The High-Efficiency Technical Specialist
DeepSeek V4 combines a one-million-token context, thinking and non-thinking modes, open weights and prices that are still low but no longer extraordinary. The peak and off-peak policy that landed in August raised every published rate: V4-Flash output went from a flat $0.28 per million tokens to $0.66 off-peak and $1.32 at peak. Two things changed in the second half of August. A third model shipped on August 21, 2026, `deepseek-v4-flash-vision-exp`, which adds image input at V4-Flash rates. And the peak window now excludes weekends, so Saturday and Sunday are off-peak all day. It remains a serious engineering option, but the cost argument is materially weaker than it was in early August, and choosing it still means making explicit decisions about hosting, data jurisdiction, evaluation and operational ownership.
Strong capability per dollar, with a different operating model
DeepSeek should be evaluated as infrastructure, not as a cheaper consumer chatbot. Its value depends on task routing, deployment choice, model evaluation and whether the organization can manage the distinct legal and operational context.
Technical, efficient and configurable
DeepSeek is designed for teams comfortable choosing modes, APIs, weights and deployment architecture rather than relying on a single polished end-user product.
- Thinking and non-thinking modes
- Agent and coding orientation
- Open weights and hosted API
High-volume technical work
Coding, tool use, document processing, long-context analysis and workloads where token economics change the business case.
- Route simple work to Flash
- Use Pro for harder cases
- Exploit caching and concurrency
Control and economics justify ownership
Use it when your team can evaluate quality, design safeguards and operate the provider or self-hosted path responsibly.
- Benchmark on your own tasks
- Resolve data and jurisdiction questions
- Budget for model operations
A third model, and a peak window that now skips the weekend
Two things changed since the August 19 edition, and both matter to a budget. The lineup is no longer two models: `deepseek-v4-flash-vision-exp` shipped on August 21, 2026 and brings image input at V4-Flash rates, so the “only two model IDs” line this guide carried is now wrong. And the pricing page has narrowed the peak window: it reads “Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday (all other hours are off-peak),” which makes Saturday and Sunday off-peak for the full day. Prices themselves are unchanged from the August reading. V4-Pro remains generally available since August 13, 2026, all three models carry 1M context and 384K maximum output, and neither retired alias appears on the pricing page.
DeepSeek V4-Flash Efficient default
Thinking and non-thinking modes, designed for speed, economics and simple agent work. The April 24 news entry described the preview as 284B total and 13B active params; DeepSeek’s weights channel states 304B params for the `DeepSeek-V4-Flash-0731` release, and the pricing page now names that same build. The two channels finally agree on which build is current.
- `deepseek-v4-flash`
- 1M context
- 2,500 listed concurrency
DeepSeek V4-Pro Generally available
DeepSeek announced the GA release on August 13, 2026, stating that “the GA release of DeepSeek-V4-Pro has been rolled out on the APP, Web, and API,” and that the “API calling method remains unchanged,” with the model name simply set to `deepseek-v4-pro`. The April 24 preview article gave 1.6T total and 49B active params; the weights channel now lists 1.7T params for the `DeepSeek-V4-Pro-0813` build, which the pricing page also names.
- `deepseek-v4-pro`
- No integration change for GA
- 500 listed concurrency
DeepSeek V4-Flash-Vision-Exp New · August 21, 2026
The third model in the lineup and the first with image input. DeepSeek says it “matches DeepSeek-V4-Flash on text capabilities, including agents, reasoning, and world knowledge,” and describes its pure-text capability as “on par with the official DeepSeek-V4-Flash.” Images are tokenized at up to 384 tokens each and charged at V4-Flash rates, so vision costs you the same per token as text. It works with the Chat Completions, Messages and Responses APIs. Images can be supplied as base64 or as URLs, or uploaded through a newly launched Files API that DeepSeek states is “available at no cost” and that lets you reuse one image across many requests.
- `deepseek-v4-flash-vision-exp`
- Up to 384 tokens per image, billed at V4-Flash rates
- Free Files API for image reuse
- Vendor benchmark claims: Terminal Bench 2.1 at 83.9 and Chartography at 64.3, unverified independently
Dual API formats
The hosted service exposes an OpenAI-compatible endpoint at `https://api.deepseek.com` and an Anthropic-compatible endpoint at `https://api.deepseek.com/anthropic`, which auto-maps Claude model names: Opus to `deepseek-v4-pro`, Sonnet and Haiku to `deepseek-v4-flash`, and anything unmapped to `deepseek-v4-flash`.
