AI Mindset · Model Cheatsheets
DeepSeek

The High-Efficiency Technical Specialist

DeepSeek still owes you an ending for V4-Pro. Its September 10 release article routes every deepseek-v4-pro request to V4.1-Flash from September 14, until V4.1-Pro launches. V4.1-Pro hasn’t launched. As of October 15, 2026, Pro is still priced and still served, and the change log adds only that DeepSeek will give further notice of any change. So the phase-out is a bridge with no far end. Meanwhile the lineup is two models, deepseek-flash and deepseek-v4-pro. Thinking is on by default at high effort, and the request formats spell the off switch differently. All ten V4-family weight repositories are MIT. And prices moved twice in under a month.

Verified October 15, 2026V4.1-Flash · V4-Pro · V4.1-Pro unshipped1M context · 384K outputPrices cut September 10, 2026MIT weights · thinking on by default
1 / Meet DeepSeek

Strong capability per dollar, with a different operating model

Judge DeepSeek as infrastructure, not as a cheaper consumer chatbot. Its value rides on task routing, deployment choice and model evaluation. And on one human question: can your organization manage the legal and operational context that comes attached? Three working defaults follow. Re-price the workload against the September 10, 2026 rates, not the August rates. Default to deepseek-flash, because the vendor no longer claims Pro performs better. And remember that peak applies Monday through Friday, minus Chinese public holidays, which DeepSeek doesn’t list.

Personality

Technical, efficient and configurable

It’s built for teams that are comfortable choosing modes, APIs, weights and deployment architecture. Not for teams that want one polished end-user product.

  • Thinking and non-thinking modes
  • Agent and coding orientation
  • MIT-licensed weights and a hosted API
Deploy it for

High-volume technical work

Coding, tool use, document and image processing, long-context analysis, and any workload where token economics change the business case.

  • Route almost everything to deepseek-flash
  • Raise thinking effort before changing model
  • Exploit caching, concurrency and the off-peak clock
Choose DeepSeek when

Control and economics justify ownership

Use it when your team can evaluate quality, design safeguards and run the hosted or self-hosted path responsibly. And when processing in the People’s Republic of China under PRC law is acceptable for the data involved.

  • Benchmark on your own tasks
  • Resolve data and jurisdiction questions first
  • Budget for model operations

Is DeepSeek the right strategic choice?

Choose the requirement that’s driving the decision.

API economics
Re-price against the September 10 rates

deepseek-flash is the cheapest way into the lineup. It’s also got the highest listed concurrency at 2,500, and it’s the only model with image input. Those rates took effect at 04:00 UTC on September 10, 2026 and replaced the August peak and off-peak list. So any budget built on an earlier price list is out of date. The August figures can’t be checked either, because DeepSeek published them only as a chart image. The numbers to budget against: output at $0.60 per million tokens off-peak and $1.20 at peak, cache-miss input at $0.15 and $0.30, cache hit at $0.003 and $0.006. One catch on the off-peak half. Peak excludes Chinese public holidays, and DeepSeek publishes no list of which days those are. So your off-peak share is a range, not a figure.

  • Model quality on your data, not the headline rate
  • Cache-hit and off-peak scheduling assumptions
  • Peak is weekdays only, and Chinese public holidays are off-peak in full
  • Total workflow cost, not token price alone
2 / What’s Current

A new default model, a price cut, and two retirements

Everything about the lineup changed on September 10, 2026. DeepSeek shipped DeepSeek-V4.1-Flash, retired two model IDs, and cut prices with effect from 04:00 UTC that day. The two retired names are still accepted, but in DeepSeek’s own wording they’re temporarily routed to V4.1-Flash and billed at the Flash price, with no stated end date. Image input is native to the default model. Pro was supposed to stop taking traffic on September 14. It didn’t. As of October 15, 2026, it’s still priced, still rate-limited, still monitored and still the model a claude-opus string buys. The release article conditions the routing on a successor, V4.1-Pro, which hasn’t shipped. Both current models carry 1M context and 384K maximum output.

Half price
Off-peak rate. Peak is 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday. Chinese public holidays are off-peak
$0.60 / $1.20
deepseek-flash output per million tokens, off-peak and peak
$1.98 / $3.96
deepseek-v4-pro output per million tokens, off-peak and peak
04:00 UTC
The hour the current price list took effect, September 10, 2026

DeepSeek-V4.1-Flash New · September 10, 2026

The new default, and the only model in the lineup that accepts images. The architecture lines below are vendor claims, not measurements.
· New: deepseek-flash. Retired: deepseek-v4-flash and deepseek-v4-flash-vision-exp. Continuing: deepseek-v4-pro. Promised and unshipped: V4.1-Pro.
· DeepSeek describes a 552B-parameter mixture-of-experts model on a new causal encoder and decoder architecture. It splits a 40-layer Transformer into a 20-layer causal encoder and a 20-layer decoder, with just 8B active parameters for input and 16B for output.
· It claims the KV cache needs a quarter of the HBM and an eighth of the SSD storage of the previous generation, and that multimodality is native through DeepSeek-ViT, trained from scratch.
· The weights channel lists a 552B backbone and 763B total parameters under the MIT license.

