The LLM Models List, August 2026: The Frontier Lineup and How to Filter It

The list of LLM models in August 2026, in one breath: Claude Opus 5, Claude Fable 5, GPT-5.6 Sol, Kimi K3, Qwen3.8 Max, Claude Sonnet 5, Grok 4.5, GLM 5.2, Gemini 3.6 Flash, GPT-5.6 Luna and DeepSeek V4 Flash are the names that matter this month — and 200+ more are callable behind a single API key. But a list of names is not an answer. It is an answer only when each row carries a price, a context window and a capability, and when the list is current enough that the price you read is the price you pay. That is what this page is: the current lineup with real numbers, how to read the columns, and the three filters that turn 200+ models into one good pick. Every figure below was read from its source on August 10, 2026.

A usable list has four columns

A page of model names — no prices, no context windows, no capability notes — is how people end up paying $50 a million tokens for a job a $0.09 model handles. A list that is actually usable carries four things per row, and none of them is optional.

• Name and vendor. The precise model, not the family. GPT-5.6 Luna and GPT-5.6 Sol are different products with a fifty-fold price gap, and a list that blurs them is worse than no list.

• Price per 1M tokens. Input and output, listed separately — the two differ by as much as twenty-fold on some models, and real bills are driven by output, so a single blended number hides the truth.

• Context window. How much of your source material fits in one call. A 200K window, a 1M window and a 1.3M window are not interchangeable, and the difference decides whether your prompt survives intact.

• Capability and modality. General reasoning, coding, agentic reliability, and whether the model accepts images. A text-only model will not debug the screenshot you paste into the thread.

The frontier lineup, right now

The leaderboard changes weekly, so here is the snapshot as of August 10, 2026. Prices are the OrcaRouter list price — provider price passed through with no markup — and the index is the Artificial Analysis Intelligence Index, read today.

• Claude Opus 5 — $5 / $25, 1M context, AA Index 63. The top general-purpose score on the index this month and the default for hard reasoning and long agent work.

• Claude Fable 5 — $10 / $50, 1M, AA Index 62. The most expensive frontier model on this list, and a proven #2 by the index — pay the premium only when the hard tail of the work justifies the meter.

• GPT-5.6 Sol — $5 / $30, 1M, AA Index 61. O​penAI's strongest general model, multimodal, with a coding profile that keeps the Claude pair honest.

• Kimi K3 — $3 / $15, 1M, AA Index 60. The best value in the top five: Moonshot's flagship at roughly a third of Claude Opus 5's price.

• Qwen3.8 Max — $2 / $6, 1M, AA Index 58. Alibaba's flagship, the strongest open-weights option at the frontier, and a price outlier for its index score.

• Claude Sonnet 5 — $2 / $10, 1M, AA Index 55. The value tier of the Claude family — the same conservative house style as its bigger siblings at a fifth of Claude Opus 5's price.

• Grok 4.5 — $2 / $6, 500K, AA Index 56. xAI's flagship, fast and cheap for its tier, with a smaller context window than the rest of this group.

• GLM 5.2 — $0.76 / $2.42, 1M, AA Index 53. Z.ai's open-weight model with a 1M window at a fraction of the closed-flagship price.

• Gemini 3.6 Flash — $1.50 / $7.50, 1M, AA Index 52. The speed pick: #2 for output speed on the index at around 238 tokens per second, built for latency-bound, high-volume loops.

• GPT-5.6 Luna — $0.10 / $0.60, 1M, AA Index 52. The same family, fifty times cheaper: a small-model price tag with a mid-tier index score.

• DeepSeek V4 Flash — $0.09 / $0.18, 1M, AA Index 52. The cheapest model on this list with a frontier-scale context window: a 284-billion-parameter, 13-billion-active open-weights mixture of experts.

What landed after this list was written

The lineup above was read on August 10, 2026. Two models have been added to the catalogue since, and both carry Intelligence Index v5.0.1 scores — a different evaluation set from the v4.3 numbers in that snapshot, so read them as a separate cohort rather than as a re-ranking of it.

• Gemini 4 Carbon — $2.00 / $10.00, 1M context, index 60, #4 of 78. Multimodal input including audio and video, a 65K output ceiling, and cache reads at $0.100 per 1M: the cheapest way on this page to put a million tokens of mixed media in front of a model that scores 60.

