DeepSeek API Signals
What teams building AI features on DeepSeek API report, with every finding tied to its evidence.
Read the findingsWhat the evidence says
Practitioner sentiment, all practitioners
+57 Strong positive from 39 sources that take a side
How this figure is madeStrong positive overall signal, net 57 points
From 37 authors in 32 discussions. Run a pilot on your own prompts to measure output quality and cost per task.
- Analyzed
- 48 total sources
- Top praise
- Price per Task & Cache Hits (25 of 25 sources) See praise evidence
- Top concern
- Output Bugs, Latency & Errors (5 of 5 sources) See concern evidence
May 2026 to Oct 2026.
43 sources from practitioners, by month and sentiment.
Across 43 sources the read is positive. Of the 39 that state a view, 29 are positive, 4 mixed, and 6 negative. Price per task and cache hits draw praise in 25 sources, and output bugs, latency, and errors are raised in 5.
Positive, mixed and negative sources
The score counts them as 27 positive, 4 mixed and 6 negative votes, one per author. See How the numbers work.
See the evidenceBelow: 1 praise theme and 1 concern.
Findings
Counted for All practitioners, from 43 sources. Change group
What practitioners praise
Every capability practitioners describe positively, ranked by how many sources say so. Open one to read who said it.
Price per Task & Cache HitsAll 25 sources on it praise itLow list prices, half-price off-peak hours, and cheap cache hits keep long agent sessions cheap.
Counted for All practitioners.
Positive
SwellJoe, Hacker News commenter
Comment on: Kimi K3.1 Pro and K3.1 Pro HighspeedHacker News comment
Keeps a DeepSeek account and almost always uses it when a task needs API access, calling it good enough.
Positive
aftbit, software developer in Ohio, at Atomic Canyon in 2025
A vibe-coded MITM test proxy built for about $0.30 of DeepSeek API spendHacker News comment
The author tested with a vibe-coded MITM proxy of the author's own that hijacks PyPI installations, created for about $0.30 of spend on the DeepSeek API.
Positive
TheArcane, autonomous driving perception research engineer
A platform.deepseek.com account beside a $100 a month Claude subscription, and DS Flash used mostHacker News comment
The author has both a platform.deepseek.com account, cheering its off-peak pricing, and a $100 a month Claude subscription, and mostly uses DS Flash, pretty much every time.
Positive
computerex, Hacker News commenter
Comment on: DeepSeek-v4.1-Flash-1017Hacker News comment
Runs out of a US$20 plan's five-hour limit in minutes and gets far more done with US$20 of DeepSeek API spend.
Positive
63stack, Hacker News commenter
Comment on: The cost of getting a job todayHacker News comment
Says five dollars of DeepSeek tokens covers about a week of constant use.
Positive
WiSaGaN, Hacker News commenter
Comment on: Why isn't the industry freaking out about DeepSeek 4.1 Flash?Hacker News comment
Uses DeepSeek V4.1 Flash on DeepSeek's own API as a baseline model for everyday tasks, keeping frontier models for research and large coding projects.
Positive
try-working, Hacker News commenter
Comment on: Why isn't the industry freaking out about DeepSeek 4.1 Flash?Hacker News comment
Used more than 40 billion tokens for over US$800 on the DeepSeek API in 30 days, mostly V4.1 Flash, splitting complex work into phases with verification.
Positive
benjiro29, Hacker News commenter
Comment on: One month coding with GLM 5.3 FlashHacker News comment
Finds V4.1 Flash much cheaper on DeepSeek's own API because its cache hit rate reaches 99% and more, which widens the cost gap on long tasks.
Positive
taylorfinley, serial builder of things
1.8bn tokens in a week bought from DeepSeek, 99.2% of them cache hits, for US$18.13Hacker News comment
The author used 1.8bn tokens in the past week from DeepSeek itself, and 99.2% were cache hits. The total cost was US$18.13.
Positive
dools, runs BenkoPhone
Weeks without topping up a DeepSeek balance, against weekly OpenAI top-upsHacker News comment
The author goes weeks without topping up the DeepSeek balance and has to top up the OpenAI balance at least every week, and uses 4.1 Flash almost exclusively for both planning and implementation.