  • deepseek-flash
  • 1M context, 384K maximum output
  • Image input supported, 2,500 listed concurrency
  • MIT-licensed weights

DeepSeek V4-Pro Still serving, past its stated date

The 0813 build reached general availability on August 13, 2026. It’s still fully served on October 15, 2026, and you can check that in four places at once rather than taking one page’s word for it. What it no longer is, on DeepSeek’s own account, is the stronger model. The September 10 article says several parties measured V4.1-Flash ahead of it on performance, cost, speed and total runtime. Pro costs 3.3 times Flash on output, 4.4 times on cache-miss input and 7.3 times on cache-hit input. It carries 500 listed concurrency against Flash’s 2,500, and the pricing page marks image input as not supported for it. That missing vision is now its only distinguishing feature.
· Checked October 15, 2026: a full price row on the pricing page, 500 concurrency on the Rate Limit page, a named model on the docs homepage, and a monitored component on the status page.

  • deepseek-v4-pro
  • No image input, 500 listed concurrency
  • 3.3x Flash on output, 4.4x on cache-miss input, 7.3x on cache-hit input
  • Read the September 14 card before you depend on it

September 14 came and went Not executed

DeepSeek set a hard date for moving Pro traffic away, and then nothing happened on it. As of October 15, 2026, every Pro request still reaches Pro and bills at the Pro rates. The sentence that says service continues with billing unchanged sits on the change log, in English and in Chinese. Both are pages DeepSeek rewrites in place. No date on them, no history behind them, and the pricing page doesn’t repeat it. DeepSeek adds that it will give further notice of any change. It names no channel and no lead time. So it’s a reading, not a commitment.
· September 10 release article: starting at 04:00 UTC on September 14, 2026 all deepseek-v4-pro requests will route to V4.1-Flash at V4.1-Flash rates, and this will continue until V4.1-Pro launches. The same article says Pro is being phased out.
· Chinese release article on deepseek.com, September 10, 2026: DeepSeek plans an orderly retirement of the V4 Pro model, with the same routing from 12:00 Beijing time on September 14, the same instant.
· Change log, English and Chinese, checked October 15, 2026: in response to user demand, DeepSeek continues V4 Pro API service after September 14, 2026, with the billing method unchanged. It will provide further notice should there be any changes.

  • The routing didn’t happen on the stated date
  • The continuation sentence lives only on rewritable pages
  • Further notice is promised, with no channel or lead time
  • A Pro dependency needs written confirmation, not a page reading

V4.1-Pro is named and unshipped Unshipped

DeepSeek tied Pro’s ending to a model it hasn’t released. The routing runs until V4.1-Pro launches, in the article’s own words. So the phase-out that reads like a shutdown is a hand-off waiting on a product, and the product isn’t there. No price row. No docs entry. No change-log entry. No published weights. The planning consequence is the awkward one. You can’t move onto the successor, and you can’t put a date on the end of the model it replaces.
· Checked October 15, 2026: no V4.1-Pro on the pricing page in English or Chinese, the docs homepage, the change log whose newest entry is September 10, the deepseek.com news index, DeepSeek’s weights channel or its GitHub organization.

  • Announced in the September 10 release articles
  • No pricing, docs, change-log or weights entry
  • Pro’s routing is conditioned on it launching
  • Plan a migration you cannot yet schedule

Retired: V4-Flash and Vision-Exp Executed

Two model IDs went out on September 10, 2026. DeepSeek states that the corresponding models have been retired, and that their requests are served by the DeepSeek-V4.1-Flash model and billed at the Flash price. The change log says the two names are temporarily routed for compatibility. Temporarily is DeepSeek’s word, and it comes with no end date. The July retirement still stands separately.
· Retired September 10, 2026: deepseek-v4-flash and deepseek-v4-flash-vision-exp.
· deepseek-chat and deepseek-reasoner: no longer listed on the pricing page or the docs model list. DeepSeek’s April 24 notice set the shutdown for 2026/07/24 15:59 UTC.

  • Pin deepseek-flash or deepseek-v4-pro
  • A temporary route isn’t a supported model ID
  • Legacy deepseek-chat and deepseek-reasoner are off the price list
  • Update tests, configuration and cost attribution

Vision is native, and the limits are hard

The old routing decision has disappeared. You no longer send images to a separate experimental build. The token cap gets quoted everywhere, and it’s the least likely thing to break you. The size, count and pixel ceilings are what stop a document pipeline cold, usually on the page nobody tested. Count your pages and weigh your payload before you design the batch.
· Vision guide, checked October 15, 2026: 1024 tokens per image after automatic resizing. JPEG, PNG, GIF and WebP, detected from the file content rather than the name or MIME type. 32 MiB per image by base64 or URL, 64 MiB through the Files API. 600 images per request. 8,192 characters maximum for a URL. 48 MiB for the whole request body. Total image size of 64 MiB per request, or up to 200 MiB when images come by file_id. 8,192 pixels per side, dropping to 4,096 once a request holds 15 or more images.
· Pricing page, checked October 15, 2026: image input supported for deepseek-flash, not supported for deepseek-v4-pro.

