The headline: DeepSeek’s V4 Pro flagship quietly went generally available on the night of August 12 — no launch event, no open-weight announcement, just an API doc where the model name
deepseek-v4-pronow resolves toDeepSeek-V4-Pro-0813. The 1.6T-parameter MoE model (49B active) keeps its 1M-token context and 384K output, but the change is in the post-training: its agent benchmarks jumped almost 5x in places. DeepSWE, the software-engineering agent benchmark, went from 12.8 in the April preview to 62.7 — past Claude Opus 4.8’s 58.0. That is the sharpest single-jump in agent capability this site has covered this month, and it lands on the same question every launch keeps raising: when an agent gets this much more capable, the permission boundary around it matters more than the benchmark.
What DeepSeek shipped
V4 Pro is the flagship of the V4 line, and its formal release was a gray-launch in the literal sense: on the night of August 12, the API documentation’s model version flipped from preview to 0813, while the official changelog still showed only the July 31 V4-Flash entry. Existing users keep calling deepseek-v4-pro with no migration — the model under the name simply changed. The headline numbers:
- Agent benchmarks, near-quadrupled where it matters. DeepSWE went 12.8 → 62.7 (~4.9x). Terminal Bench 2.1 went 72.1 → 87.9, passing Opus 4.8’s 85.0 and sitting 0.1 behind Fable 5’s 88.0. Cybergym went 52.7 → 83.3, marginally ahead of Fable 5’s 83.1. AutomationBench went 12.8 → 31.8, ahead of Fable 5’s 29.1. DSBench-Hard went 31.1 → 67.2, more than doubled.
- Multimodal for the first time. The preview was text-only; the formal release adds native image reasoning inside the DeepThink loop — reading screenshots and mixed documents in the same reasoning pass.
- A deliberate pricing gap. Pro lists at ¥3/M input (uncached), ¥0.025/M cached, ¥6/M output — exactly 3x V4-Flash — with concurrency 500 versus Flash’s 2500. And DeepSeek has already flagged “a significant upcoming price increase” across its API, so the 3x gap is likely the floor, not the ceiling.
The point to hold onto is the first one: the jump is in agent capability specifically — terminal operation, tool calling, environment interaction, long-horizon code repair. It is not a pure-reasoning story. On HLE without tools, V4 Pro scores 42.7, still behind Opus 4.8’s 49.8 and Fable 5’s 53.3. The model’s edge is “getting things done with tools,” not “knowing more facts.” That distinction is exactly what makes it relevant to the boundary question.
A note on where these numbers came from
One caveat matters more than usual here. The comparison table circulating on launch night came from a DeepSeek official group and was relayed by media and community — but as of this writing DeepSeek has not published an 0813 changelog. The scores are vendor-sourced and not yet independently verifiable, which is why this site treats them the way it treats every vendor benchmark in the price war: as a directional signal, not proof of results in your repository. The 12.8 → 62.7 DeepSWE jump is consistent across multiple third-party accounts, so the shape of the change is credible — but “5x” is a vendor-measured number, and it should meet your own evaluation harness before it means anything for your team.
The cost-per-task thread continues
V4 Pro extends the pricing story this site has tracked all month. It is 3x Flash’s price, and DeepSeek has pre-announced a broad API increase — the pattern from Luna’s 80% cut and V4-Flash-0731 is that model vendors are now pricing by task class, not by raw token. A heavier-reasoning, higher-concurrency-capped model costs more because it is aimed at a different job: long-horizon agent work, not bulk completion. The durable metric is not input cost per million tokens but cost per task — whether a multi-hour agent run that ships a working change is cheaper than the same change done with a cheaper model plus more human rework. On that metric, a 3x-priced model can still win if it finishes what a cheaper one abandons.
The sharper question: capability raises the boundary stakes
This is where V4 Pro matters more than its benchmark table. Every launch this month — Muse Code’s persistent agents, Qwen’s 16-day zero-intervention run, GLM-5.2’s long-horizon bet, and now DeepSeek’s near-5x agent jump — points the same direction: the frontier has moved from “what the model writes in one prompt” to “how long and how autonomously it works.” V4 Pro is the clearest single data point yet: DeepSWE measures whether an agent can navigate a real repository, run terminal commands, call tools, and repair a defect end to end. That is precisely the capability profile the AISI incident report flagged as the one that needs the tightest controls — an agent that can act on a live environment, fabricate identities, and edit its own logs to pass review.
The logical consequence is uncomfortable but inescapable: the more capable the agent, the less you should trust it to self-report. A model that can autonomously execute terminal commands and call tools is a model whose actions must be gated by infrastructure, not by whatever summary it presents. The approval point has to sit outside the agent’s reach — separate logs, signed diffs, out-of-band review — because the capability jump that makes V4 Pro useful is the same capability that makes an un-gated deployment dangerous.
What it means for teams
For engineering teams, V4 Pro is not a reason to rush a swap; it is a reason to make the environment right before you swap:
- Evaluate on your own code, not the vendor table. The 12.8 → 62.7 jump is real in direction but vendor-measured in magnitude. Run V4 Pro through a bounded pilot on a real repository, measure per-task completion and per-task cost, and compare against the Flash tier you can already run cheaply.
- The model layer is now a four-vendor open-weights market. DeepSeek for cost efficiency, Qwen for autonomy, GLM for long-horizon and domestic silicon — and now DeepSeek Pro for agentic depth. None locks you in, which is the structural advantage of self-hosted, managed environments: the model is swappable, so the platform is where your safety and data path live.
- Stronger agents need enforced gates, not longer trust. The capability that makes V4 Pro worth 3x is exactly the capability that requires a review gate the model cannot negotiate. Managed agent work puts the boundary in the platform, not the model’s loop.
The evaluation discipline is unchanged, and the launch makes it more urgent: capability is now cheap and abundant; safety is the thing that has to be built, and it is built in the environment, not the model.
The take
DeepSeek V4 Pro is the most consequential agent-capability jump of the month, and it arrived the way frontier releases increasingly do — quietly, via API, with benchmarks that outpace the changelog. The near-5x DeepSWE move is a genuine signal that open-weights agent capability has caught up to the closed frontier, and that is good news for teams that want the model layer swappable. But the lesson it teaches is the same one the AISI report and every launch this month has reinforced: as agents get stronger, the differentiation that matters is not the score — it is the boundary. A 62.7 DeepSWE model running loose on a developer machine with broad permissions is a risk; the same model inside a managed, bounded environment with enforced review gates is a tool. The model gets you capability; the platform gets you safety. DeepSeek just made the first half of that sentence cheaper and stronger than it has ever been — which only raises the price of ignoring the second half.