Mixed
tiluha, Hacker News commenter
Comment on: Ask HN: What default model do you use and why?Hacker News comment
Runs DeepSeek V4.1 Flash agents for speed and price with a frontier model as advisor, while calling the API's privacy policy not great.
Positive
Ekky Armandi, full-stack Python developer (DEV profile)
DeepSeek as the default model on ChatGPT CodexDeveloper blog post, DEV Community
Sets DeepSeek Flash as Codex's default model on the official API to drop a flat subscription, and prices a moderate coding week at US$0.18 off-peak against US$1.63 on GPT-5.
Positive
irthomasthomas, Hacker News commenter
Comment on: DeepSeek V4 Pro 0813Hacker News comment
Recommends replacing OpenRouter with direct calls to the DeepSeek API to skip the fee and get the better caching rate.
Positive
nchmy, Hacker News commenter
Comment on: DeepSeek V4 Pro 0813Hacker News comment
Says the official API's cache reads are about ten times cheaper than other hosts', so long chats cost about a dollar.
Positive
stavros, founder of Stochastic Technologies
Uses the DeepSeek API, pays peanuts, and is very satisfiedHacker News comment
The author uses the DeepSeek API, pays peanuts and is very satisfied.
Positive
Bayart, SRE
$5 a month going directly through DeepSeek's API all dayHacker News comment
Going directly through DeepSeek's API and hitting it all day long with light and medium tasks, the author is at $5 a month.
Positive
eulers_secret, works on embedded Linux for avionics, formerly at Intel for more than 10 years
The DeepSeek API for all coding at home, through OpenCode or Pi, for $5 over six monthsHacker News comment
The author uses OpenCode or the Pi harness with the DeepSeek API for all coding at home, and the cost was $5 for six months of use.
Positive
lionkor, software engineer in the electrical power industry
287 million tokens in a couple of days on DeepSeek's API, for US$5.77Hacker News comment
Using DeepSeek's API directly, the author burnt 287 million tokens in a couple of days, at a cost of US$5.77.
Positive
atreids, Software Engineer in Scotland
DeepSeek's platform API with V4 Flash in OpenCode, more than acceptable for maybe $10 over a couple of weeksHacker News comment
The author finds going directly through DeepSeek's platform API, with its V4 Flash model hooked into a harness such as OpenCode, more than acceptable, and thinks the spend was maybe $10 over a couple of weeks.
Positive
Kuyawa, programmer
15 cents this month on the DeepSeek API with a home-built coding agent, and three apps delivered to clientsHacker News comment
This month the author has spent only 15 cents using the DeepSeek API with a coding agent of the author's own, and delivered three apps to clients.
Mixed
CJefferson (Christopher Jefferson), Professor in the School of Computer Science at the University of St Andrews
DeepSeek's Anthropic endpoint for bigger reading jobs near the limitHacker News comment
The author advises putting $5 into deepseek.com and setting ANTHROPIC_BASE_URL to https://api.deepseek.com/anthropic. The author started with bigger reading jobs near the limit and describes the service as somewhat lower quality; its lower cost lets work continue.
Positive
SwellJoe, works on Virtualmin
The DeepSeek API caches raw API calls well, so it is hard to spend a lot fastHacker News comment
The DeepSeek API seems really good at caching, so it is hard to spend a lot of money fast. The author had been concerned about needing to do something specific in a simple agent harness to make caching effective, and DeepSeek seems to handle caching very effectively for raw API calls.
Positive
sourcecodeplz, full-stack engineer
DeepSeek API credits never expire, and the prices are ridiculously low for the qualityHacker News comment
With the DeepSeek API, the author's credits never expire and can be used whenever and in whatever amount, and the prices are ridiculously low for the quality, intelligence and performance.
Positive
vinhnx, iOS & Applied AI Engineer
US$1.15 used after two full weekends on a 2 USD DeepSeek API top-upHacker News comment
The author registered a DeepSeek API key, topped up 2 USD and used only DeepSeek V4 Pro and Flash with max thinking. Usage stood at US$1.15 after two full weekends.
Positive
xulingfeng, QA engineer of 15 years building an AI test framework
3.4 Billion Tokens for $27: A Real DeepSeek Cost AnalysisDeveloper blog post, DEV Community
From the author's actual DeepSeek bills for one month: 3,452,248,487 tokens and 37,882 API calls for $27, a monthly spend of ¥194.09, which the author puts down to a cache hit rate of about 95%.