- Lower migration friction
- Explicit base URLs
- Auto-mapping means an unrecognized name silently lands on Flash
Thinking control
Both models support thinking and non-thinking modes, selected with `{"thinking": {"type": "enabled"}}` or `"disabled"` in `extra_body`. Thinking is enabled by default. The August 13 release adds published effort levels: DeepSeek describes “flexible reasoning effort for V4-Pro & V4-Flash: low for simple tasks, high for daily Agent workflows, max for complex tasks.” The pricing page itself does not mention effort levels, so do not expect a per-level rate.
- Thinking is the default, so opt out deliberately
- Low, high and max effort are now documented
- Effort is not a separate line on the price list
Open weights Released
DeepSeek published V4 weight collections and technical material, enabling self-managed deployment and research.
- Hosting control
- Model operations responsibility
- Hardware and quantization planning
Legacy aliases: retired Executed
DeepSeek retired `deepseek-chat` and `deepseek-reasoner` on schedule at 2026/07/24 15:59 UTC. The pricing page no longer carries the deprecation notice and neither alias appears anywhere on it. Anyone still sending those model strings is calling a dead API, so migrate to `deepseek-v4-flash` or `deepseek-v4-pro`.
- Pin explicit V4 IDs
- Update tests and configuration
- Legacy strings now fail rather than degrade
Peak and off-peak pricing Weekends now off-peak
The pricing page now reads: “Off-peak rates are half of the peak rates. Peak hours are 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday (all other hours are off-peak).” The “Monday through Friday” qualifier is new since the August 19 reading and it changes the scheduling advice: a batch job moved to a Saturday is off-peak for the whole twenty-four hours, not just outside the seven weekday windows. Note also what the page no longer says. It carries no statement of when peak and off-peak pricing took effect. The 16:00 UTC, August 16, 2026 time this guide has quoted survives only in the August 13 changelog entry, as “pricing adjustments effective August 16, 2026,” without the hour.
- Peak is thirty-five hours a week, not forty-nine
- Weekends are off-peak end to end
- The exact 16:00 UTC start time is no longer confirmable on any current page
Builds are named in the docs Gap stays closed
The pricing page names `DeepSeek-V4-Flash-0731`, `DeepSeek-V4-Pro-0813` and now `DeepSeek-V4-Flash-Vision-Exp` directly, and the weights channel carries the same names. On Hugging Face the 305B `DeepSeek-V4-Flash-Vision-Exp` is now the newest release, ahead of `DeepSeek-V4-Pro-0813`. Docs and weights channel remain in sync on build identity.
- All three build strings are documented, not only inferred
- The vision build is the newest weights release at 305B params
- Still pin the model ID and re-evaluate on a schedule
Responses API coverage
The gap closed on schedule and has since widened to cover the new model. Responses API support now spans all three model IDs, and the August 21 release states the vision model works with the Chat Completions, Messages and Responses APIs. The August 13 release also claims “Native OpenAI Responses API support, optimized for Codex with one-click setup.”
- All three models are supported
- The early-August promise was kept
- Codex setup is a vendor claim, so test it
| Model and rate window | Input: cache hit | Input: cache miss | Output | Concurrency |
|---|---|---|---|---|
| V4-Flash · off-peak | $0.007 / 1M | $0.22 / 1M | $0.66 / 1M | 2,500 |
| V4-Flash · peak | $0.014 / 1M | $0.44 / 1M | $1.32 / 1M | 2,500 |
| V4-Flash-Vision-Exp · off-peak | $0.007 / 1M | $0.22 / 1M | $0.66 / 1M | Not separately listed |
| V4-Flash-Vision-Exp · peak | $0.014 / 1M | $0.44 / 1M | $1.32 / 1M | Not separately listed |
| V4-Pro · off-peak | $0.022 / 1M | $0.66 / 1M | $1.98 / 1M | 500 |
| V4-Pro · peak | $0.044 / 1M | $1.32 / 1M | $3.96 / 1M | 500 |
Use the least expensive mode that meets the task’s failure cost
DeepSeek’s product design rewards routing, and since the peak and off-peak split arrived in August it rewards scheduling too. Flash versus Pro versus the vision build, thinking versus non-thinking, and weekday peak versus everything else should all be evaluated policy rather than user preference.