• Claude Fable 5.2 — $8.00 / $40.00, 1M context, index 65, #1 of 78. The strongest score on the current evaluation set and the most expensive entry on this page, with a 128K output ceiling, always-on thinking and cache reads at $0.200 per 1M.

How to filter the list

Nobody wants 300 models; they want the one that fits. Three filters do most of the work, and the live catalog is built around exactly these.

• Filter by capability first. Start with what the model has to do. Hard reasoning and long agent loops → Claude Opus 5 or Claude Fable 5. Repo-scale coding → the coding-index leaders, Claude Opus 5 and GPT-5.6 Sol. Latency-bound autocomplete → Gemini 3.6 Flash. And if a prompt will ever include an image, drop every text-only model before doing anything else — that single filter removes DeepSeek V4 Flash immediately.

• Filter by price second. The spread this month runs from free (NVIDIA's Nemotron tiers, Google's Gemma 4 31B free tier) to $10 / $50 for Claude Fable 5 — more than a hundred times the bottom. The useful middle is where most production sits: DeepSeek V4 Flash and GPT-5.6 Luna under $0.20 / $0.60, Claude Sonnet 5 and Qwen3.8 Max around $2, the flagships at $3–$10 input. A price filter that cannot show you that spread is not a price filter.

• Filter by context window third. Claude Haiku 4.5 is capped at 200K; most of the frontier sits at 1M; Llama 4 Scout stretches to 1.3M and Grok 4.20 to 2M. But advertised context is a ceiling, not a guarantee — quality degrades as you approach it, so buy headroom. If your payload is 150K tokens, a 200K model is a trap, not a fit.

Where the live list lives

Every table on a blog post — including this one — is a snapshot, and in this market a snapshot is stale within a week. The live version of the list lives on OrcaRouter's /models page: 200+ models behind one API key, with provider, context and price filters in the sidebar and modality tabs on top. The property that page has and the listicles all lack is that it is the actual catalog, not a paste of one. Frontier releases land there within about a day of the vendor announcing them, open-weights within a few days, and the price shown is the live upstream rate passed through at 0% markup — what the page says is what your account is billed. That is also the point of routing 200+ models on one key: when the list changes, and it changes weekly, your code does not. One contract, one API shape, and the model-id is the only thing that moves.

When the list misleads you

The columns make a list usable, but they also manufacture false confidence. Four traps are worth naming before you trust any row.

• Newest is not best. A model released last week is not automatically the right call. The index is recomputed as variants and effort levels change, and a proven model with failover baked in beats a brand-new one on the day it ships.

• Price is not capability. GPT-5.6 Luna at $0.10 / $0.60 posts an index of 52 — within a point of Gemini 3.6 Flash at fifteen times the price — for the tasks both handle. The premium tiers buy the hard tail, not the average case.

• Context length is advertised, not earned. A 1M window does not mean a model uses 1M tokens well. Treat the number as an upper bound and buy headroom, not as a quality promise.

• One index hides variance. Two models at the same index score differently across reasoning, coding and agentic suites. The list gives you the starting lineup; the per-task ranking decides the game — which is the entire argument of the best LLM for coding guide, and why the same model does not win every job.

• Open vs closed changes the question. For self-hosting, air-gapped data or hard cost caps, the ranking flips: GLM 5.2, Qwen3.8 Max and DeepSeek V4 Flash become more attractive than their index position suggests, because you are not paying per token at all.

The short version

A list of LLM models is only worth what its columns carry. The current frontier — Claude Opus 5, Claude Fable 5, GPT-5.6 Sol, Kimi K3, Qwen3.8 Max, Claude Sonnet 5, Grok 4.5, GLM 5.2, Gemini 3.6 Flash, GPT-5.6 Luna and DeepSeek V4 Flash — spans a price range of roughly a hundred to one and context windows from 200K to 2M. Filter by capability first, price second, context third. And when the answer to "which model" keeps changing, stop memorizing a snapshot: use the live /models catalog, where 200+ models sit behind one key at passed-through list price. This page will be stale in a week. The list on the /models page will not.