Key Takeaways
-
On August 13, DeepSeek swapped its flagship API model to the
DeepSeek-V4-Pro-0813 build. Same API name
(
deepseek-v4-pro), same price ($0.435 / $0.87 per 1M tokens official), same 1M context. The model underneath changed a lot. - The agent and coding gains over the preview build are large. Terminal-Bench 2.1 went from 72.1 to 87.9, DeepSWE from 12.8 to 62.7, DSBench-FullStack from 41.8 to 71.1. All vendor-reported.
- Against the frontier it is now within 0.1 points of Claude Fable 5 on Terminal-Bench (87.9 vs 88.0) and beats Claude Opus 4.8 on four agent benchmarks. It still trails clearly on knowledge-heavy work, scoring 42.7 on HLE vs Opus 5's 64.7, and 53 on the AA Intelligence Index vs Fable 5's 62.
- DeepSeek has officially pre-announced a significant API price increase. The current rates are the cheapest this model will ever be.
-
Because DeepSeek upgraded the model behind the existing API name,
the 0813 build is already what you get when you call
deepseek-v4-proon iCreat ($0.437 / $0.87 per 1M, 1M context), alongside V4 Flash at $0.14 / $0.28.
1. What actually shipped
There was no launch event. At around 1 p.m. Beijing time on August 13, DeepSeek's
API documentation
changed the model version behind deepseek-v4-pro from the preview build
to DeepSeek-V4-Pro-0813. Existing integrations got the new model
without changing a line of code, which is how DeepSeek has shipped flagship updates
since the V3 days.
The confirmed specs, per the official docs:
-
Model version DeepSeek-V4-Pro-0813, called as
deepseek-v4-pro. The smallerdeepseek-v4-flashremains on the 0731 build. - Context 1M tokens, with a maximum output of 384K tokens.
- Thinking mode on by default, switchable to non-thinking per request.
- Official pricing $0.435 per 1M input tokens (cache miss), $0.87 per 1M output, cache hits at $0.003625. Concurrency is capped at 500 for Pro on the official API.
- API compatibility OpenAI format and Anthropic format both supported.
One line in the official pricing page deserves more attention than the benchmarks. Quoting it in full, because it will age well:
DeepSeek rarely pre-announces anything. A pre-announced significant price increase means the current rates have a shelf life, and every cost calculation in this post should be read with that in mind.
On open weights, be careful with the popular claim. The V4 family is on Hugging Face under an MIT license (V4-Pro at 1.6T total parameters with 49B active, V4-Flash at 284B with 13B active, per the Hugging Face release notes), but as of this writing the model card still reflects the preview checkpoint. Weights for the hosted 0813 build have not been published yet.
2. Benchmarks, the leap and the limits
The most informative comparison is against DeepSeek's own preview build, because it shows what the 0813 update actually bought. Vendor-reported numbers, from the release table:
| Benchmark | V4 Pro Preview | V4 Pro 0813 |
|---|---|---|
| Terminal-Bench 2.1 | 72.1 | 87.9 |
| DeepSWE | 12.8 | 62.7 |
| DSBench-FullStack | 41.8 | 71.1 |
| CyberGym | 52.7 | 83.3 |
| HLE (with tools) | 48.2 | 60.0 |
These are not incremental gains. DeepSWE quintupled. A model that stumbled on multi-step coding agents in July is now competitive on them in August.
Against the frontier, the picture splits cleanly in two. The full release table below puts V4 Pro 0813 next to Claude Opus 4.8 and Claude Fable 5 across all ten agent benchmarks:
*The full vendor-reported benchmark table for the 0813 release. Infographic by DeepSeek V4 Hub, data from DeepSeek's release notes.*
Where it stands level or wins. On Terminal-Bench 2.1 it posts 87.9 against Claude Fable 5's 88.0, effectively a tie with a model that costs an order of magnitude more, and clearly ahead of Claude Opus 4.8's 85.0. It also beats Opus 4.8 on CyberGym (83.3 vs 78.3), DeepSWE (62.7 vs 58.0), and AutomationBench (31.8 vs 27.2), and edges Fable 5 on CyberGym (83.3 vs 83.1) and AutomationBench.
Where it clearly loses. Knowledge-heavy and long-horizon reasoning. HLE at 42.7 against Fable 5's 53.3 and Opus 4.8's 49.8. NL2Repo at 61.5 against Opus 4.8's 69.7. Toolathlon-Verified at 74.1 against Fable 5's 77.9. On the composite Artificial Analysis Intelligence Index it scores 53, well above the open-weight median of 27 but a tier below Fable 5 Max (62), Grok 4.6, and GPT-5.6 Sol Max (61 each). AA also measured it as somewhat verbose, generating 130M tokens on the index run against a median of 100M, which matters when you pay per output token.
Two caveats before you quote any of these numbers. First, the comparison scores are measured and published by DeepSeek itself, and the testing harness has not been disclosed (the company says a "DeepSeek Harness minimal mode" is coming). Second, DSBench-FullStack and DSBench-Hard are DeepSeek's internal test sets with no external leaderboard, so treat wins there as unverifiable for now. Independent numbers from Artificial Analysis and open evals will settle the picture over the coming weeks.
The honest one-line read, echoed by industry coverage: on shared agent benchmarks Fable 5 leads by roughly 5 percent on average, and the blended API price gap is roughly 46x.
3. It is already live, no waiting
Here is the practical consequence of DeepSeek's in-place upgrade. Any platform serving DeepSeek's official API got the 0813 build the moment DeepSeek flipped it. That includes iCreat, where both V4 models are served from the official upstream:
| Model | iCreat price (per 1M in/out) | Official price | Context |
|---|---|---|---|
| DeepSeek V4 Pro (0813) | $0.437 / $0.87 | $0.435 / $0.87 | 1M |
| DeepSeek V4 Flash (0731) | $0.14 / $0.28 | $0.14 / $0.28 | 1M |
Unlike the Fable 5 and GPT-5.6 Sol listings on iCreat (which run about 90% below the labs' list prices, see our Grok 4.6 cost breakdown, the DeepSeek rates sit at official parity. There is no discount to advertise here, and when DeepSeek's announced increase lands, rates will move with the upstream. The pitch is a different one. You get the 0813 build today, on the same key that already covers Fable 5, GPT-5.6 Sol, and image, video, audio, and 3D APIs, without opening yet another provider account or queueing on the official platform during launch-week load.
4. The cost math
Same scenario we used for the Grok 4.6 analysis, an agentic workflow burning 10M input and 2M output tokens per day:
| Daily cost | Monthly (30d) | |
|---|---|---|
| Claude Fable 5 @ official | $200.00 | $6,000 |
| Grok 4.6 @ official | $32.00 | $960 |
| Fable 5 @ iCreat | $20.00 | $600 |
| DeepSeek V4 Pro 0813 @ iCreat | $6.11 | $183 |
| DeepSeek V4 Flash @ iCreat | $1.96 | $59 |
For terminal-heavy agent work, where V4 Pro 0813 now benchmarks within a rounding error of Fable 5, that is a 33x monthly cost difference against Fable 5 at list price. For knowledge-heavy work the benchmark gap says the cheap option will cost you in failed runs and retries, which brings us to the actual decision.
5. Route by task, not by loyalty
Terminal and coding agents, cost-sensitive pipelines. V4 Pro 0813 is the value play of the moment. Terminal-Bench 87.9 at $0.87 output is unmatched, and the pre-announced price increase is a reason to benchmark it this week rather than next month.
Knowledge work, complex reasoning, highest-stakes agent runs. The HLE and Intelligence Index gaps are real. Fable 5 Max remains the stronger model overall, and at iCreat's $1 / $5 it costs a fraction of its own list price. Run the expensive model where failure is expensive.
High-volume, low-complexity traffic. Classification, extraction, summarization at scale. V4 Flash at $0.14 / $0.28 with the same 1M context is hard to argue with.
The realistic setup is all three behind one key, routed per task. Model leadership has flipped four times this summer (Opus 5, Sol, Grok 4.6, now DeepSeek on the agent benchmarks), and single-provider bets keep aging badly.
6. Calling it takes one model string
iCreat's endpoints are OpenAI-compatible, so the 0813 build is a base URL and a model name away:
# DeepSeek V4 Pro 0813 — standard OpenAI SDK
from openai import OpenAI
client = OpenAI(
base_url="https://api.icreat.ai/v1/llm",
api_key="ICREAT_API_KEY",
)
completion = client.chat.completions.create(
model="deepseek-v4-pro",
messages=[{"role": "user", "content": "Hello"}],
)
Thinking mode is on by default and switchable per request ("thinking": {"type": "enabled"}
with adjustable reasoning_effort). Swap the model string to
deepseek-v4-flash for the cheaper tier, or to
claude-fable-5 when the task calls for it. There is no free tier, but
top-ups start at $1, which at V4 Pro rates buys roughly 2M input tokens of testing.
FAQ
deepseek-v4-pro; no code change is needed.
