accurate as of that lesson's last_reviewed date; the method for reading the numbers is the durable part.

Sources

• Artificial Analysis — LLM Leaderboard & Methodology (2026)
• Roediger & Karpicke — Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention (2006)