What a month of coding-agent tokens costs: subscriptions vs API
One heavy month of agentic coding, priced across Anthropic, OpenAI, the open-model subscriptions and pay-per-token providers.
The uncensored DeepSeek V4.1 Flash post priced a single session. Coding agents don’t run single sessions, so here’s a full month, priced across subscriptions and APIs.
Prices and limits as of 3 November 2026.
The workload
A heavy month of agentic coding, rounded up from real tku stats:
| Tokens | |
|---|---|
| Input, uncached | 500M |
| Cache reads | 15B |
| Output | 50M |
| Requests | 100K |
About 97% of that is cache reads; coding agents resend the same context on every turn. For a casual user, the math looks very different - pay per token is often cheaper than any subscription.
The short version
- Strongest models: Anthropic and OpenAI subscriptions, by far. At API rates this month is about $6,500 on Opus 5.5 and $12,500 on Fable 5.1. Claude Max 20x ($200) carries about $2,000 of API usage a week (measured with tku). ChatGPT Pro 200 plausibly carries it on GPT-6.1 Sol, which is $3,000 at API rates.
- Weaker models: 2-4x cheaper on subscription, at about the same token volume. Alibaba’s Coding Plan Pro ($50) almost fits; Ollama Cloud Max ($100) fits. Weaker models take more turns per task, so the same tokens buy less finished work.
- Uncensored: pay per token is the only real option. No subscription carries this volume with an uncensored model. Maple has plans, but doesn’t publish what a credit is worth.
- Frontier models on the API cost 15-110x the subscription. For cheap models, the gap mostly disappears: DeepSeek V4.1 Flash on the API is $150-355 for the month.
How the models compare
Independent scores only (Vals, Artificial Analysis, Scale):
| Model | Terminal-Bench 4.0 | AA index | SWE-Bench Pro V2 |
|---|---|---|---|
| Claude Opus 5.5 | 65.2 | 58 | - |
| Claude Sonnet 5.5 | 64.1 | 54 | 70.6 |
| Kimi K3.1 Pro | 63.6 | 46 | - |
| GPT-6 Astra | 59.6 | 57 | 90.2 |
| Claude Fable 5.1 | 58.1 | 49 | 92.2 |
| GLM-5.3 | 38.9 | 32 | 84.3 |
| Qwen3.8-Max | 38.9 | 34 | - |
| DeepSeek V4.1 Flash 1017 | 31.8 | 32 | - |
| MiniMax M3.1 Flash | 25.3 | 29 | 84.3 |
| Kimi K3 | 17.2 | - | 88.2 |
- On long terminal tasks, the open models trail Opus 5.5 by 26-48 points. On single-issue repo fixes (SWE-Bench Pro V2), they’re 4-8 points behind Fable 5.1.
- Kimi K3.1 Pro is the exception, about 2 points behind Opus 5.5. Its weights are due on 4 November.
- Vendor-reported scores often run 10-13 points above the independent runs. The harness matters as much as the model.
Of course, benchmarks aren’t a uniform measure. In my experience, Fable 5.1 often does better than Opus 5.5, despite the scores. An uncensored DeepSeek V4.1 Flash does surprisingly well on long-running tasks, and it doesn’t stop to question security work the way the Claude models do. Pick the model for the task.
Subscriptions
Against the month above:
| Plan | Price | Models | Limit | Fits? |
|---|---|---|---|---|
| Claude Max 20x | $200 | Opus 5.5, Fable 5.1, Sonnet 5.5 | weekly | yes on Opus 5.5, not all-Fable |
| ChatGPT Pro 200 | $200 | GPT-6.1 Sol, GPT-6 Astra | weekly, size unpublished | plausibly, on Sol |
| ChatGPT Pro 100 / 500 | $100 / $500 | same | weekly | unlikely / plausibly |
| Ollama Cloud Max | $100 | DeepSeek V4.1 Flash, GLM-5.3, Kimi K3, MiniMax | $300 of credits | yes, on V4.1 Flash ($150-300) |
| Synthetic Big | $90 | DeepSeek, GLM-5.3, Kimi K3, Qwen | “Infinite mode”, much slower | maybe |
| Alibaba Coding Plan Pro | $50 | Qwen 3.7 Plus, Kimi K2.5, GLM-5, MiniMax M2.5 | 90K requests/month | almost |
| OpenCode Go Plus | $40 | DeepSeek, GLM-5.3, Kimi K3, Qwen | $200/month on V4.1 Flash, during a 4x promo | maybe, while it lasts |
| Z.ai GLM Coding Max | $168 | GLM-5.3 | 140K credits/week | no, ~20% (40% off-peak) |
| NanoGPT Pro | $12 | open models | 60M input tokens/week | no |
| Kimi Code Ultra | ¥1,399 (~$195) | Kimi | pool size unpublished | unknown |
Not worth it at this volume: GitHub Copilot (Max $100 = $200 of credits at API rates), Kiro (Power $200 = 10,000 credits), and Cursor Ultra ($200, third-party models billed separately). Google’s AI Pro and Ultra no longer work in Gemini CLI, and Antigravity’s terms ban third-party tools.