GILD's fact checks, which don't count as sources
Checked Oct 28, 2026.deepseek-flash costs US$0.30 per 1M cache-miss input tokens, US$0.006 per 1M cache-hit input tokens, and US$1.20 per 1M output tokens at peak, and off-peak rates are half. Models and pricing
Open them in the record
What concerns them
Every concern practitioners raise, most raised first.
Output Bugs, Latency & ErrorsAll 5 sources on it raise it as a concernEmpty answers, leaked tool-call markup, agent loops that stop early, TLS errors, and slow periods.
Counted for All practitioners.
Negative
Joseph Cake, AI builder using Kilo Code
Intermittent EPROTO: The TLS handshake failed. Does the server speak TLS on this port?GitHub issue
Kilo Code agent sessions on the DeepSeek API intermittently stop with TLS handshake failures.
Negative
Doll-sensei, operator of a QQ bot on the DeepSeek API
From a production bot, reports DeepSeek Flash writing raw internal tool-call markup into message content when a request omits tools, and ships a defensive parser.
Negative
Amiyanshu-web, developer replaying production conversations
Replaying three captured production conversations 45 times, 13 ended with the whole answer in reasoning_content and empty content on a normal stop.
Negative
Lady-Lin, developer running a coding agent on the API
deepseek-v4-flash terminates prematurely in agent tool-call loops via OpenAI-compatible APIGitHub issue
In a tool-call loop, V4 Flash often answers with one short sentence and a stop finish, ending the agent's turn and skipping the tool call.
Negative
shortformblog, Hacker News commenter
Comment on: DeepSeek planning to significantly raise pricesHacker News comment
Noticed the DeepSeek API running slowly for several days and points to an alternative model with similar price and performance.
Where the evidence agrees
Statements this group's sources agree on, then the ones close to agreement.
For all practitioners, no statement reaches 5 sources from 3 authors.
Where it disagrees
Questions the sources split on, with the count on each side.
For all practitioners, no question reaches 5 sources from 3 authors, with 2 on each side.
The evidence
Whose experience this is
The three groups, then every practitioner by the context they work in, and how emphasis changes with scale.
Three groups drawn from the same sources. Choose a group to select whose experience counts. Within that group, each author counts once in the practitioner sentiment, with their latest view. Signals open on All practitioners. Teams and leaders have 0 accounts; GILD prints their sentiment from 20.
Teams and technical leaders
0 sources. No team account remains after the October 30, 2026 check.
Individual builders
43 sources. Developers describe the API in personal projects, coding harnesses, and comparisons with subscriptions.
All practitioners
43 sources. Together, these accounts are mostly individual developers comparing DeepSeek's price with frontier models and subscriptions.
43 sources from practitioners, by the context they work in. Filled rows count for this group.
- Individual builderDevelopers building on the platform on their own43 sources, counted here
Emphasizes Price per task, cache hits, and agent-loop bugs
- Small product team2 to 10 engineers, or an unnamed startup teamNo sources from this setting yet
- Product or platform team11 to 50 engineers, or a named company's teamNo sources from this setting yet
- Technical leadershipLeaders setting model and platform choices across teamsNo sources from this setting yet
- Organization-wideEnterprise or company-wide AI platformNo sources from this setting yet
What changes with scale. Individual developers choose DeepSeek's API for price per task and its cache-hit rate, and report output bugs in agent loops. No team account remains among the records, so team-scale reads rest on the Hub's operator cards.
Where the sources come from
Where this group's sources were published.
Counted for All practitioners, from 43 sources. Change group
48 sources for this group, 5 of them documentation
- Developer Communities41, 85% of 48 sources
- Product Documentation5, 10% of 48 sources
- Engineering Blogs & Reviews2, 4% of 48 sources
Earlier evidence remains when it still applies. Records tied to retired model names carry Needs a recheck.
See sources by originThe record
Who each source is, how it knows, whether it still applies and what it counts toward.
Newest first, on its own page. Every count above opens its sources there.
Read all 48 sourcesAbout this read
What changed, and how we count
What changed
- Oct 31, 2026: 20 records added from a search for practitioners whose role or employer a page states; a second model checked every record against its stored source and accepted 20 of 21. 43 remain.
