DeepSeek V4 Pro 0813: GA Benchmarks, Pricing, and API Guide
DeepSeek V4 Pro did not suddenly become a 1.6-trillion-parameter model today. That machine has existed in preview since April. What changed on August 13 is more useful: DeepSeek finished turning it into an agent product.
The new DeepSeek-V4-Pro-0813 checkpoint is live in the app, on the web, and through the API. It arrives with much stronger repository and tool-use results, three reasoning-effort levels, a Responses API path shaped for Codex, a versioned MIT-licensed weight release, speculative decoding support, and public DeepSeek Harness code.
Then there is the price reset.
DeepSeek advertises off-peak rates at half the new peak tariff. That is true, but incomplete. When the new schedule begins on August 16, even off-peak Pro output will cost 128% more than it does under the pre-August 16 tariff. Peak output will cost 355% more. The cheapest serious way to use V4 Pro is therefore no longer simply “send everything to Pro.” It is to route work across four dimensions: model tier, reasoning effort, cache state, and time of day.
That is the central idea of this release. V4 Pro 0813 is the escalation lane in an agent scheduler—not the automatic default for every coding task.
Calling this “the V4 Pro launch” erases the most important part of the story: the model has been moving from research artifact to deployable agent stack in stages.
DeepSeek’s GA announcement is short. The changelog supplies the actual benchmark table, while the new 0813 model repository and DeepSeek Harness repository reveal what “GA” means operationally.
Our earlier V4 Flash 0731 analysis argued that DeepSeek was assembling an open agent stack, not merely releasing another chatbot. Today completes that thought—and changes the economics enough that the old Pro-versus-Flash cost comparison is now obsolete.
The headline parameter count did not move. The V4 family architecture underneath this release remains a text-only mixture-of-experts design with 1.6 trillion total parameters and 49 billion active per token, a 1-million-token context window, and up to 384,000 output tokens.
The change is post-training: better repository navigation, longer action sequences, tool use, terminal work, and recovery inside agent loops.
DeepSeek’s official numbers make the distinction unusually clear.
Here is the comparison that matters for buyers. Pro 0813 is a huge repair relative to April’s Pro Preview, but only a selective premium over the already improved Flash 0731.
| Official benchmark | Pro 0813 | Flash 0731 | Pro gain | Pro Preview | GA gain |
|---|---|---|---|---|---|
| Terminal Bench 2.1 | 87.9 | 82.7 | +5.2 | 72.1 | +15.8 |
| NL2Repo | 61.5 | 54.2 | +7.3 | 38.5 | +23.0 |
| Cybergym | 83.3 | 76.7 | +6.6 | 52.7 | +30.6 |
| DeepSWE | 62.7 | 54.4 | +8.3 | 12.8 | +49.9 |
| Toolathlon-Verified | 74.1 | 70.3 | +3.8 | 55.9 | +18.2 |
| Agents’ Last Exam | 25.7 | 25.2 | +0.5 | 16.5 | +9.2 |
| DSBench-FullStack | 71.1 | 68.7 | +2.4 | 41.8 | +29.3 |
| DSBench-Hard | 67.2 | 59.6 | +7.6 | 31.1 | +36.1 |
Those numbers support a precise verdict: major GA upgrade, modest tier upgrade.
If a hard task is worth paying more for, the 5–8 point gaps on repository, cyber, and software-engineering work can be meaningful. For cheap parallel exploration, routine edits, or tasks where a failed attempt is reversible, Flash has already captured most of the new post-training value.
Self-reported launch tables are a starting point, not a procurement decision.
Artificial Analysis gave Pro 0813 an Intelligence Index score of 53 in its August 13 measurement, placing it second within its 104-model comparable large open-weight class—not second among every model the site tracks. Its independent Terminal-Bench 2.1 run measured 78.65% for both Pro and Flash, versus DeepSeek’s own 87.9 and 82.7. Its hosted test also reported roughly 76.8 output tokens per second. That throughput sounds excellent until you look at the full trace: the page showed about 26 seconds of reasoning before the first answer token and roughly 34 seconds total. In other words, the model can decode quickly after a long deliberation phase. Token speed and task latency are not interchangeable.
Its cost result also has a short shelf life. Artificial Analysis tested the pre-August 16 $0.435 input and $0.87 output tariff. Any “value per task” chart built from those prices needs recalculation after August 16.
The most useful independent warning comes from an older model. NIST’s CAISI evaluation of the Pro Preview, not 0813, found that apparent parity on developer-selected tests weakened on CAISI’s held-out suite, where the model looked closer to GPT-5 and roughly eight months behind the then-frontier. That does not disprove the GA gains. It tells us exactly what is still missing: fresh, held-out, multi-run evaluation of 0813 under production-like tool and context conditions.