OpenAI officially lets you sign in with ChatGPT from OpenCode, Amp, Kilo Code and others; that usage comes out of the same plan.
Pay per token
The same month on the API:
| Provider | Model | Month | Privacy (their claim) |
|---|---|---|---|
| DeepSeek | V4.1 Flash | $150 off-peak / $300 peak | China-hosted |
| NEAR AI | V4.1 Flash 1017 abliterated | $300 | TEE, attested |
| Venice | V4.1 Flash uncensored | $355 | zero retention (Private) |
| Refuseless | V4.1 Flash abliterated | $620 | policy |
| heabsy | V4.1 Flash uncensored | $870 | policy |
| Maple | V4.1 Flash uncensored | $2,350 | TEE, E2EE |
| OpenAI | GPT-6.1 Sol | $3,000 | - |
| Anthropic | Sonnet 5.5 | $4,750 | - |
| Anthropic | Opus 5.5 | $6,500 | - |
| Anthropic | Fable 5.1 | $12,500 | - |
| OpenAI | GPT-6 Astra | $22,500 | - |
Anthropic rows bill the uncached input as cache writes. The cache-read price decides almost everything at this volume: Refuseless is the cheapest uncensored provider per input token ($0.14), but its cache reads cost 4x Venice’s, so the month comes out at nearly double.
NEAR AI is new since the last post: the same abliterated 1017 build, in an attested TEE, at stock DeepSeek peak prices.
Who serves what
| Provider | V4.1 Flash | Uncensored | GLM-5.3 | Kimi | Qwen | MiniMax | Access |
|---|---|---|---|---|---|---|---|
| Venice | yes | yes | yes | K3 | yes | yes | per token |
| NEAR AI | yes | yes | yes | K3 | yes | - | per token, TEE |
| Maple | yes | yes | yes | K3 | - | - | plans + credits, TEE |
| Phala / Redpill | yes | Qwen, Gemma | yes | K3 | yes | - | per token, TEE |
| DeepSeek | yes | - | - | - | - | - | per token |
| Ollama Cloud | yes | - | yes | K3 | - | yes | plans + credits |
| OpenCode Go | yes | - | yes | K3 | yes | M3 | plans |
| Synthetic | yes | - | yes | K3 | yes | - | plans, per token |
| NanoGPT | yes | GLM-5.3 | yes | yes | yes | yes | plans, per token |
| Alibaba Coding Plan | - | - | GLM-5 | K2.5 | yes | M2.5 | plans |
| Z.ai | - | - | yes | - | - | - | plans |
Caveats
- Plan limits are a pricing decision. Flat plans average heavy users against light ones, and API list prices aren’t what inference costs to run (cache reads are cheap to serve). Limits still move on short notice: Copilot moved to metered billing in June, Google pulled its subscriptions from Gemini CLI, and OpenAI cut and restored Pro allowances within a month.
- A token isn’t a token. Different tokenizers, and weaker models take more turns. In one hands-on test, Kimi K3.1 Pro needed 28 turns where Opus 5.5 needed 7.
- Third-party hosting varies. Moonshot’s vendor verifier found tool-call accuracy as low as 72-83% on some OpenRouter routes.
- Abliteration costs a bit. The dealignai card reports MMLU dropping from 87.63% to 83.77%; nobody has published coding benchmarks for it.
If you want frontier models, the Anthropic and OpenAI subscriptions are the best deal by an order of magnitude. If you need uncensored or private inference, you pay for it per token - and with DeepSeek V4.1 Flash, that’s a few hundred dollars a month, even at this volume.