  • 1024 tokens per image, 600 images per request
  • 4,096 pixels per side once a request holds 15 images
  • Body, image-total and per-image caps all bind
  • Files API storage has no published price

Files API: kept until you say otherwise Permanent by default

Upload a file to DeepSeek and it stays. Permanently, unless the upload said otherwise. That’s the default most teams never notice, because nobody reads the expiry field on an upload that works. And the storage sits in the People’s Republic of China with the rest of the hosted service. So set an expiry on every upload, as policy, not as a tidy-up. One more gap. No DeepSeek page prices file storage. Unpublished isn’t free.
· Files API guide, checked October 15, 2026: omit both expires_after fields to keep the file permanently. A custom lifetime runs from 3,600 to 2,592,000 seconds, one hour to 30 days. JPEG, PNG, GIF and WebP. 64 MiB per upload, 512-character filenames, 25 GiB and 10,000 files per user. An upload must complete within 10 minutes. Upload through POST /files or POST /anthropic/v1/files.

  • Default retention is permanent
  • Set expires_after on every upload
  • 25 GiB and 10,000 files per user
  • No published price for storage

Dual API formats

Two request formats, one service. Auto-mapping is convenient, and it also means a Claude model string sitting in your configuration quietly decides which DeepSeek model you pay for.
· OpenAI-compatible endpoint at https://api.deepseek.com. Anthropic-compatible endpoint at https://api.deepseek.com/anthropic.
· Anthropic API guide, checked October 15, 2026: model names starting with claude-haiku or claude-sonnet are mapped to deepseek-flash. claude-opus maps to deepseek-v4-pro and is billed at the V4 Pro price.

  • Lower migration friction
  • Explicit base URLs
  • claude-haiku and claude-sonnet land on deepseek-flash
  • claude-opus lands on Pro and bills at the Pro price

Thinking control, and the trap in it

Here’s the part that breaks code copied between formats. Thinking is on by default, at high effort. So a job that never touched the setting is paying for reasoning, right? Three request formats, three spellings for effort, and two different ways to turn thinking off. Write the setting per format, and write it explicitly. The pricing page lists no per-effort rate, so effort moves your token count, not your rate. And DeepSeek quietly remaps effort names borrowed from other vendors. Port code that says medium and it runs at high.
· Thinking Mode guide, checked October 15, 2026: thinking mode is enabled by default, with the default effort being high. Effort names map minimal and low to low, medium, high and xhigh to high, and max and ultra to max.
· OpenAI-compatible format, the default base URL: reasoning_effort takes low, high or max. A separate thinking object set to disabled turns thinking off.
· Anthropic-compatible format, per the Anthropic API guide: a thinking object turns thinking on or off, and budget_tokens is ignored. Effort goes in output_config.effort, the only output_config field supported, and takes low, high or max.
· Responses API, per the Responses API guide: effort goes in reasoning.effort. The Thinking Mode guide gives that parameter none, low, high or max, with none disabling thinking.
· The Thinking Mode guide’s parameter table, in English and Chinese, doesn’t tie each parameter to one format, so the per-format names rest on the Anthropic API and Responses API guides.

  • Default: thinking on, effort high
  • OpenAI and Anthropic formats switch off with a thinking object
  • none belongs to the Responses API’s reasoning.effort
  • medium silently runs as high

Responses API coverage

Two DeepSeek pages disagree on which models it serves. The Responses API guide names only deepseek-flash. The pricing page’s feature table marks the Responses API as supported for both deepseek-flash and deepseek-v4-pro. Standardized on the Responses API and also depend on Pro? Test it, then get it in writing. Two more traits bite here. This surface ignores parameters it doesn’t support, silently, with no error. So an effort field written for another format quietly does nothing, and you get the default: thinking on, effort high. And it takes no file inputs at all. The August 13 release article also claims native Responses API support optimized for Codex, with one-click setup. That’s a vendor claim to test.
· Responses API guide, checked October 15, 2026: model support names deepseek-flash. reasoning is partially supported, with effort supported and summary accepted but no summary generated. Unsupported parameters are silently ignored. File inputs are not supported, and an image in a system or assistant message returns a 400.
· Pricing page, checked October 15, 2026: Responses API marked supported for both models.

  • The guide names only Flash, the pricing page lists both
  • Unsupported parameters are dropped with no error
  • Effort is reasoning.effort on this surface
  • No file inputs on this surface

Peak and off-peak pricing Holiday carve-out

Thirty-five hours a week is the ceiling now, not the number. Peak used to be every weekday hour inside the two windows. It now excludes Chinese public holidays, and DeepSeek neither lists those days nor links a calendar for them. So you can bound your peak exposure from above, and you can’t compute it from first-party material. National Day week, October 1 to 7, 2026, was the first long holiday under the rule. DeepSeek said nothing about how it billed, on the pricing page, the change log, the docs homepage, deepseek.com or the status page. Both price regimes do carry a published effective hour, which is the only reason either can be dated: the pricing page itself has no last-updated stamp.
· Pricing page, checked October 15, 2026: off-peak rates are half of the peak rates. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, excluding Chinese public holidays. All other hours are off-peak, including weekends and Chinese public holidays in full.
· Chinese pricing page, checked October 15, 2026: the carve-out reads 法定节假日, statutory holidays. It doesn’t say whether adjusted rest days count, or how a make-up working day on a weekend is billed.
· August 13 release article: the peak and off-peak regime took effect at 16:00 UTC on August 16, 2026. September 10 release article: the current rates took effect at 04:00 UTC on September 10, 2026, which is 24 days and 12 hours later.