My read is straightforward: the official uplift is too broad to dismiss, but the evidence is not strong enough to replace your own traces.
There is a second release hiding underneath the model launch. DeepSeek published DeepSeek Harness, an MIT-licensed developer preview whose design slogan is “Everything is a Plugin.” The model card says its public code-agent scores used the local harness’s minimal mode, max reasoning effort, temperature 1.0, and top-p 0.95. Those sampling settings describe the local evaluation; the hosted thinking API accepts but ignores temperature and top-p.
That matters because the previous Flash release described this harness before outsiders could inspect it. Now the evaluation wrapper is visible.
Minimal mode is intentionally bare. Its default prompt is simply “You are a helpful software engineer assistant,” followed by this tool surface:
This is valuable for reproducibility. It is not a sensible production security policy. DeepSeek itself tells users to run the preset only against a disposable checkout or container.
The preset uses danger-full-access: its persistent Bash tool is unconfined, and the editor can address absolute paths visible to the harness process. It deliberately avoids the approvals, sandbox context, memory, and orchestration layers a real developer agent should have. Copying “minimal mode” into an internal coding product would optimize for matching the score while stripping away the controls that make an agent deployable.
This distinction is easy to miss in AI launches. A model benchmark measures the joint system:
model × prompt × effort × tools × harness × environment × retry policy
DeepSeek exposing more of that multiplication is a real contribution. It also makes the limitations easier to see.
DeepSeek now gives a builder four knobs that materially change cost and behavior:
Together, those four controls make V4 a workload scheduler.
This creates useful routing patterns that did not exist in the old flat tariff:
The same shift is happening elsewhere—our GPT-5.6 Codex rate-card analysis reached a similar conclusion from a different billing system. Once agents can run for minutes and fan out into subagents, pricing becomes a scheduling problem.
DeepSeek’s new pricing table takes effect at 16:00 UTC on August 16, 2026. The company describes off-peak rates as a 50% discount from peak. Mathematically, that is correct. Economically, the more useful baseline is what customers pay before the change.
Against the pre-August 16 tariff, every Pro rate rises—even off-peak.
| Pro, per 1M tokens | Until Aug. 16 | New off-peak | Change | New peak | Change |
|---|---|---|---|---|---|
| Cache-hit input | $0.003625 | $0.022 | +507% | $0.044 | +1,114% |
| Cache-miss input | $0.435 | $0.66 | +51.7% | $1.32 | +203% |
| Output | $0.87 | $1.98 | +128% | $3.96 | +355% |
The cache-hit line deserves attention. DeepSeek’s almost-free cached input helped long-running repository agents keep enormous stable prefixes economically. That rate goes from $0.003625 to $0.022 off-peak—about 6.1 times higher—and to $0.044 at peak, about 12.1 times higher.
Flash rises too. Its new off-peak miss/output rates are $0.22/$0.66; peak is $0.44/$1.32. Under the new schedule, Pro costs exactly 3× Flash for cache misses and output, and about 3.14× on cache hits.
That replaces the conclusion from our July analysis, where Pro cached input was only about 1.29× Flash. Warm-cache Pro is no longer a near-free upgrade.
The two daily peak windows are 01:00–04:00 UTC and 06:00–10:00 UTC. All other hours are off-peak.
DeepSeek does not state why it chose these hours. My inference is capacity management around Asian demand: both windows map closely to Chinese working hours and overlap Indian and European mornings. Regardless of motive, batch CI, overnight research queues, synthetic-data jobs, and scheduled code review should become clock-aware.
One common launch-day comparison also expires on August 16. At the pre-August 16 $0.87/M Pro output versus Fable 5’s $50/M, “57× cheaper” is roughly true. Under the new rate it becomes about 25× cheaper off-peak and 12.6× cheaper at peak. Still cheap—just not the same headline.
DeepSeek now exposes a native Responses API and an official Codex setup path. Existing API callers continue using the deepseek-v4-pro alias. In Codex-style clients, the important model and base settings are conceptually:
model = "deepseek-v4-pro"
model_provider = "deepseek"
preferred_auth_method = "apikey"
forced_login_method = "api"
model_reasoning_effort = "high"
[model_providers.deepseek]
name = "deepseek"
base_url = "https://api.deepseek.com/"
wire_api = "responses"
Use DeepSeek’s current setup documentation for the complete client-specific configuration, authentication, and model catalog, and review any installer before executing it. OpenAI’s Responses schema is becoming a distribution channel for competing models.
