Every token-priced model in our index, ordered by input price. The spread is roughly seventy to one between the cheapest usable tier and the frontier, which is wide enough that routing by task difficulty is not an optimisation, it is the architecture. The cached-input column is the one that decides agent economics, and it is the one most comparisons leave out.
List prices per million tokens from each vendor's own price list, standard tier, checked on 8 September 2026. Batch and flex discounts, long-context surcharges and regional variation are noted below rather than in the table, because folding them into a single number would hide the thing that actually moves your bill. Self-hosted models are excluded here and covered in the open-weight collection, since their cost is hardware and engineering rather than tokens.
| Model | Vendor | $ / 1M in · out | Context | Weights | Released |
|---|---|---|---|---|---|
01 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 | $0.14 · $0.42 | 1M | Closed | 26 Aug 2026 |
02 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 | $0.20 · $1.20$0.02 cached | 1.05M | Closed | 26 Jun 2026 |
03 Gemini 3.8 Flash Google's newest Flash, introductory pricing at $0.75 / $3.75 through 31 December 2026 and double that from January. | $0.75 · $3.75 | — | Closed | Sep 2026 | |
04 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 | $1.40 · $4.40$0.26 cached | 1M | Open | 25 Aug 2026 |
05 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 | $2 · $12$0.20 cached | 1.05M | Closed | 26 Jun 2026 |
06 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 | $2 · $6$0.25 cached | 1M | Closed | 3 Aug 2026 |
07 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 | $4 · $20$0.40 cached | 1.05M | Closed | 26 Jun 2026 |
08 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 | $5 · $25 | 1M | Closed | 24 Jul 2026 |
09 GPT-Image-2 Token-priced rather than per-image, so you cannot quote a cost per picture until you measure your own prompts. | OpenAI | $5 · $30$1.25 cached | — | Closed | Apr 2026 |
10 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 | $10 · $50$1 cached | 1.05M | Closed | 3 Sep 2026 |
11 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 | $10 · $50$0.25 cached | 1M | Closed | 1 Sep 2026 |
12 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 | $10 · $50$0.25 cached | 1M | Closed | 1 Sep 2026 |
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.
Claude Fable 5.1 reads cached input at $0.25 per million against GPT-6 Astra's $1, on identical $10 headline pricing. GLM-5.3 charges $0.26 against a $1.40 base. An agent re-reads its stable prefix on every turn, so on a long session the cache rate matters more than the base rate does. Measure your own cache-hit ratio before modelling any saving.
Qwen3.8-Flash at $0.14 input against GPT-6 Astra at $10. At the cheap end, steps become worth doing that were not worth doing before: classifying every inbound message, extracting structure from every document, reviewing every call. Treating one frontier model as the default for all of it is the most common and most expensive architectural mistake we see.
Batch and flex run at 50% on most vendors. Fast modes run at 2x. GPT-6 Astra reprices the entire request above 272,000 input tokens. Alibaba's Beijing endpoint is 60 to 70% cheaper than Singapore, which is a data-residency decision before it is a pricing one. Model your actual traffic shape rather than multiplying volume by a list price.
Z.ai's Coding Plan runs $18 to $168 a month. Almost nobody else offers a flat rate, and unpredictable agent spend causes more internal friction than high agent spend does, because one bad loop can cost more than a fortnight of normal use. For a team running coding agents daily this is often worth more than a lower per-token price.
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.
