On August 13 I ran a single long agent task through ChatOSS on DeepSeek V4 Pro 0813. It ran for about an hour, made 157 API calls, and carried a large context the whole way — 31.2 million input tokens and 107,000 output tokens in total. When it finished I opened the billing log expecting a few dollars. The charge was $0.279594. Twenty-eight cents.
That number only means something next to an alternative, so I priced the exact same token volume against the three closed frontier models people actually compare it to: Claude Opus 5, GPT-5.6 Sol, and Claude Fable 5. The short version: $19.05, $19.58, and $38.09.
Where the numbers come from
These are not estimates from a pricing calculator. ChatOSS writes one row per model call into a usage_events table — model, prompt tokens, completion tokens, total tokens, and the dollar cost the provider actually charged. I summed my own rows for a one-hour window, 20:54 to 21:54 UTC on 2026-08-13. Every event in that window was deepseek/deepseek-v4-pro-0813.
| Metric | Value |
|---|---|
| API calls | 157 |
| Input tokens | 31,171,613 |
| Output tokens | 106,938 |
| Wall-clock time | ~1 hour |
| Total billed | $0.279594 |
Why an hour of work costs 28 cents
The whole answer is the cache. An agent loop re-sends the same growing context on every turn, so after the first call almost none of those 31 million input tokens are new — they are the same prefix, read back out of DeepSeek's cache. My billing log records total cost but not the cache split, so I reconstructed it from the price: the only way $0.28 works across 31.2M input tokens is a cache-hit rate around 99.4%, billed at DeepSeek's $0.003625 per million. The reconstruction lands within 0.6% of the real charge.
| Component | Tokens | Rate / 1M | Cost |
|---|---|---|---|
| Cache read | ~31.0M (99.4%) | $0.003625 | $0.11 |
| New input (cache miss) | ~175K | $0.435 | $0.08 |
| Output | 106,938 | $0.87 | $0.09 |
| Total | 31.3M | — | ~$0.28 |
The same tokens on Opus 5, Sol, and Fable 5
Now hold my token volumes constant — same 31.0M cache reads, same 175K of new input, same 107K of output — and apply each model's published list price. Nothing else changes.
| Model | Cached / 1M | Input / 1M | Output / 1M | Cost for my hour | vs DeepSeek |
|---|---|---|---|---|---|
| DeepSeek V4 Pro 0813 | $0.0036 | $0.435 | $0.87 | $0.28 | 1× |
| Claude Opus 5 | $0.50 | $5.00 | $25.00 | $19.05 | 68× |
| GPT-5.6 Sol | $0.50 | $5.00 | $30.00 | $19.58 | 70× |
| Claude Fable 5 | $1.00 | $10.00 | $50.00 | $38.09 | 136× |
Read the cached column, not the input column. That is where the entire gap lives: 31 million cache reads cost me eleven cents and would cost $15.50 on Opus 5 or Sol, and $31 on Fable 5. On long agent runs the cache-read rate is the price of the model.
GPT-5.6 Sol, line by line
Sol deserves its own breakdown, because it is the model most people are actually paying for. At $0.50 per million cached, $5 per million input, and $30 per million output, my hour comes to $15.50 of cache reads, $0.87 of new input, and $3.21 of output — $19.58, or 70× what DeepSeek charged. Artificial Analysis puts the same gap on its own terms: about $1.23 per Intelligence Index task for Sol versus $0.06 for V4 Pro 0813, roughly 20×.
The awkward part for Sol is that the premium does not buy a better coding score. On SWE-bench Verified, scored independently by vals.ai, Sol lands around 96.0% and V4 Pro 0813 lands at 96.40%.
500 real GitHub issues resolved with a bash tool only. Bars scaled from a 93% floor to make the spread visible.
Two caveats, one in each direction
The comparison flatters the expensive models. I handed Opus 5, Sol, and Fable 5 the same 99.4% cache-hit rate DeepSeek earned. They would not get it: Anthropic's prompt cache has a five-minute default lifetime, while DeepSeek's disk cache persists far longer. Over a full hour those models would re-pay full input price on a chunk of that context, so $19.05, $19.58, and $38.09 are floors, not ceilings.
And DeepSeek's price is about to move. A peak/off-peak billing change takes effect August 16, 2026, with peak rates roughly double off-peak. Even doubled, my hour would have cost around 56 cents.
What I had wrong before I checked
I had been telling people this task would cost $5–$10 on Opus and $10–$20 on Fable 5. Both were too low, by a lot — the real figures are ~$19 and ~$38. My “100× cheaper” line turned out to be conservative against Fable 5 (136×) and a little generous against Opus 5 (68×) and Sol (70×).
One claim I have to split in half. “Same level” holds on coding: 96.40 against 97.00 on SWE-bench Verified is inside the noise. It does not hold in general. On the Artificial Analysis Intelligence Index — a composite of nine reasoning, coding, and knowledge evals — Opus 5 scores 63, Fable 5 62, Sol 61, and V4 Pro 0813 scores 53. On Vibe Code Bench v1.1, which asks a model to build a working web app from a spec, Fable 5 leads at 90.35% and V4 Pro sits around fifth. The closed frontier is still ahead on broad reasoning and from-scratch app building.
The takeaway
For long-running, cache-heavy agent work — the workload where bills actually get scary — V4 Pro 0813 is 68–136× cheaper than the closed frontier and gives up almost nothing on the coding benchmark that matters. The sensible policy is to default to it and escalate to Opus 5 or Fable 5 on the specific tasks where those last few points of composite reasoning earn their price.
Do not take my word for the numbers, though. Run one of your own long-context tasks, then open your billing log and read the cache-read line. That is the whole argument.