But compatibility has boundaries:
| Responses feature | DeepSeek behavior | Builder implication |
|---|---|---|
| Function tools | Supported | Normal structured tool loops work. |
| Built-in web search | Supported | Server-side search is available. |
| Custom tools | Only apply_patch | Do not assume arbitrary custom tool parity. |
previous_response_id, conversations, storage | Unsupported | Your client must own state and replay context. |
| Background mode | Unsupported | Long jobs need your own queue and worker lifecycle. |
| Images and files | Unsupported; images become placeholder text | V4 Pro remains text-only despite protocol compatibility. |
| MCP, computer use, code interpreter | Ignored | Feature detection is mandatory. |
| Unsupported parameters | Often silently ignored | A 200 response does not prove your requested behavior happened. |
The same caution applies to DeepSeek’s Anthropic-compatible endpoint. Model names that look like Claude tiers are mapped to DeepSeek Pro or Flash, but unsupported features can be ignored. This is valuable wire compatibility, not behavioral or capability equivalence.
V4 Pro 0813 supports low, high, and max. Thinking defaults to high. Under the Responses API, medium, high, and xhigh all map to high; max is its own mode, and none disables thinking.
Two implementation details are easy to get wrong:
reasoning_content must be sent back with subsequent messages or the API returns a 400.That second rule makes “reasoning state” part of the orchestration contract. If a gateway strips the field, the model may look flaky when the actual bug lives in your middleware.
The exact DeepSeek V4 Pro 0813 checkpoint is now downloadable under the MIT license. This is an important same-day correction to early coverage that found only the older Preview repository.
The release is substantial:
Open weights therefore provide portability, auditability, and a negotiating hedge against hosted API changes. They do not make this a casual local model. Download size alone is close to a terabyte; serving it well requires serious memory capacity, fast interconnects, and an inference team that understands expert parallelism, compressed attention, quantized caches, and long-context capacity planning.
The underlying V4 technical report describes the family architecture: hybrid compressed attention, mHC residual connections, DeepSeekMoE, Muon optimization for most parameters, and staged context training from 4K to 1 million tokens. DeepSeek says the Pro base saw 33 trillion training tokens and that its attention design uses 27% of V3.2’s single-token inference FLOPs and 10% of its KV cache at 1-million-token context.
Today’s model card says 0813 keeps the Preview model structure and attaches a DSpark speculative decoding module. DeepSeek’s DSpark paper describes lossless speculative decoding—faster generation without changing the target distribution. The checkpoint supports it through vLLM and SGLang. The GA package couples the checkpoint to a tuned runtime.
There was a brief, verifiable repository configuration issue: the new repository carried a Flash-shaped configuration before DeepSeek corrected it to the 61-layer, 384-expert Pro configuration. That is visible in the public commit history. It is not evidence that the hosted API rolled back; those separate community claims remain unconfirmed. It is evidence that teams should pin revisions and wait for their own smoke tests before allocating expensive infrastructure.
That simultaneity is strategically useful. In our Kimi K3 analysis, we argued that an open-model promise only becomes meaningful when the artifact, license, and runnable path align. DeepSeek has aligned them here, even if the hardware bar remains high.
Judge successful work per dollar and minute, including retries and cleanup.
Artificial Analysis’s first-day measurements make the routing case stronger. Pro scored about 53 on its Intelligence Index versus roughly 52 for Flash, while Flash generated near 115 tokens/second versus Pro’s 77. Both landed at 78.65% on its Terminal-Bench run, materially below DeepSeek’s own 87.9 Pro figure. Pro used fewer total output tokens across the suite, so its measured total eval cost was about 1.9× Flash rather than the roughly 3× list-price gap.
Those results are early and may move, but they show why neither “Pro is always better” nor “Flash is always cheaper” is rigorous. A more disciplined Pro trace can partially offset its token tariff. A faster Flash run can win when exploration is the task.
Use Flash for search, repository mapping, repetitive edits, first-pass bug sweeps, cheap parallel workers, and tasks where another attempt costs little.
Use Pro for architecture, ambiguous repairs, long dependency chains, high-stakes review, uncertainty resolution, and the final pass before merge.
Send difficult but non-urgent migrations, evaluation suites, and research jobs outside the two peak windows. Preserve stable prefixes to improve cache economics.
My preferred workflow is a ladder:
Flash/low scouts → Flash/high attempts → Pro/high repairs → Pro/max verifies
Do not invoke every rung automatically. Escalate when confidence drops, tests fail, the change radius expands, or the cost of a wrong answer crosses a threshold.
Do not migrate a coding fleet from a launch table. Shadow real work first.
One privacy point belongs in the engineering decision, not a legal footnote. DeepSeek’s public Privacy Policy says data is processed and stored in China and that prompts can be used to improve services. Consumer users can disable “Improve the model for everyone,” but that setting should not be treated as an API zero-data-retention commitment. Its Open Platform Terms place responsibility for permissions, consent, and downstream disclosure on API developers. As of August 13, I found no new GA-specific safety report or explicit public zero-data-retention commitment accompanying 0813. Enterprises should confirm their own contractual terms rather than extrapolate from the open-weight license.
