“Which is cheapest for a coding agent: DeepSeek, Qwen or GLM?” sounds like a question you can answer from a price table. Usually you can’t. A coding agent doesn’t send one prompt and get one answer. It sends the whole growing conversation back on every turn, so the bill depends far more on how many turns you run, how big the context gets, and how much of it the provider caches than on the headline price per million tokens.
Below I work through one realistic agent task with official list prices, then scale it to a month. I build a flat-rate gateway (disclosure at the end), so I’ve also included the comparison with flat pricing. In some cases that comparison doesn’t come out in my favour.
All prices are USD per 1M tokens, taken from each vendor’s own pricing page:
| Model | Input (cache miss) | Input (cache hit) | Output | Source |
|---|---|---|---|---|
| DeepSeek V4 Pro, peak | $1.32 | $0.044 | $3.96 | DeepSeek |
| DeepSeek V4 Pro, off-peak | $0.66 | $0.022 | $1.98 | same |
| DeepSeek Flash (V4.1), peak | $0.30 | $0.006 | $1.20 | same |
| Qwen3.8-Max (International / Singapore) | $2.00 | $0.25 | $6.00 | Alibaba Model Studio |
| GLM-5.3 | $1.40 | $0.26 | $4.40 | Z.ai |
| GLM-5.3-Flash | $0.15 | $0.03 | $0.50 | same |
Some details that matter for agents:
Take one typical “fix this failing test” run in an agent like Cline or Claude Code:
That gives:
input = sum(15,000 + 3,000 × n) for n = 0..24 = 1,275,000 tokens
output = 25 × 1,500 = 37,500 tokens
Look at the ratio. Input is 97% of the tokens. In agent loops, the input price and the cache-hit price decide the bill, not the output price.
Each call starts with the exact text of the previous call, so in theory almost all of it can be a cache hit. In this example about 1.19M of the 1.275M input tokens are a repeated prefix, roughly 93%. Real hit rates are lower. Caches expire, some agents rewrite earlier messages, and context compaction breaks the prefix. So I’ve costed both ends:
def task_cost(miss, hit, out, cached_share):
inp, outp = 1_275_000, 37_500
return (inp * (1 - cached_share) * miss + inp * cached_share * hit + outp * out) / 1e6
| Model | Per task, no cache | Per task, ~93% cached |
|---|---|---|
| DeepSeek V4 Pro, peak | $1.83 | $0.32 |
| DeepSeek V4 Pro, off-peak | $0.92 | $0.16 |
| DeepSeek Flash, peak | $0.43 | $0.08 |
| Qwen3.8-Max (Intl) | $2.78 | $0.70 |
| GLM-5.3 | $1.95 | $0.60 |
| GLM-5.3-Flash | $0.21 | $0.07 |
Caching alone changes the per-task cost by 3–6x. The difference between “DeepSeek vs Qwen vs GLM” is smaller than the difference between “cached vs not cached” for any one of them.
Now say you run 12 of these tasks per working day, for 22 days a month. That’s 264 tasks and 6,600 model calls, or about 300 calls a day.
| Model | Month, no cache | Month, ~93% cached |
|---|---|---|
| DeepSeek V4 Pro, peak | $484 | $83 |
| DeepSeek V4 Pro, off-peak | $242 | $42 |
| DeepSeek Flash, peak | $113 | $21 |
| Qwen3.8-Max (Intl) | $733 | $184 |
| GLM-5.3 | $515 | $157 |
| GLM-5.3-Flash | $55 | $18 |
Two takeaways:
Flat plans meter something other than tokens. For a single vendor, Z.ai sells a GLM Coding Plan from $18/month, with 5-hour and weekly credit limits that are themselves computed from tokens. For one model family that’s often the cheapest predictable option.
For several vendors behind one key, my own gateway counts requests per day instead: $19/month for 500 a day, $39/month for 1,000. Heavier models count as more than one request. The public model list currently gives deepseek-v4-pro a weight of 1, and qwen3.8-max, glm-5.3 and glm-5.3-flash a weight of 2. For the 300 calls a day above:
| Model | Weighted requests/day | Flat plan needed | Per-token range (month) |
|---|---|---|---|
| DeepSeek V4 Pro | 300 | $19 (500/day) | $42–$484 |
| Qwen3.8-Max | 600 | $39 (1,000/day) | $184–$733 |
| GLM-5.3 | 600 | $39 (1,000/day) | $157–$515 |
| GLM-5.3-Flash | 600 | $39 (1,000/day) | $18–$55 |
Read honestly, that table says:
Don’t trust my 25 calls and 3,000 tokens per turn. Your agent, repo and habits will be different. Most agents and gateways log token usage per request. Take one real day, add up the cache-miss, cache-hit and output tokens per model, multiply by the table at the top, and count your calls. If your monthly per-token estimate is under about $20, stay per-token. If it’s well above your flat-plan price, or it swings a lot from week to week, the flat meter is probably the better fit.
Disclosure: I build APIClaw, the request-based gateway mentioned above. Vendor prices are from their official pricing pages as of October 2026, and vendors change prices, so recheck them before relying on these numbers.
Tools:
Direct answers:
Guides: