DeepSeek V4 Pro 0813: frontier-class benchmarks at 100× less than Opus 5

DeepSeek shipped the finished version of its flagship on August 13, 2026 with no blog post and no fanfare — just a quietly changed version string on its API pricing page: DeepSeek-V4-Pro-0813. The general-availability build closes a preview period that ran nearly four months, and the numbers that came back are uncomfortable for the closed-weight incumbents. On SWE-bench Verified — the field's most-cited real-world coding benchmark — the open-weight V4 Pro 0813 lands at 96.40%, second overall and within 0.60 points of Claude Opus 5 at 97.00%. And it does it for roughly one-twentieth the input price and one-fiftieth the output price.

What the 0813 build actually is

V4 Pro 0813 is the production release of the model DeepSeek previewed on April 24, 2026. It is a mixture-of-experts system with 1.6 trillion total parameters and 49 billion active per token, supporting a one-million-token context window and a maximum output of 384,000 tokens. The architecture pairs two attention variants DeepSeek calls Compressed Sparse Attention and Heavily Compressed Attention, which the company says cut single-token inference compute to 27% and KV cache to 10% of what its V3.2 generation needed at the million-token setting. Both V4 models were pre-trained on more than 32 trillion tokens, and the weights ship on Hugging Face under the MIT license— meaning you can run, modify, and commercially deploy them yourself. The economics carry over from preview unchanged: $0.435 per million input tokens on a cache miss, $0.003625 per million on a cache hit, and $0.87 per million output tokens.

The benchmark story: it matches Fable 5, beats GPT-5.6 Sol

Independent evaluators at vals.ai ran V4 Pro 0813 through the two benchmarks that best separate frontier coding models, and the results upend the usual open-vs-closed ordering.

On SWE-bench Verified — 500 real GitHub issues that a model must resolve by editing code in unfamiliar repositories, using only a bash tool — the leaderboard now reads: Claude Opus 5 at 97.00%, DeepSeek V4 Pro 0813 at 96.40%, GPT-5.6 Sol at roughly 96%, and Claude Fable 5 at roughly 95.4%. V4 Pro 0813 is the first open-weight model to reach the very top of SWE-bench Verified, sitting ahead of closed models like Claude Opus 4.8 (88.60%) and Grok 4.5 (86.60%), and ahead of every other open model including Kimi K3 (93.40%).

On Vibe Code Bench v1.1 — a harder benchmark that asks a model to build a complete working web app from a natural-language spec in a full Docker environment, then checks it with point-and-click testing — Claude Fable 5 leads at 90.35% with Claude Opus 5 second at 88.40%. V4 Pro 0813 lands in the top group (around fifth), ahead of GPT-5.6 Sol in sixth. On the broader Artificial Analysis Intelligence Index — a composite of nine reasoning, coding, and knowledge evals — Claude Opus 5 leads at 63, Claude Fable 5 at 62, GPT-5.6 Sol at 61, Kimi K3 at 57, and V4 Pro 0813 at 53. So on pure composite intelligence the closed frontier is still ahead. But on the coding benchmarks that actually drive buying decisions, the open-weight model is now within rounding error of the top — and on SWE-bench Verified it beats GPT-5.6 Sol.

Reading the chart: the open-weight value pocket

The scatter above plots every model's blended price (the 7:2:1 cache-input-output blend Artificial Analysis uses) against its Intelligence Index. The log-scale x-axis is doing the heavy lifting here, because the price range is absurd: DeepSeek V4 Pro 0813 blends to about $0.65/M while Claude Fable 5 blends to about $30/M — a 46-fold gap for an 8.8% intelligence advantage. Claude Opus 5 at roughly $10/M is the smarter-but-pricier closed option, and GPT-5.6 Sol sits at the same $10/M blended point. The two solid bubbles in the bottom-left are the open-weight models (V4 Pro 0813 and Kimi K3), and they occupy a value pocket no closed model can touch: frontier-adjacent intelligence at cents-per-task pricing.

The headline number from Artificial Analysis is the cost per Intelligence Index task — the average dollars spent to complete one composite benchmark task. There the gap is even more stark: V4 Pro 0813 costs about $0.06 per task, versus $1.23 for GPT-5.6 Sol, $2.34 for Claude Opus 5, and $3.14 for Claude Fable 5. Fable 5 is roughly 52× more expensive per task than V4 Pro 0813; Opus 5 about 39×; GPT-5.6 Sol about 20×. That is the “10–100× cheaper” range in concrete terms.

Cost per task: the dramatic price gap

List prices tell part of the story; per-task cost tells the rest. A typical coding-agent interaction might burn ~10,000 input tokens and a few thousand output tokens, and on a long-running agentic job the closed models also tend to reason longer and generate more. Here are the published per-million-token rates side by side, plus an estimated cost for one representative SWE-bench-style task (~8,000 input + ~1,000 output tokens, 50% cache hit rate):

* Estimated per-task cost for a coding-agent interaction (~8,000 input + ~1,000 output tokens, 50% cache hit rate). On a cache miss the gap is wider still: V4 Pro 0813's cached-input rate of $0.0036/M is roughly 139× cheaper than Fable 5's $1.00/M and 278× cheaper than Opus 5's $0.50/M on the cache-hit path that dominates long-running agents.

At a cache-hit rate of over 93% in production (per DeepSeek's own OpenRouter telemetry), the effective input price V4 Pro 0813 actually charges drops to about $0.032/M — putting the real-world per-task cost even lower. For a team running high-volume agentic workloads, the math is transformative: roughly 10,000 SWE-bench tasks on V4 Pro 0813 cost about the same as 170 on Claude Fable 5.

Full comparison table

The open-vs-closed dynamic just flipped

For most of the last two years, the story was simple: closed models led on every benchmark, open models were cheap. V4 Pro 0813 scrambles that. On the single most-watched coding benchmark (SWE-bench Verified), an open-weight, MIT-licensed model now sits second overall, beating three closed frontier models (GPT-5.6 Sol, Claude Fable 5, and every Opus 4.x model). The closed frontier's remaining edge — a few points on composite intelligence and a ~5% lead on the hardest app-building task — now costs 20× to 52× more per task to obtain.

That changes the buying calculus. When the open option is good enough on the benchmark that matters to you and 1–2 orders of magnitude cheaper, the rational move for most high-volume workloads is to default to the open model and escalate to a closed one only on the tasks where the few extra points actually matter. DeepSeek's own nine-benchmark comparison table concedes that Claude Fable 5 still leads by an average of ~5.3% across the benchmarks where both are scored — and even leads on two of them. The honest reading is that the closed frontier is still ahead on raw composite intelligence. But “ahead by 5% at 4,500% the price,” as one outlet put it, is no longer an unambiguous win for the incumbent.

Caveats worth knowing

Two honest caveats. First, several of DeepSeek's own headline agent numbers (including its DSBench-FullStack and DSBench-Hard scores) come from internal test sets with no external leaderboard, so those particular claims await independent reproduction. That is why this article leans on vals.ai for SWE-bench Verified and Vibe Code Bench — both are run independently with a fixed, minimal bash-only harness, which is the apples-to-apples comparison the field trusts. Second, DeepSeek has signaled a peak/off-peak billing change taking effect August 16, 2026, with peak rates roughly double the off-peak ones — so the sticker price advantage can narrow during peak hours (01:00–04:00 and 06:00–10:00 UTC). Even at peak pricing, though, V4 Pro would still blend to about $1.30/M, an order of magnitude below the closed frontier.

The takeaway

DeepSeek V4 Pro 0813 is the model that makes the open-weight question non-optional. If you are building coding agents, running high-volume agentic workloads, or just watching your API bill, the practical move is to default to V4 Pro 0813 for the cost savings and escalate to a closed frontier model (Opus 5 for the hardest reasoning, Fable 5 for from-scratch app building) only when the task demands those last few points. That two-tier strategy — open by default, closed on demand — is exactly the workflow that's making open-weight models irresistible to cost-conscious teams. And with the weights shipped under MIT, you can always run it yourself and pay nothing per token at all.