  • Thirty-five hours a week is an upper bound
  • Weekends are off-peak end to end
  • Chinese public holidays are off-peak and aren’t listed
  • No word on how National Day week was billed

Open weights, MIT licensed Released

DeepSeek publishes its V4 weights and technical material on its own weights channel under the MIT license. That permits commercial use, modification and redistribution with attribution, and it’s much more permissive than the custom community licenses common elsewhere. All ten V4-family repositories carry it, both DSpark builds included. DeepSeek describes DSpark as the same checkpoint with a speculative decoding module attached, not a new model. One oddity worth knowing. The retired DeepSeek-V4-Flash-Vision-Exp repository is still published on October 15, 2026, 35 days after the API retired the model on September 10. So the weights channel and the API don’t move together. And self-hosters get a different effort dial. The V4.1-Flash model card describes reasoning effort as an integer from 1 to 100, with its instruct benchmarks run at 100. That isn’t the API’s low, high and max, so hosted and self-hosted scores need care before you compare them.
· Weights channel, checked October 15, 2026: V4.1-Flash at 763B total, V4-Flash-Vision-Exp at 305B, V4-Pro-0813 at 1.7T, V4-Flash-0731 at 304B, V4-Flash-DSpark at 165B, V4-Pro-DSpark at 1.7T, V4-Pro at 1.6T, V4-Flash at 291B, V4-Flash-Base at 292B and V4-Pro-Base at 1.6T.
· MIT on all ten. The two Base repositories show no model card and no license tag, and each holds a LICENSE file that opens with MIT License.

  • MIT on all ten V4-family repositories
  • DSpark adds speculative decoding to the same checkpoint
  • Self-hosted effort runs 1 to 100, not low, high and max
  • A retired API model can still be self-hosted

Service status

DeepSeek publishes a status page with per-service uptime. Those are the vendor’s own figures, on a live page rewritten continuously, over a period the page doesn’t state. So they’re a snapshot, not a commitment, and no SLA sits behind them. Read the layout right, too. Chat Service is a group of three components, Chatservice, File Upload and Search. The group figure and the Chatservice figure sit side by side and differ, because one contains the other.
· Checked October 15, 2026: uptimes between 99.53% and 100.00% across the V4 Pro API, the V4.1 Flash API and the Chat Service group, with no period stated.

  • Vendor-published uptime, not a contractual SLA
  • No stated period behind the percentages
  • Chat Service is a group of three components
  • The overall state changes by the hour, so read it live
Model and rate windowInput: cache hitInput: cache missOutputConcurrency
deepseek-flash (DeepSeek-V4.1-Flash) · off-peak$0.003 / 1M$0.15 / 1M$0.60 / 1M2,500
deepseek-flash (DeepSeek-V4.1-Flash) · peak$0.006 / 1M$0.30 / 1M$1.20 / 1M2,500
deepseek-v4-pro (DeepSeek-V4-Pro-0813) · off-peak$0.022 / 1M$0.66 / 1M$1.98 / 1M500
deepseek-v4-pro (DeepSeek-V4-Pro-0813) · peak$0.044 / 1M$1.32 / 1M$3.96 / 1M500
deepseek-v4-flash · retired, legacy aliasBilled at the Flash rate aboveBilled at the Flash rate aboveBilled at the Flash rate aboveServed by V4.1-Flash
deepseek-v4-flash-vision-exp · retired, legacy aliasBilled at the Flash rate aboveBilled at the Flash rate aboveBilled at the Flash rate aboveServed by V4.1-Flash
Model IDBuild namedContextMax outputImage inputThinkingConcurrency
deepseek-flashDeepSeek-V4.1-Flash1M384KSupportedSupported, on by default2,500
deepseek-v4-proDeepSeek-V4-Pro-08131M384KNot supportedSupported, on by default500
3 / Model Router

Use the least expensive mode that meets the task’s failure cost

The routing map got simpler on September 10. Images no longer route anywhere special, because the default model has vision. Hard work no longer routes to Pro, because the vendor no longer claims Pro performs better. What’s left is one model, an effort dial and a clock. Run deepseek-flash at the right effort, scheduled against the weekday peak windows. Keep Pro for the narrow cases that really call for it. And check which request format you’re on before you write the effort value. Each format spells it differently, and the Responses API drops a misspelled field without a word.

Route the workload

Select the dominant task property.

Everyday agent task
deepseek-flash · effort high

The right economic default. Flash with thinking at high covers planning, tool calling and coding where the model needs a deliberate plan. Its output costs under a third of Pro’s, and it has five times the listed concurrency. High is also DeepSeek’s published default. Set it explicitly anyway, because the page that says so is rewritten in place, and a field written for the wrong format can be dropped without an error.

  • Strong economic default
  • Matches the published default, so write it down anyway
  • Set maximum steps
  • Image input works here too, capped at 1024 tokens per image

Reasoning effort versus operational cost

Three active thinking levels, low, high and max, plus an off state. Thinking is on by default at high. The OpenAI-compatible format sets effort through reasoning_effort and turns thinking off with a separate thinking object. The Anthropic-compatible format sets effort through output_config.effort and uses the same kind of thinking object. The Responses API sets effort through reasoning.effort, where none turns thinking off according to the Thinking Mode guide. Raise effort only when the task actually benefits from deeper search or tool use.

ThroughputMaximum reasoning
Effort high
Everyday complex work

High effort fits daily agent workflows: analysis, tool calling and coding where the model needs a deliberate plan. It’s also what you get if you set nothing, because high is DeepSeek’s published default. Write it explicitly anyway.

  • The working default for agents
  • Good balance of cost and depth
  • Set completion criteria
  • Check whether your workload actually needs it
Routing evaluation
Run this task suite on deepseek-flash with thinking disabled, then at effort low, high and max, and on deepseek-v4-pro at high. State which request format you used and how you disabled thinking in it. Compare exact-task success, latency, output tokens, retries and human correction time. Recommend a routing policy and state the cost per successful task at both peak and off-peak rates.
Agent boundary
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.
4 / Deployment

Hosted API and open weights solve different problems

Open weights improve control. They don’t remove cost. They swap per-token pricing for infrastructure, model serving, security, monitoring and lifecycle work. What DeepSeek does remove is the licensing obstacle. All ten V4-family weight repositories are MIT, the two DSpark builds and the two Base checkpoints included.

Choose the deployment path

Start with the organizational requirement, not the ideology.

Fastest path to production
Hosted DeepSeek API

Use the official API when speed of integration, provider-operated serving and listed token economics are the priority. Accept that content is processed and stored in the People’s Republic of China, and that uploaded files are kept permanently unless you set an expiry. And accept that the same Open Platform terms cover individual and enterprise developers.

  • OpenAI or Anthropic request format
  • Provider data and jurisdiction review
  • Monitor rate and pricing changes, which moved twice in a month
DimensionHosted APISelf-managed weights
Time to startFastInfrastructure project
Variable costPublished token pricingCompute and staff, paid for even when idle
Data controlProcessed and stored in the PRC under PRC law. Uploaded files kept permanently unless an expiry is setOrganization controls environment
LicensingOpen Platform terms, no enterprise variant. Training other models on the outputs, including distillation, is explicitly permittedMIT on all ten V4-family weight repositories, as read October 15, 2026
ScalingProvider-managed within listed concurrency. Capacity expansion at no extra costOrganization provisions capacity
Model updatesProvider-led, and a model can retire on the day it’s announcedOrganization tests and deploys
ObservabilityAPI-level signalsFull stack if implemented
Availability promiseNone. The terms disclaim availability by jurisdiction and reserve broad suspension rightsOrganization owns uptime
Operational burdenLowerHigher

Context caching

Caching on disk is enabled by default for all users, with no code change. And the discount isn’t a published percentage. It’s the cache-hit column of the price list. A cache hit costs 2 percent of the cache-miss rate on deepseek-flash and 3.3 percent on deepseek-v4-pro. So prompt architecture is part of the economics, not a micro-optimization.

  • Keep stable instructions stable
  • Measure the real hit rate through prompt_cache_hit_tokens
  • Don’t assume every request qualifies

MIT-licensed weights

Permissive licensing removes one common blocker to a self-hosted path. And the Vision-Exp repo shows a second use. It’s still published on October 15, 2026, 35 days after the API retired the model, so a model DeepSeek has pulled from the API is still available for you to run.
· Ten V4-family repositories on DeepSeek’s weights channel, checked October 15, 2026, all MIT: V4.1-Flash, V4-Flash-Vision-Exp, V4-Pro-0813, V4-Flash-0731, V4-Flash-DSpark, V4-Pro-DSpark, V4-Pro, V4-Flash, V4-Flash-Base and V4-Pro-Base.

  • Commercial use, modification and redistribution with attribution
  • A retirement escape hatch for a model you depend on
  • The weights channel and the API move separately, so verify against both

User isolation

The user identifier isn’t just a label. On an account with expanded quota, each user_id gets its own concurrency allowance, 2,500 on deepseek-flash and 500 on deepseek-v4-pro. So the field is a capacity lever as well as a safety and cache boundary. Leave it empty and the empty value becomes its own user_id, which means every anonymous caller shares one bucket.
· Rate Limit page, checked October 15, 2026: the identifier must match the regex [a-zA-Z0-9\-_]+, with a maximum length of 512.

  • Don’t place personal data in the identifier
  • Use stable internal pseudonyms, 512 characters maximum
  • An empty identifier is itself one shared bucket
  • Expanded-quota accounts get the limit per user_id

Concurrency planning

Published account-level limits differ sharply: 2,500 for deepseek-flash and 500 for deepseek-v4-pro. Flash is cheaper, and on DeepSeek’s own account it measures better too. So the concurrency gap is one more reason the migration runs toward Flash rather than away from it. And if the cap pinches, ask. DeepSeek says there is no additional cost for capacity expansion, so don’t engineer around a limit you could have raised for free.
· Rate Limit page, checked October 15, 2026.

  • Model peak demand
  • Handle 429 responses
  • Request capacity expansion, which carries no extra cost

Off-peak scheduling

The clock is a cost lever, and it’s one notch less predictable than it looks. Off-peak runs at half the peak rate, and peak covers just seven weekday hours in two UTC windows. Batch, backfill and evaluation runs that don’t need to be interactive belong in the other seventeen hours of a weekday. Or anywhere at all on a Saturday or Sunday. The catch for a forward budget is that Chinese public holidays are off-peak too, and DeepSeek doesn’t publish the holiday list, so your peak share has a ceiling and no exact value. Build the forecast on the ceiling and treat the holidays as upside.
· Pricing page, checked October 15, 2026, for the window text and the holiday carve-out. No calendar is listed or linked.

  • Queue deferrable work for off-peak
  • A weekend run is off-peak for the full day
  • Track the UTC windows, not local time
  • Budget peak at the ceiling, because holidays are unlisted

Service status and fallback

DeepSeek publishes per-service uptime on a live status page, and no SLA alongside it. The terms go further than silence: they disclaim liability for inaccessibility, disruption, delay and suspension, and reserve the right to suspend service. One dated reading shows why the fallback has to sit outside DeepSeek. On October 15, 2026 the page carried an open incident that listed the V4 Pro API and the V4.1 Flash API together. Switching from one DeepSeek model to the other wouldn’t have helped. So the fallback provider isn’t a nice-to-have.
· Status page, read October 15, 2026: incident titled DeepSeek Web/API Partially Unavailable, opened 00:11 UTC, marked Identified at 01:06 UTC and unresolved when read. Affected: V4 Pro API, V4.1 Flash API and Chatservice.

  • Subscribe to the feed rather than checking manually
  • Uptime disclosure isn’t an SLA
  • One incident can hit Flash and Pro together
  • Design an approved fallback provider outside DeepSeek
5 / Agentic Work

Thinking inside tool use is useful, and needs hard boundaries

Cheap tokens make a bad loop cheap to run, right? That’s the whole risk here. DeepSeek’s docs list agent integrations from Claude Code and Codex to OpenCode, and put DeepSeek Harness first, which the docs homepage says is in developer preview for agent harness developers worldwide. Its September 10 article says OpenCode and WorkBuddy fully support V4.1-Flash. The operational risk hasn’t changed, and it’s slightly worse now that tokens are cheaper. An efficient model can take a lot of inexpensive wrong steps before anyone notices.

The bounded DeepSeek agent loop

Make the stop conditions part of the task.

Brief
Define outcome and maximum authority

State the finish line, the available tools, the prohibited actions, the step limit and the validation method.

  • Maximum turns or budget
  • No implicit external side effects
  • Named fallback
Tool-using analysis
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.
Agentic coding
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.
Structured verification
Return the result as the required JSON schema, then run an independent validation pass that checks completeness, allowed values and source coverage.
Commit rule
After two verification passes, commit to the strongest supported answer. Do not continue rechecking unless new evidence appears.
6 / Behavior Playbook

Nine habits for extracting the economic advantage safely

DeepSeek rewards engineering discipline. Explicit routing. Structured outputs. Hard agent limits. Scheduling against the weekday peak windows. And evaluation on your organization’s real tasks, not on a price list that changed twice in one month.

01

Pin a current model ID

Send deepseek-flash or deepseek-v4-pro. deepseek-v4-flash and deepseek-v4-flash-vision-exp were retired on September 10, 2026 and are only temporarily routed to V4.1-Flash. deepseek-chat and deepseek-reasoner are no longer listed on the pricing page or the docs model list. DeepSeek’s April 24 notice set the shutdown for 2026/07/24 15:59 UTC.

02

Route by effort, not by model

Flash covers the range. Escalate low, high, max before you consider a different model. Write the effort field for the format you’re actually on: reasoning_effort, output_config.effort or reasoning.effort. And use DeepSeek’s own names. medium and xhigh both run as high.

03

Turn thinking off deliberately

Thinking is on by default at high effort, so every job pays for reasoning until someone switches it off. Don’t pay reasoning latency for transformations a schema can validate. In the OpenAI and Anthropic formats, send a thinking object set to disabled. On the Responses API, send reasoning.effort set to none, then check the token usage, because that surface drops what it doesn’t support without an error.

04

Exploit stable prefixes

A cache hit costs 2 percent of the cache-miss rate on Flash. Design repeated context to benefit from caching without hiding changing instructions.

05

Set hard agent limits

Maximum steps, budget, tools and stop conditions belong in the task contract.

06

Evaluate cultural and political domains

Test behavior on topics that matter to your users and markets, not only on code benchmarks.

07

Separate hosted and self-managed risk

The same model can create different legal and security profiles by deployment. Hosted means PRC processing, files kept until you set an expiry, and inputs that can be used to improve the models unless someone turns that setting off. MIT weights mean self-hosting is a real option for you.

08

Measure correction cost

Include retries and human review when you compare providers.

09

Schedule deferrable work off-peak

Off-peak runs at half the peak rate. Move batch and evaluation runs out of the two weekday UTC windows, and treat weekends and Chinese public holidays as off-peak in full. Budget peak at the ceiling, because the holiday list isn’t published.

7 / Watch Outs

The risk profile belongs in the architecture decision

Capability answers none of the questions that actually stall a deal. Jurisdiction. Hosted-service data handling. Political behavior. Licensing. Operational responsibility. DeepSeek states the jurisdiction facts plainly in its own policies, so you can decide on them rather than guess at them.

Data residency and governing law

Stated plainly in DeepSeek’s own policies. So decide whether it’s acceptable before the technical evaluation, not after.
· Privacy Policy, last updated February 10, 2026: DeepSeek directly collects, processes and stores your personal data in the People’s Republic of China.
· Open Platform Terms of Service, effective April 29, 2026: the agreement is governed by the laws of the People’s Republic of China in the mainland. Suits are filed at the location of the registered office of Hangzhou DeepSeek Artificial Intelligence Co., Ltd.

No promise the service reaches you

Before DeepSeek goes on a critical path, read the availability clause. The terms make no warranty that the service is available, or stays available, in any given jurisdiction. They disclaim liability for inaccessibility, disruption, delay and suspension from causes outside reasonable control. And they reserve broad rights to suspend. Put that beside the missing SLA and you have a provider that has written down in advance that it owes you nothing on uptime or on access. That isn’t a reason to walk away. It’s a reason to make the fallback real and tested rather than named on a slide.
· Open Platform Terms of Service, effective April 29, 2026, sections 1.4, 7.2 and 7.5.

Retention is open-ended

The Privacy Policy commits to keeping personal data for as long as necessary to provide the services. No fixed deletion window is published. It also gives users the right to opt out of the use of their personal data for training DeepSeek’s models or optimizing its technologies. That’s a control you have to exercise, not one that applies by default. Uploaded files are the sharper case. The Files API keeps a file permanently unless the upload sets an expiry, and the longest expiry on offer is 30 days.
· The Open Platform Terms of Service link out to exactly three documents: the general Terms of Use, the general Privacy Policy and a page headed Model Mechanism and Training Methods of DeepSeek. None of them is an API-specific privacy policy. So the general policy is the applicable one by DeepSeek’s own linking, not by inference.
· Files API guide, checked October 15, 2026: omit both expires_after fields to keep the file permanently.

Your inputs can train the model

The clause isn’t in the API terms. It’s in the general Terms of Use the API terms link to, and that document names APIs explicitly. DeepSeek reserves the right to use inputs and outputs, to a minimal extent, to provide, maintain, operate, develop or improve the services and the technologies behind them, subject to encryption, de-identification and irreversibility. There’s an opt-out and it’s named: a setting called Improve the model for everyone. Which reads like a consumer app switch. The Model Mechanism page points to the Privacy Policy for opting out. Have your legal review ask DeepSeek in writing where a developer account turns it off.
· Terms of Use, last updated March 27, 2026.

Your end users are not covered

Build a product on this API and DeepSeek’s privacy policy stops at you. It says the processing rules for personal data collected from end users of downstream systems built by developers on the open platform fall outside that policy. So you write your own disclosures. You answer your own data-subject requests. And you can’t hand a customer a link to DeepSeek’s page as the answer. Teams usually find this during a security review, which is the expensive place to find it.
· Privacy Policy, last updated February 10, 2026.

No separate enterprise terms

The same Open Platform Terms of Service cover individual and enterprise developers. There’s no enterprise agreement, no separate data-processing addendum on any first-party page, and no published SLA behind the status page uptimes. So if your procurement process assumes a negotiated contract exists to fall back on, check that assumption early.

Political sensitivity

Evaluate behavior on politically sensitive and culturally specific topics that matter in your markets. This is an evaluation task on your own prompts. No benchmark answers it for you.

Provider concentration

Don’t build a critical system without model abstraction, fallback and exit planning. A provider that retired two model IDs on the day it announced their replacement is a provider whose lineup you should be able to leave quickly. And an outage can take both of its API models down at once.

Open-weight operations

Self-hosting transfers patching, serving, scaling, safety and monitoring to the organization. The MIT license makes that legally simple. Legally simple isn’t the same as operationally cheap.

Long-context confidence

A 1M window doesn’t guarantee full coverage, reliable retrieval or balanced attention. DeepSeek’s efficiency claims for the new KV cache design are vendor claims about resource use, not about retrieval quality.

Agent loops

Cheap tokens can make a runaway verification or tool loop inexpensive and still operationally harmful. And tokens just got cheaper.

Pro outlived its own shutdown date

A dated article gave Pro a September 14 ending. As of October 15, 2026, Pro is still fully served. That sounds like good news and it’s really a planning problem, because nothing dated says the reprieve is permanent. The article conditions the routing on V4.1-Pro launching, and V4.1-Pro hasn’t launched. The only sentence saying service continues with billing unchanged lives on the change log, in English and Chinese, pages rewritten in place with no version history. The pricing page doesn’t repeat it. DeepSeek promises further notice of any change, and names no channel and no lead time. So anyone with a production dependency on Pro should get written confirmation from DeepSeek rather than rely on any page at all. Diarize the date, re-read the source page on the same cadence, and keep a dated copy of the row you depend on. A date written down once and trusted is how you miss a deadline that moved.
· September 10 release article: all deepseek-v4-pro requests route to V4.1-Flash from 04:00 UTC on September 14, 2026, and this continues until V4.1-Pro launches.
· Change log, checked October 15, 2026: Pro continues after that date with unchanged billing, and DeepSeek will provide further notice should there be any changes.

The price moved twice in a month

This is the one to say out loud to a client, right? The peak and off-peak policy went live at 16:00 UTC on August 16, 2026. Twenty-four days and twelve hours later, at 04:00 UTC on September 10, the whole list was replaced again. DeepSeek published the August rates only as a chart image, and superseded rates leave the pricing page, so how far the September list moved from August is unconfirmed. Two full repricings in under a month. That’s the pattern to plan for. A vendor that can do that twice can do it again, in either direction. So budget with headroom, keep an abstraction layer, and re-read the pricing page before any commitment that rests on a published number.

Weights can move under a fixed model ID

Credit where it’s due: DeepSeek names builds where buyers can see them. The pricing page identifies DeepSeek-V4.1-Flash and DeepSeek-V4-Pro-0813, matching the weights channel. The caution survives the improvement, though. A model ID is still an alias over weights that can be replaced. Naming a build on a live page isn’t a versioned endpoint you can pin to. And deepseek-flash is a generational name rather than a dated one, so the next replacement may not change the string you send at all. This isn’t hypothetical. The V4.1-Flash repository has taken commits since release: a chat template, an encoding fix that preserves tool namespaces in prompts and completions, README edits and a technical report. None of them visibly replaces a weight shard. But a chat template and an encoding fix change what the model actually sees, and that can move your scores as surely as new weights. So re-run your evaluation suite on a schedule, not only on a version bump.
· V4.1-Flash commit log on Hugging Face, checked October 15, 2026. It shows relative dates only, and they disagree with the repository listing, so date nothing from them.

Retired in the API, still on the weights channel

The DeepSeek-V4-Flash-Vision-Exp repository is still published on DeepSeek’s weights channel on October 15, 2026, 35 days after the API retired the model on September 10. Read that two ways. It’s a real escape hatch, because an MIT-licensed model you can host doesn’t disappear when the vendor withdraws the endpoint. It’s also evidence that DeepSeek’s channels don’t update together. So a live weights page isn’t proof that a model is still served.

Benchmark scores with no citable text

The September 10 release article publishes its benchmark comparison and its price chart as images, so its figures can’t be quoted from it. Any text version sits on the change log, a page DeepSeek rewrites in place. You can read a page like that and you can’t cite it, and that distinction turns real the moment a procurement document needs a source line. Same trap on the August 21 article, whose benchmark figures are chart-only too. So screenshot anything you intend to rely on, treat the claim that V4.1-Flash outperforms V4-Pro as a vendor claim, and measure it on your own tasks.

Risk decisionHosted APISelf-managed weights
DataContent processed and stored in the PRC. Uploaded files kept until an expiry is setAccess to model inputs, logs and infrastructure
SecurityProvider and network controlsProtection for the serving stack and the weights
SafetyProvider policies and isolationFilters, evaluation and abuse controls, and who runs them
ReliabilityListed concurrency and a status page. No published SLA, and terms that disclaim availability by jurisdictionCapacity scaling and recovery
LegalPRC governing law, Hangzhou jurisdiction, one set of terms for everyone, and distillation from outputs expressly allowedMIT license, plus internal acceptable-use and export review
PrivacyInputs can be used to improve the models unless the setting is turned off. End users of a downstream product sit outside DeepSeek’s policyThe organization holds the inputs and still owes its users a policy
LifecycleTwo models retired on the day of announcementModel testing and rollout owned in-house
8 / Sources

Where these facts come from

Every source listed below is a dated first-party DeepSeek article or policy document. Several load-bearing facts live only on pages DeepSeek rewrites in place. Those can’t anchor a dated claim, so they’re attributed inline instead. Confirm each against the live page before you rely on it, and keep a timestamped copy of anything you’ll need to cite later. One trap worth knowing on DeepSeek’s documentation host: it serves its documentation homepage for unknown paths rather than a 404. So a page that renders isn’t proof the path exists.

DeepSeek-V4.1-Flash: Smarter, Faster, More EfficientDeepSeek · September 10, 2026 · V4.1-Flash launch, two retirements, the 04:00 UTC price change and the still-live routing of V4-Pro until V4.1-Pro launchesDeepSeek V4.1 Flash:更强、更快、更普惠DeepSeek · September 10, 2026 · Chinese release article: an orderly V4 Pro retirement, with routing and pricing times in Beijing time matching the English articleDeepSeek-V4-Pro GA ReleaseDeepSeek · August 13, 2026 · V4-Pro general availability and the 16:00 UTC August 16 start of peak and off-peak pricing, with rates shown only as a chart imageDeepSeek-V4-Flash-Vision-Exp Release: Multimodal API Now LiveDeepSeek · August 21, 2026 · history only · announces a model retired September 10, 2026, with chart-only benchmark figuresDeepSeek V4 Preview ReleaseDeepSeek · April 24, 2026 · history only · preview parameter counts and the alias retirement it still describes in the future tenseDeepSeek Open Platform Terms of ServiceDeepSeek · effective April 29, 2026 · PRC governing law, one set of terms for all developers, no availability warranty, suspension rights and permitted distillationDeepSeek Privacy PolicyDeepSeek · last updated February 10, 2026 · PRC storage, open-ended retention, the training opt-out and the carve-out for end users of developer-built systemsDeepSeek Terms of UseDeepSeek · last updated March 27, 2026 · names APIs, reserves minimal use of inputs and outputs to improve the services, and names the Improve the model for everyone setting

AI Mindset

Explore the model cheatsheets