- Oct 30, 2026: a second model checked every record against its stored source. 20 of the 43 records were removed: 8 not about DeepSeek's own API, 7 that fail GILD's source standard or state no outcome a buyer acts on, 4 whose summaries claimed more than their source says, and 1 no route could fetch. 7 more keep their place without the role or project their author's profile gives, since the source itself does not state it. 23 remain.
- Oct 28, 2026: first edition, built from the Hacker News, Reddit, GitHub, and blog accounts GILD's research missions stored on October 28, 2026, favoring sources that name their author's role, company, or project.
- Every account links to its public source and names its author and publication date.
- A finding shows only when at least 5 sources from 3 different authors discuss it. A disagreement also requires at least 2 sources on each side.
- Counts practitioners who describe their own use of api.deepseek.com; benchmark leaderboards and press coverage are excluded.
- Checks the price, limit, data, and model-name findings against DeepSeek's pricing page, rate limit page, privacy policy, and terms on the review date.
Method
A running review of what teams building AI features on DeepSeek API report, all from individual builders. Each finding links to its sources, and each conflict shows both sides.
Counts cover only the sources GILD reviewed.
Who counts
- Teams and technical leaders
- Small product team, Product or platform team, Technical leadership, Organization-wide
- Individual builders
- Individual builder
- All practitioners
- Everyone in the teams and leaders and individual builders groups
Choose a group to select whose experience counts. Within that group, each author counts once in the practitioner sentiment, with their latest view. GILD files each source in a group from how the account describes its author. A role or employer marked self-described is what the source states about themselves. Documentation checks facts for every group and never counts toward sentiment.
How the numbers work
- Overall signal. Positive votes minus negative ones, per 100 votes from the sources that take a side. Each author counts once, with their latest view, whoever they are and wherever they posted. The older weighted reading, where accounts GILD matched to a named person count three times when GILD lists the author among technical leaders and twice otherwise, is shown beside it as a second view. Where an author wrote more than one source, the line beside the split gives the votes, one per author, so the score follows from it. When one author's sources from the same day disagree, that author counts as mixed, and a half point rounds away from zero. Sources with no opinion are listed and not counted. Strong positive is +50 or higher, positive +20 to +49, mixed −19 to +19, negative −20 to −49, strong negative −50 or lower.
- Weighted reading, second view. Counted the older way, where an account GILD matched to a named person counts two or three times, all practitioners read +58 strong positive (method sentiment/S1). GILD printed that number as "GILD score" until October 29, 2026; most of those matches rest on a name or a self-described role.
- Unscored. A group gets a score only with 20 sources from 15 authors. Below that, one source moves the score about 5 points, so the read says "Unscored until 20 sources from 15 authors" and shows its counts.
- Independence. A source whose author works for the vendor, for one of its partners or programs, or for a competitor, and a page the vendor published or ran, is listed with that tie and counts toward nothing. One author counts on at most two records per product. The line under the score says how many authors and how many separate discussions the counted sources come from: comments in one thread are one discussion.
- Praise and concerns. How many of the sources that discuss a theme praise it or raise it.
- Consensus. A statement joins when 20% of the group's sources support it and fewer contradict it. Contradicting sources are counted separately. Below that line a statement shows as building. At 40% or more it reads Strong agreement; below, Recurring point.
- Disagreements. The count on each side. GILD names no winner and gives no share of all users.
- Minimum evidence. A finding needs 5 sources from 3 authors; a disagreement also needs 2 on each side. Smaller counts stay in the table below, marked under the minimum.
Labels on a source
A source's date and whether it still applies are separate: a recent post can describe an old version, and an old one can still hold.
- Documented
- Stated in official documentation.
- Reported
- A practitioner's account of their own use.
- Measured
- Backed by the source's own measurements, such as a comparative test or tracked team metrics.
- GILD analysis
- GILD's check of a claim against official documentation.
- Current
- Describes the product as it is now.
- Still applicable
- Older, and the behavior or lesson still holds.
- Needs a recheck
- May depend on a plan, version or setting that changed since it was written. It keeps counting until GILD rechecks it, since the rest of the account usually holds, and its card says so.
- Out of date
- No longer accurate. Kept in the record; counts toward nothing current.
Flagged for a recheck, of the sources that take a side: teams and leaders 0 of 0, individual builders 1 of 39, all practitioners 1 of 39.
Limits
- These figures describe reviewed sources; they do not represent all DeepSeek API users. Hacker News supplies 33 of the 43 practitioner sources, and its threads compare the API's price with coding subscriptions, which tilts the set toward individual developers and cost. GitHub issues supply 6, 4 of them bug reports, which tilt toward problems. Accounts of DeepSeek models served by other hosts, and government restrictions on the DeepSeek chatbot, are kept out of the count.
- Sources run from May to October 2026, from V4 Pro and V4 Flash through V4 Flash 0731, V4 Pro 0813, and V4.1 Flash 1017, which now answers the deepseek-flash name. One account about V4 Flash before V4.1 carries Needs a recheck.
Every count, and the sources behind it
9 findings for 3 groups
Each row opens its sources for the group you are reading.
| Finding | Teams and technical leaders | Individual builders | All practitioners | Sources |
|---|---|---|---|---|
| Current readCurrent sentiment | 0 positive, 0 mixed, 0 negative | 29 positive, 4 mixed, 6 negative | 29 positive, 4 mixed, 6 negative | Sources |
| PraisePrice per Task & Cache Hits | 0 of 0 positive (under the minimum) | 25 of 25 positive | 25 of 25 positive | Sources |
| PraiseSpeed | 0 of 0 positive (under the minimum) | 4 of 4 positive (under the minimum) | 4 of 4 positive (under the minimum) | Sources |
| PraiseModel Quality for the Price | 0 of 0 positive (under the minimum) | 2 of 3 positive (under the minimum) | 2 of 3 positive (under the minimum) | Sources |
| PraiseOpenAI and Anthropic Compatible | 0 of 0 positive (under the minimum) | 3 of 3 positive (under the minimum) | 3 of 3 positive (under the minimum) | Sources |
| ConcernOutput Bugs, Latency & Errors | 0 of 0 raise it (under the minimum) | 5 of 5 raise it | 5 of 5 raise it | Sources |
| ConcernCost That Grows With Tokens | 0 of 0 raise it (under the minimum) | 1 of 3 raise it (under the minimum) | 1 of 3 raise it (under the minimum) | Sources |
| ConcernData Location & Training | 0 of 0 raise it (under the minimum) | 2 of 2 raise it (under the minimum) | 2 of 2 raise it (under the minimum) | Sources |
| Current readEvidence mix | 5 sources | 48 sources | 48 sources | Sources |
Fact checks
GILD's checks of claims against official documentation, kept apart from the source counts.
- Price per Task & Cache Hits. deepseek-flash costs US$0.30 per 1M cache-miss input tokens, US$0.006 per 1M cache-hit input tokens, and US$1.20 per 1M output tokens at peak, and off-peak rates are half. Models and pricing, checked Oct 28, 2026
- OpenAI and Anthropic Compatible. The DeepSeek API uses formats compatible with OpenAI and Anthropic, with an Anthropic base URL at api.deepseek.com/anthropic and guides for Claude Code, Codex, and OpenCode. Your first API call, checked Oct 28, 2026
- OpenAI and Anthropic Compatible. DeepSeek's documentation lists integrations for coding agents: Claude Code, GitHub Copilot, and OpenCode can use DeepSeek as their backend model directly. Your first API call, checked Oct 28, 2026
- OpenAI and Anthropic Compatible. Because the API format is compatible with OpenAI and Anthropic, integrations built on the OpenAI or Anthropic SDKs reach DeepSeek by changing the base URL and API key. Your first API call, checked Oct 28, 2026
- Cost That Grows With Tokens. Model usage is billed per million tokens at the listed price and Search and Fetch per 1,000 successful requests, with fees deducted from a topped-up or granted balance. DeepSeek reserves the right to adjust prices. Models and pricing, checked Oct 28, 2026
- Data Location & Training. DeepSeek's privacy policy states that it directly collects, processes, and stores personal data in the People's Republic of China. Depending on where the user lives, the rights it lists may include opting out of the use of personal data to train its models. Privacy policy, checked Oct 28, 2026