Run this task suite on V4-Flash non-thinking, V4-Flash thinking and V4-Pro thinking. Compare exact-task success, latency, output tokens, retries and human correction time. Recommend a routing policy.
Complete this repository task with a maximum of 12 tool steps. Stop before any external network or deployment action. Report tests, files changed and unresolved risk.
Hosted API and open weights solve different problems
Open weights improve control but do not remove cost. They replace per-token pricing with infrastructure, model serving, security, monitoring and lifecycle work.
| Dimension | Hosted API | Self-managed weights |
|---|---|---|
| Time to start | Fast | Infrastructure project |
| Variable cost | Published token pricing | Compute, staffing and utilization |
| Data control | Provider architecture and terms | Organization controls environment |
| Scaling | Provider-managed within limits | Organization provisions capacity |
| Model updates | Provider-led | Organization tests and deploys |
| Observability | API-level signals | Full stack if implemented |
| Operational burden | Lower | Higher |
Context caching
Repeated prefixes can reduce input cost dramatically, making prompt architecture part of the economics.
- Keep stable instructions stable
- Measure actual cache hit rate
- Do not assume every request qualifies
User isolation
The API supports a user identifier for content-safety, KV-cache and scheduling isolation.
- Do not place personal data in the identifier
- Use stable internal pseudonyms
- Understand account-level limits
Concurrency planning
Published account-level limits differ between Pro and Flash and can create throughput architecture decisions.
- Model peak demand
- Handle 429 responses
- Request capacity expansion when justified
Off-peak scheduling
The clock is a cost lever: off-peak rates are half of peak, and peak is 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday. Batch, backfill and evaluation runs that do not need to be interactive belong in the other seventeen hours of a weekday, or anywhere at all on a Saturday or Sunday.
- Queue deferrable work for off-peak
- A weekend run is off-peak for the full day
- Track the UTC windows, not local time
- Interactive weekday traffic will still land at peak
Thinking inside tool use is useful, and needs hard boundaries
DeepSeek’s agent direction began with V3.2 and continues with V4. The operational risk is an efficient model taking many inexpensive wrong steps before anyone notices.
Research this technical decision using only the supplied documentation and approved web domains. Maximum 10 tool calls. Cite every conclusion and stop if the sources conflict.
Implement the change, run the tests and inspect the diff. Maximum 15 tool steps. Do not modify deployment, credentials or network configuration. If blocked, report the exact evidence.
Return the result as the required JSON schema, then run an independent validation pass that checks completeness, allowed values and source coverage.
After two verification passes, commit to the strongest supported answer. Do not continue rechecking unless new evidence appears.
Nine habits for extracting the economic advantage safely
DeepSeek rewards engineering discipline: explicit routing, structured outputs, hard agent limits, scheduling against the new peak windows, and evaluation on the organization’s real tasks.
Pin explicit V4 IDs
The retirement executed at 2026/07/24 15:59 UTC. `deepseek-chat` and `deepseek-reasoner` are gone; any call still using them is hitting a dead API. Send `deepseek-v4-flash`, `deepseek-v4-pro` or `deepseek-v4-flash-vision-exp`.
Route by task
Flash for throughput; Pro when measured quality gains justify it.
Use non-thinking deliberately
Do not pay reasoning latency for transformations that schemas can validate.
Exploit stable prefixes
Design repeated context to benefit from caching without hiding changing instructions.
Set hard agent limits
Maximum steps, budget, tools and stop conditions belong in the task contract.
Evaluate cultural and political domains
Test behavior on topics relevant to your users and markets, not only code benchmarks.
Separate hosted and self-managed risk
The same model can create different legal and security profiles by deployment.
Measure correction cost
Include retries and human review when comparing providers.
Schedule deferrable work off-peak
Off-peak is half of peak. Peak is 01:00 - 04:00 and 06:00 - 10:00 UTC, Monday through Friday, so move batch and evaluation runs out of those windows or onto a weekend, which is now off-peak all day.
The risk profile belongs in the architecture decision
Technical capability does not answer questions about jurisdiction, hosted-service data handling, political behavior, licensing or operational responsibility.
Data jurisdiction
Hosted API use may not meet every sovereignty, contractual or regulatory requirement.