Today most routers examine task complexity, latency, and price. DeepSeek adds the clock. Expect agent queues to tag jobs with urgency and automatically choose between immediate Flash, immediate Pro, or deferred Pro. Observability vendors will add “effective model price at execution time” beside token counts.
Publishing a score without the agent loop is increasingly like publishing a database benchmark without the query plan. DeepSeek Harness is unfinished, but its existence pressures competitors to reveal prompts, tools, retry rules, effort settings, and filesystem assumptions. Buyers should reward that transparency while still demanding independent replication.
The stable deepseek-v4-pro name is convenient, but it means an existing deployment changed behavior without a configuration change. Versioned 0813 weights give sophisticated teams a reference point. I expect more providers to expose resolved model revisions in response metadata, because “same alias, different agent” is otherwise painful to debug.
DeepSeek V4 Pro 0813 is a serious GA release, but “new frontier king” is the wrong summary.
The base architecture is largely familiar. The important improvement is agent discipline: the ability to navigate repositories, use tools, sustain terminal work, and finish multi-step tasks. Official scores show enormous repair versus Pro Preview, while independent results and the narrower gap to Flash argue against routing every request to Pro.
The exact weights are available, the harness is inspectable, and the Codex path reduces switching friction. Those are real wins. The Responses API is still partial, the model is text-only, fresh independent evidence is thin, and self-hosting roughly 893 GB of checkpoint data is infrastructure work.
Most of all, the price table changes the deployment decision. Beginning August 16, builders are not choosing only between Flash and Pro. They are choosing tier, effort, cache strategy, and execution time. Teams that treat those as routing inputs will get much more value than teams that treat Pro as a shiny replacement for Flash.
Use Flash to search. Use Pro to decide. Use max effort sparingly. Move queued hard work off-peak. Measure completed tasks, not benchmark aura.
It is the August 13, 2026 general-availability checkpoint of DeepSeek V4 Pro. It supersedes April’s Pro Preview and focuses on stronger agent, coding, repository, terminal, and tool-use performance.
Yes. DeepSeek says it is live in the app, web Expert Mode, and API. The stable API model name remains deepseek-v4-pro.
It is an MIT-licensed open-weight release. The checkpoint and encoding code are public, though “open source” can imply disclosure beyond weights and inference code. The repository contains roughly 893 GB of safetensors, so local deployment requires server-class infrastructure.
The hosted model supports a 1-million-token context window and up to 384,000 output tokens. DeepSeek recommends a large output allowance for high and max reasoning during local inference, but applications should still impose task-appropriate limits.
Under the pre-August 16 tariff, Pro costs $0.003625/M cached input tokens, $0.435/M uncached input tokens, and $0.87/M output tokens. From 16:00 UTC on August 16, off-peak rates become $0.022/$0.66/$1.98 and peak rates become $0.044/$1.32/$3.96.
Peak pricing applies daily from 01:00–04:00 UTC and 06:00–10:00 UTC. All other times are off-peak.
Use Flash as the default for volume, exploration, routine edits, and parallel workers. Escalate to Pro for difficult repairs, architecture, final review, long dependency chains, and tasks where failure is expensive. Validate that policy on your own repositories.
Configure DeepSeek as a custom Codex provider using its official model catalog, https://api.deepseek.com/ base URL, and wire_api = "responses"; select deepseek-v4-pro as the model. Compatibility is incomplete: persistent response state, background mode, images, files, computer use, MCP, and several other features are unsupported or ignored.
Yes, but not on a typical workstation. The official checkpoint is roughly 893 GB, and DeepSeek’s vLLM example uses one four-GB300 node. Smaller or alternative hardware plans still require expert-parallel inference engineering and careful context-capacity tradeoffs.
No. V4 Pro 0813 is text-only. Image and document inputs are not natively processed through the current API.
It depends on the task. DeepSeek’s table is highly competitive on Terminal Bench, Cybergym, and AutomationBench, but it does not lead every benchmark. In Artificial Analysis’s August 13 snapshot, V4 Pro scored 53, while Kimi K3 scored 60 and Fable 5 scored 62. These aggregates are not universal rankings. Pro’s strongest argument is price/performance, open weights, long context, and agent integration—not blanket dominance.
The setting changes how much deliberation the model performs before answering. Low is appropriate for cheap, reversible work; high is the default general agent mode; max should be reserved for hard tasks where extra latency and output are justified. Benchmark it per workflow rather than assuming more reasoning always helps.