Models holding 250,000 tokens or more, ordered by window size. The window is the easy part. What separates these is whether retrieval still works at the far end of it, and whether the price changes when you cross a threshold. GPT-6 Astra reprices the entire request above 272,000 input tokens, which turns a $10 model into a $20 one without anyone changing a setting.
Selected as every model in our index documenting a context window of 250,000 tokens or more, ordered by that number. Context figures are as published by the vendor or the model card. Published context and useful context are different numbers more often than vendors acknowledge, so we would run a retrieval test at your real context length before relying on any row here.
| Model | Vendor | Weights | Context | $ / 1M in · out | Licence |
|---|---|---|---|---|---|
01 GPT-6 Astra First model rated Critical for cyber under the Preparedness Framework. Records on computer use and terminals, fourth on a third-party intelligence index. | OpenAI | Closed | 1.05M | $10 · $50$1 cached | Proprietary |
02 GPT-5.6 Sol The prior OpenAI flagship and still the right default for most production work at two and a half times less than Astra. | OpenAI | Closed | 1.05M | $4 · $20$0.40 cached | Proprietary |
03 GPT-5.6 Luna OpenAI's cheap tier. Classification, routing and extraction, which is where most of an agent's token volume actually goes once you measure it. | OpenAI | Closed | 1.05M | $0.20 · $1.20$0.02 cached | Proprietary |
04 GPT-5.6 Terra The balanced OpenAI tier, and where most of the token volume belongs in a working system. Same context and tool surface as Sol. | OpenAI | Closed | 1.05M | $2 · $12$0.20 cached | Proprietary |
05 Claude Fable 5.1 Tops the Artificial Analysis Intelligence Index at roughly 66. Cache reads at $0.25, a quarter of Fable 5 and of every other Claude model. | Anthropic | Closed | 1M | $10 · $50$0.25 cached | Proprietary |
06 Claude Opus 5 Anthropic's recommended starting tier and the coding arena leader. 63.1 on the Artificial Analysis index at half Fable pricing. | Anthropic | Closed | 1M | $5 · $25 | Proprietary |
07 Gemini 3.1 Pro Leads OCR and visual question answering on the Nanonets document leaderboard, and handles sparse tables at 94% where most pipelines break. | Closed | 1M | Preview | Proprietary | |
08 Claude Mythos 5.1 Fable 5.1's specs and price, restricted to Project Glasswing participants. Scores 60.9 on Terminal-Bench 4.0 against Fable's 55.8. | Anthropic | Closed | 1M | $10 · $50$0.25 cached | Proprietary, invitation only |
09 Qwen3.8-Flash The cheapest usable tier we would put in production, flat across a 1M context with no long-context surcharge. Roughly seventy times cheaper than the frontier on input. | Alibaba | Closed | 1M | $0.14 · $0.42 | Proprietary |
10 Qwen3.8-Max Alibaba's hosted flagship with no public weights, at a fifth of Claude Opus 5 on input. The Beijing endpoint runs 60 to 70% cheaper than Singapore. | Alibaba | Closed | 1M | $2 · $6$0.25 cached | Proprietary |
11 GLM-5.3 zai-org/GLM-5.3 Level with GPT-5.6 Sol on Terminal-Bench 2.1 at a seventh of the input price, and state of the art on CyberGym vulnerability discovery. | Z.ai | Open | 1M | $1.40 · $4.40$0.26 cached | glm-5.3, bespoke |
12 Kimi K3 moonshotai/Kimi-K3 The largest open-weight model shipped, at 2.8 trillion parameters, and third on the Artificial Analysis index by third-party trackers. | Moonshot | Open | 1M | Self-hosted | Modified MIT |
13 GLM-5.3-Flash zai-org/GLM-5.3-Flash MIT at 320B, multimodal, and four points behind its own flagship. The most permissive licence attached to a model this size. | Z.ai | Open | 300K | Self-hosted | MIT |
14 Qwen3.8-27B Qwen/Qwen3.8-27B Beats Claude Opus 4.6 Max on SWE-bench Pro and OSWorld in Alibaba's table, at 27B under Apache 2.0. Second most-liked model on Hugging Face. | Alibaba | Open | 262K, 1M max | Self-hosted | Apache 2.0 |
15 Qwen3.8-Flash-Next Qwen/Qwen3.8-Flash-Next Beats Claude Opus by 22 points on AndroidWorld device automation. Frontier scores at roughly 6B inference cost, on a 180B machine. | Alibaba | Open | 262K, 1M max | Self-hosted | qwen-community-1.0 |
Every row carries the date we last verified it. Prices are list rates at the standard tier and exclude batch discounts, long-context surcharges and regional variation. Hugging Face download and like counts are pulled from the API rather than retyped, and they measure adoption rather than quality.
GPT-6 Astra bills at 2x input and 1.5x output above 272,000 input tokens, applied to the whole request rather than the excess. That makes it a $20 / $75 model rather than a $10 / $50 one for any genuinely long call. Claude Fable 5.1 prices its full 1M window at the standard rate with no surcharge, and on long-context work that is the deciding difference between two models with identical headline pricing.
GPT-6 Astra reports 96.3% on eight-needle MRCR retrieval in the 512K to 1M band against 73.8% for GPT-5.6 Sol. That twenty-two point gap is between two models with the same 1.05M window. Ask for retrieval numbers at your context length rather than accepting the headline, and run a needle test yourself before you build on it.
Long context is a capability, not a strategy. Feeding a model 800,000 tokens because you can is slower, more expensive and often less accurate than retrieving the right 8,000. We build for context hygiene first and reach for the full window when the task genuinely needs everything in view at once, which is rarer than it sounds.
GLM-5.3 and Kimi K3 both document 1M windows, and Qwen extends to 1M from a 262K native window. Self-hosting long context means the KV cache is your problem, and it is the thing that runs deployments out of memory in production after testing fine on short prompts. Size against p95 context and concurrency.
Every model we integrate runs through the same operating system. Three pillars, sixteen layers, one Compound Growth Loop. The methodology that keeps AI work from rotting after the first ship.
Read the K-FrameworkDirect API integration with the model. No LangChain, no orchestration vendor, no agent framework built on quicksand. Typed contracts, the same way we wire up Postgres.
An eval suite built from your real tasks gates every prompt and model change. Quality is measured before it ships, not vibed in a demo.
Governance, audit, and oversight wired in from day one. Who called what, with which prompt version, at what cost. Your auditors get answers, not screenshots.
A model in production without observability is roulette. We instrument every integration so engineering and finance can see the same numbers, and so a regression at 3am surfaces before a customer opens a ticket.
Tokens in, tokens out, dollars spent. Sliced by feature, tenant, and route. Budgets enforced where it matters.
Real distributions, not averages. We know which routes are slow, and why.
The same eval suite that gates a release runs continuously in production. A regression on real traffic surfaces fast.
PII scrubbed at the proxy, shipped to your SIEM. Retention controls match your compliance window.
Dashboards your team owns, not ours. At handoff you get the queries, the alerts, and the runbook. We are not in the path to read your metrics.