Political sensitivity
Evaluate behavior on politically sensitive and culturally specific topics relevant to your markets.
Provider concentration
Do not build a critical system without model abstraction, fallback and exit planning.
Open-weight operations
Self-hosting transfers patching, serving, scaling, safety and monitoring to the organization.
Long-context confidence
A 1M window does not guarantee full coverage, reliable retrieval or balanced attention.
Agent loops
Cheap tokens can make runaway verification or tool loops inexpensive but still operationally harmful.
Alias retirement executed
The 2026/07/24 15:59 UTC retirement of `deepseek-chat` and `deepseek-reasoner` has happened. Neither alias is listed on the pricing page any more. This is no longer a migration deadline to plan for. It is a live outage for anything that still sends those strings.
Weights can move under a fixed model ID
Credit where it is due: DeepSeek has started naming builds where buyers can see them. The pricing page now identifies `DeepSeek-V4-Flash-0731` and `DeepSeek-V4-Pro-0813`, matching the weights channel, and the changelog is publishing again after three months of silence. The underlying caution survives the improvement. A model ID is still an alias over weights that can be replaced, naming a build on a live page is not the same as a versioned endpoint you can pin to, and the page can be updated again without anyone telling you. Re-run your evaluation suite on a schedule, not only on a version bump.
Dated announcements are not live status
This edition is the third illustration in a row, and this time the drift runs the other way. The August 13 post announced new pricing in the future tense, “New pricing takes effect at 16:00 UTC, Aug 16, 2026,” and earlier archival pages still said the retired aliases “will be discontinued” long after they had gone. Now the pricing page has dropped its effective-date statement altogether, so a fact this guide could verify on August 19 can only be cited from a changelog entry today, and only to the day rather than the hour. Read the live pages for what is true now, confirm against the API, treat every dated post as a description of the day it was written, and archive anything you will need to prove later.
The price already moved once, and the rules moved again
The peak and off-peak policy this guide flagged as pending in early August went live thirteen days later and raised every billing item: off-peak is half of peak, and off-peak alone is more than double the previous flat output rates. Then the definition of peak changed again, narrowing to 01:00 - 04:00 and 06:00 - 10:00 UTC on weekdays only. Two policy moves in quick succession, though only the first one is dated: the weekday qualifier was absent from the August 19 reading and present on September 3, and DeepSeek dates its arrival nowhere, so it landed somewhere in that fifteen-day window. That second change happens to be in your favor, which is exactly why it is easy to miss. A vendor that can reprice its whole list on three days’ notice, and then quietly redraw the billing calendar, can do either again. Budget with headroom, keep an abstraction layer, and re-read the pricing page before any commitment that assumes a number from this guide.
A model named Exp
The new vision model ships as `deepseek-v4-flash-vision-exp`, and the weights build is `DeepSeek-V4-Flash-Vision-Exp`. DeepSeek publishes no stability guarantee, support window or deprecation policy alongside it that this guide could find. The benchmark figures it cites, Terminal Bench 2.1 at 83.9 and Chartography at 64.3, are vendor claims with no independent replication. Prototype against it freely; think twice before a customer-facing dependency rests on a build whose own name says experimental.
| Risk decision | Hosted API question | Self-managed question |
|---|---|---|
| Data | Where is content processed and retained? | Who can access model inputs, logs and infrastructure? |
| Security | What provider and network controls apply? | How are serving stack and weights protected? |
| Safety | What provider policies and isolation apply? | What filters, evaluation and abuse controls will you operate? |
| Reliability | What are limits, SLAs and fallback options? | How will you scale and recover? |
| Lifecycle | How are provider upgrades communicated? | Who owns model testing and rollout? |
First-party evidence behind this guide
These DeepSeek documentation, news, pricing and weights pages anchor the current V4, price, API and migration claims. All were re-checked on September 3, 2026. The pricing page is the evidence for the peak and off-peak rates, for the new Monday-through-Friday peak window and for the build strings shared with the weights channel. The August 21 release article is the evidence for the vision model, its image tokenization, its pricing parity with V4-Flash and the free Files API. The August 13 release article remains the evidence for V4-Pro general availability, and it is now the only surviving source for the 16:00 UTC, August 16 effective time, which the pricing page no longer states. August 21 is the most recent changelog entry; nothing has been published after it.
AI Mindset