DEV Community: daniel

DeepSeek Context Caching: The 50x Hit/Miss Spread, Explained Properly

daniel — Fri, 30 Oct 2026 14:28:29 +0000

Disclosure: I run ai-info.fastget.link, an independent reference site for AI API pricing and comparisons. No affiliate links in this post. All figures below were checked on the date noted — treat it as a snapshot, not gospel. This post is a mirror; the canonical version lives on my site.

DeepSeek Context Caching Explained

Written 19 Sep 2026 · prices verified 19 Sep 2026 19 Sep 2026

Short answer

DeepSeek's context caching is on by default for every account — no flag, no opt-in, no code change. You are already being billed at cache-hit prices whenever you hit.
A cache hit on deepseek-flash costs $0.003 per 1M input tokens (off-peak) vs $0.15 missed — a 50× difference. On deepseek-v4-pro it's $0.022 vs $0.66 — 30×.
But hits require an exact match against a complete persisted cache unit. If your prompt changes at the front, nothing hits, and you won't know unless you read prompt_cache_hit_tokens from the response.

Most people's mental model of prompt caching is "if I send something similar again, it's cheaper." DeepSeek's actual implementation is stricter than that, and the gap between the mental model and the real rule is where the money leaks. Here's how the disk cache actually decides what counts as a hit — and how to tell whether you're getting the 50× rate or silently paying full price.

What "cache" means here: disk, not memory

DeepSeek calls the feature Context Caching on Disk. Every request you send triggers the construction of a hard-disk cache of the input prefix. When a later request shares that prefix, the overlapping part is fetched from disk instead of being recomputed, and those tokens are billed at the cache-hit price.

Two consequences follow from the disk-based design. First, it's account-scoped and automatic — the docs are explicit that it "is enabled by default for all users, allowing them to benefit without needing to modify their code." Second, it's prefix-only: the cache matches the front of your input. Change one token near the top of a long prompt and everything after it is recomputed and billed as a miss, even if the rest is byte-identical.

The hit rule: complete units, not fuzzy prefixes

Because DeepSeek-V4 models use Sliding Window Attention, the cache isn't one continuous blob you can join mid-way. Each cached prefix is an independent, complete unit, and a request only counts as a hit when it fully matches one of those units. The docs' own example is worth internalizing:

Say your first request is A + B (a long document A, then question B). Your second request is A + C — same document, different question. The second request does not hit, because A + C doesn't fully match the stored unit A + B. A partial overlap isn't enough.

But here's the part most write-ups miss: after seeing A + B and A + C, the system detects the common prefix A and persists it as its own cache unit. When your third request A + D arrives, it fully matches A — and hits. So in a same-document workflow, you typically pay full price on the first one or two calls, then the hit rate snaps up on everything after.

The three persistence points

Cache units get written to disk at three kinds of moments, per the official guide:

Persistence point	What gets cached	When it helps
Request boundaries	Each request produces two units: one ending at the end of the user input, one ending at the end of the model output	Multi-turn chat — the next round that appends to either boundary hits
Common-prefix detection	When the system spots a shared prefix across multiple requests, it persists that shared part as its own unit	Stateless jobs re-sending the same document or system prompt
Fixed token intervals	For very long inputs or outputs, units are carved out at fixed token intervals	Long generations that never "reach an end position" would otherwise never cache at all

The multi-turn case is the easy win: request two of a conversation contains request one's input plus the assistant's reply, so it fully matches the unit ending at the model output — conversation history rides the hit price. The stateless case (fresh request per call, same big prefix) takes one or two calls to warm up, as above.

How to check what actually hit

The response usage object carries two fields that settle the question:

prompt_cache_hit_tokens — input tokens served from cache, billed at the hit price
prompt_cache_miss_tokens — input tokens recomputed, billed at the miss price

If you're doing anything nontrivial with the API, watch these numbers on your first few real calls. The miss pattern tells you exactly what's breaking the prefix: hit = 0 on every call means something at the very front of your prompt is changing each time — a timestamp in the system prompt is the classic culprit. A hit rate that collapses partway through a long conversation usually means the change sits mid-history.

What a hit is actually worth

Current official pricing, per 1M tokens, with the hit and miss rates side by side:

Model	Input (cache hit)	Input (cache miss)	Output
deepseek-flash, off-peak	$0.003	$0.15	$0.6
deepseek-flash, peak	$0.006	$0.30	$1.2
deepseek-v4-pro, off-peak	$0.022	$0.66	$1.98
deepseek-v4-pro, peak	$0.044	$1.32	$3.96

Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday — everything else is off-peak, and off-peak is a flat 50% of peak across the board. The caching discount stacks on top of that: hit tokens are 1/50th of the miss price on flash ($0.003 vs $0.15) and 1/30th on pro ($0.022 vs $0.66).

Run the numbers on a realistic job: a 500K-token document, 50 independent questions against it, no conversation state — each call re-sends the document. The first two calls miss (and the second one is what triggers the common-prefix persist). Calls 3 through 50 hit. Ignoring the small per-question tokens:

Without caching behavior (all 25M input tokens at miss price, off-peak): 25 × $0.15 = $3.75
With the cache warm (~1M miss + 24M hit): $0.15 + 24 × $0.003 ≈ $0.22

That's roughly 94% off the input bill, from a feature you never configured. Output tokens are billed the same either way. If you want to model this for your own workload, the cost calculator does the arithmetic.

What silently kills your hits

The failure modes all reduce to one idea: the front of your prompt is not as stable as you think it is.

Dynamic content at the front. A timestamp, request ID, or "today's date" injected into the system prompt makes every request a brand-new prefix. Keep static content first; put anything variable as late in the message list as possible.
Cache expiry. The cache is cleared once it stops being used — the docs say "usually within a few hours to a few days." A job that runs every night usually stays warm; one that runs weekly after a quiet weekend may pay the miss price on the first call again. That's expected, not a bug.
Best-effort, not guaranteed. The docs state plainly that the system doesn't promise a 100% hit rate. Build cost models on observed prompt_cache_hit_tokens, not on theoretical hits.
Caching doesn't make outputs reproducible. The disk cache only covers the input prefix; generation still runs through inference with your temperature and other sampling settings. Same cache hit, different wording — that's normal.
Third-party endpoints may not play along. The 50× spread is the official API's structure. A hosting provider reselling DeepSeek models doesn't automatically mirror it — if the cache math matters to you, check their hit pricing before assuming, and compare the effective rates against the cheapest DeepSeek API providers.

Where caching pays off most: coding agents and other harness-driven workloads. A fixed system prompt plus a growing but append-only conversation is exactly the shape the prefix rule rewards, which is one reason agent backends are a good fit for the API — see DeepSeek as a coding agent backend.

Stacking it with off-peak pricing

The two discounts are independent and compound multiplicatively. Off-peak hours alone halve every rate; caching then divides the input price by another 50. A flash-model input token that lands in an off-peak cache hit costs $0.003 per 1M — for practical purposes the floor of what DeepSeek input costs. Conversely, a peak-hour miss on v4-pro pays $1.32 per 1M, 440× the floor rate for the same token. If your workload is schedulable, shifting it into the off-peak window is the cheapest optimization available; the window rules are covered in DeepSeek off-peak pricing explained.

One more number worth knowing as context: V4.1-Flash and V4 Pro both take up to a 1M-token context. The larger the context you actually use, the more the hit/miss spread dominates your bill — at 1M tokens of prefix, the difference between a hit and a miss is $0.003 versus $0.15 per call before you've generated a single output token.

Prices and hit rules checked against DeepSeek's official documentation on 2026-09-19. DeepSeek revises pricing periodically — verify against the official pages before committing a budget.

Sources: DeepSeek API Docs — Context Caching, DeepSeek API Docs — Models & Pricing

Related on my site

Cheapest DeepSeek V4 API: 4 options compared — if you run through a reseller, check their hit pricing against the official spread.
DeepSeek V4 API cost calculator — model your own workload with hit/miss inputs.
Off-peak pricing, windows in your timezone — the other independent discount that stacks with caching.

Originally published on ai-info.fastget.link — more DeepSeek pricing and comparison pieces there.

DeepSeek V4.1 Flash vs V4 Pro: The Two Benchmarks It Lost

daniel — Fri, 30 Oct 2026 14:27:38 +0000

Disclosure: I run ai-info.fastget.link, an independent reference site for AI API pricing and comparisons. No affiliate links in this post. All figures below were checked on the date noted — treat it as a snapshot, not gospel. This post is a mirror; the canonical version lives on my site.

DeepSeek V4.1 Flash vs V4 Pro: We Checked the “Comprehensive Win” Claim

Written 10 Sep 2026 · prices verified 10 Sep 2026 10 Sep 2026

DeepSeek's September 10 release notes say V4.1 Flash has "comprehensively surpassed V4 Pro in performance, cost, speed, and total time." The company's own benchmark table mostly agrees: of the 16 benchmarks where both models have published scores, V4.1 Flash wins 14. But not all 16. On GPQA Diamond, the retiring flagship still leads 92.4 to 90.9. On HLE (text-only subset), it leads 42.7 to 39.1. Both holdouts are pure knowledge tests — every agentic, coding, terminal, and security benchmark goes to the new model, which also costs 70–86% less per token.

The short version

V4.1 Flash beats V4-Pro-0813 on 14 of 16 shared benchmarks — and undercuts it on every pricing line (cheapest: $0.003/1M cached input, off-peak)
The two exceptions are knowledge tests: GPQA Diamond (90.9 vs 92.4) and HLE text-only (39.1 vs 42.7)
Switch the model name to deepseek-flash; the old Flash names are retired but temporarily routed
From 04:00 UTC, September 14, all deepseek-v4-pro requests route to V4.1 Flash at Flash prices

1. What DeepSeek claimed, exactly

The launch post makes two claims worth separating:

"Benchmark results ahead of flagship models, including DeepSeek-V4-Pro" — from new pre-training methods plus larger-scale RL post-training.
"V4.1 Flash has comprehensively surpassed V4 Pro in performance, cost, speed, and total time" — the justification for retiring V4 Pro entirely.

"Comprehensively" is doing some work in that second sentence. DeepSeek's own table (linked above) compares V4.1 Flash against V4-Pro-0813, the previous V4-Flash-0731, GLM 5.3, Kimi K3, GPT-5.6-Sol, and Claude Opus 5 across 19 benchmarks. Against V4 Pro specifically, 16 rows have scores for both models. Flash wins 14.

2. The two benchmarks V4 Pro still wins

Benchmark	V4.1-Flash	V4-Pro-0813
GPQA Diamond	90.9	92.4
HLE (text-only subset)	39.1	42.7

Source: DeepSeek announcement, Sept 10, 2026.

HLE scores marked with an asterisk in DeepSeek's table are text-only subsets. V4-Pro-0813 has no full multimodal HLE score because V4 Pro doesn't support vision input — V4.1 Flash does, scoring 36.8 on the full HLE. So the fair reading: on text-only HLE, the old flagship wins 42.7 to 39.1; on vision benchmarks, there's no comparison to make.

Why the holdouts are knowledge tests

Both losses fit a pattern. GPQA Diamond and HLE measure dense scientific and academic knowledge recall. V4.1 Flash is built differently on purpose: a new Causal Encoder–Decoder architecture with just 8B active parameters for input and 16B for output (552B total, MoE) — an asymmetric design optimized for agent workloads that read far more than they write. Knowledge-heavy benchmarks reward something the new architecture visibly trades away. Whether that gap closes in the upcoming V4.1 Pro is the open question.

3. The 14 it wins — and by how much

Benchmark	V4.1-Flash	V4-Pro-0813	Δ
Terminal-Bench 2.1	90.6	87.9	+2.7
Terminal-Bench 3.0	30.0	11.8	+18.2
Terminal-Bench 4.0	31.2	12.4	+18.8
DeepSWE v1.1	74.2	62.7	+11.5
CyberGym	88.1	83.3	+4.8
SEC-Bench Pro	62.8	56.4	+6.4
ExploitGym	15.3	5.4	+9.9
Automation-Bench	54.8	43.2	+11.6
Agents' Last Exam	31.8	25.7	+6.1
NL2Repo-Bench	65.4	61.5	+3.9
ProgramBench	20.3	15.5	+4.8
HLE (w/ tools)	63.9	60.0	+3.9
Codeforces (rating)	3471	3348	+123
MathArena Apex	65.6	65.3	+0.3

The Terminal-Bench jumps are the headline: on the two newest versions, Flash scores roughly 2.5× the retiring flagship. It also beats its own predecessor (V4-Flash-0731) on all 16 shared rows.

Against the Western frontier models in the same table, the picture is mixed rather than dominant: Flash tops Terminal-Bench 2.1 (90.6 vs 89.1 for Claude Opus 5 and 88.8 for GPT-5.6-Sol) and edges both on DeepSWE v1.1 (74.2 vs 74.0 and 73.0). But Claude Opus 5 leads on Terminal-Bench 3.0/4.0 (43.3/51.8), NL2Repo-Bench (75.3), and ProgramBench (37.0), and GPT-5.6-Sol leads on ExploitGym (33.7) and HLE (44.5). At 3–10× the price, they should.

4. Pricing (USD card, effective 10 Sep 2026): Flash is cheaper on every line

Official USD pricing (source):

Per 1M tokens	deepseek-flash (cheapest)	deepseek-v4-pro
Input, cache hit — off-peak	$0.003	$0.022
Input, cache hit — peak	$0.006	$0.044
Input, cache miss — off-peak	$0.15	$0.66
Input, cache miss — peak	$0.30	$1.32
Output — off-peak	$0.60	$1.98
Output — peak	$1.20	$3.96

That's 86% cheaper on cache hits, 77% on cache misses, and 70% on output. Peak hours are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday; off-peak rates are 50% of peak. Cache-heavy agent pipelines — the ones that re-read long context every turn — capture most of the saving, which is exactly the workload the architecture targets.

Everything else the card tells you

Both models run 1M context with 384K max output, but only Flash supports vision, and Flash's concurrency limit is 2,500 versus Pro's 500. The cost cuts trace back to architecture: V4.1 Flash's KV cache needs just 1/4 the HBM and 1/8 the SSD storage of the previous generation (announcement).

5. Migration: what to change, and by when

Switch to deepseek-flash now. The legacy names deepseek-v4-flash and deepseek-v4-flash-vision-exp are already retired and route to V4.1 Flash at Flash prices.

deepseek-v4-pro is effectively dead at 04:00 UTC on September 14. Until V4.1 Pro ships, requests to it silently route to V4.1 Flash and bill at Flash rates.

After September 14, your "V4 Pro" responses will come from a different model. If any production flow relies on the old flagship's stronger knowledge recall (the GPQA/HLE gap above), re-run your evals this week — don't discover the regression in user reports.

Official partners WorkBuddy (including CodeBuddy) and OpenCode have fully integrated V4.1 Flash. Open weights are on Hugging Face, with the technical report (PDF) alongside.

6. FAQ

Is V4.1 Flash actually better than V4 Pro?

On 14 of the 16 benchmarks both models have published scores for, yes — often by a lot. It loses exactly two, both knowledge tests: GPQA Diamond (90.9 vs 92.4) and text-only HLE (39.1 vs 42.7). If your workload is graduate-level science recall rather than agentic coding, the retiring flagship still scores higher.

How much cheaper is it than V4 Pro?

On every line of the official USD card: 86% cheaper on cached input ($0.003 vs $0.022 per 1M off-peak), 77% on uncached input ($0.15 vs $0.66), 70% on output ($0.60 vs $1.98). Peak rates are double the off-peak rates on both models.

What happens to my app on September 14?

Nothing breaks. Requests to deepseek-v4-pro are routed to V4.1 Flash and billed at Flash rates until a V4.1 Pro ships. But your outputs will change, so re-run evals before the date — especially if you chose Pro for its knowledge recall.

Sources

DeepSeek API Docs — DeepSeek-V4.1-Flash release announcement — the two claims quoted in §1, the asymmetric architecture figures (8B/16B active, 552B total), the KV-cache reductions (1/4 HBM, 1/8 SSD), and the official benchmark table behind every number in §2–§3.
DeepSeek API Docs — Models & Pricing — the USD card in §4, peak-hour windows, concurrency limits, and the Sept 14 reroute-and-billing footnote.
DeepSeek on Hugging Face — DeepSeek-V4.1-Flash — open weights; the technical report (PDF) sits alongside.

Caveats, stated plainly. Benchmark figures are DeepSeek's own, read from its launch-day table; they have not been independently reproduced. Pricing is the official USD card as of 10 September 2026; prices can change, so re-check the console before committing a budget. This page will be updated if figures move — check the verified date at the top.

We have no affiliation with DeepSeek and carry no affiliate links on this page.

Related on my site

DeepSeek off-peak pricing explained — how to get V4.1 Flash at half price more of the week.
DeepSeek context caching explained — the $0.003/1M cached-input rate behind the cheapest line in this comparison.
DeepSeek V4 API cost calculator — compare the two models on your own volume.

Originally published on ai-info.fastget.link — more DeepSeek pricing and comparison pieces there.

Cheapest DeepSeek V4 API: 4 Options, Real Numbers ($9–$77/mo)

daniel — Fri, 30 Oct 2026 14:27:28 +0000

Disclosure: I run ai-info.fastget.link, an independent reference site for AI API pricing and comparisons. No affiliate links in this post. All figures below were checked on the date noted — treat it as a snapshot, not gospel. This post is a mirror; the canonical version lives on my site.

Cheapest DeepSeek V4 API — 4 Options Compared

Written 3 Sep 2026 · prices verified 3 Sep 2026 3 Sep 2026

Running DeepSeek V4 Flash on 100M input / 25M output tokens a month costs $9 on one platform and $77 on another — an 8x spread on the same model. Here are all four options with real numbers, and why most price quotes you'll find online are measuring the wrong thing.

The short answer

Cheapest overall: OpenRouter — $9.00/mo for Flash, $115.50 for Pro at the volumes above. It has no affiliate deal with us, which makes it a clean recommendation.
Cheapest fixed rate: Novita AI — $21.00/mo for Flash, at $0.14/$0.28 around the clock. That is 45% below DeepSeek's own off-peak rate and 73% below peak, with no scheduler to build and no route to re-check. It is still $12/mo more than OpenRouter — you are paying for a rate that does not move.
On Pro, do not use Novita — it costs roughly double DeepSeek's off-peak rate. Use the official endpoint off-peak, or route through OpenRouter.
The quotes you're comparing are probably for different builds. Both marketplaces list each model twice, undated and dated, and the two disagree on price and context window. See §4.

1. What each option actually costs

Enter your own volume. Rows are sorted by what you would actually pay, cheapest first.

Model

DeepSeek V4 Flash
DeepSeek V4 Pro

Monthly INPUT tokens (millions)

Monthly OUTPUT tokens (millions)

OpenRouter rows are the cheapest route for the given model ID, checked 3 Sep 2026. Routes are re-priced continuously — the cheapest host today is not guaranteed to be cheapest tomorrow (see §3).
DeepSeek Official splits into peak and off-peak: peak is UTC 01:00–04:00 and 06:00–10:00 Monday–Friday, 35 of 168 hours. The other 133 hours, including all weekend, are half price. Check those windows in your timezone →
Novita quotes a flat rate at every hour. DeepSeek's own endpoint appears as a host inside OpenRouter at exactly its off-peak rate, which is why the Pro rows tie.

2. Flash: where the flat rate wins

On Flash, both resale options beat DeepSeek's own endpoint at every hour of the day:

Option	Input /1M	Output /1M	Monthly	vs official peak
OpenRouter deepseek-v4-flash-0731 Sign up →	$0.05	$0.16	$9.00	−88%
Novita AI Sign up →	$0.14	$0.28	$21.00	−73%
DeepSeek Official — off-peak Sign up →	$0.22	$0.66	$38.50	−50%
DeepSeek Official — peak Sign up →	$0.44	$1.32	$77.00	—

OpenRouter is the cheapest of the four, and it is also the one we earn nothing from. But it is a routing marketplace: it sends your request to whichever host is cheapest at that moment, and that host changes. Novita charges a flat rate you can put in a budget — at the cost of paying more than the router.

Where to start
→ Want a flat rate you can forecast? Novita AI charges $0.14/$0.28 for Flash around the clock — 45% below DeepSeek's own off-peak rate and 73% below peak, with no scheduler to build.
→ Chasing the lowest possible number? OpenRouter's DeepSeek V4 Flash routes from $0.05/$0.16 — cheaper, but re-check it before you commit spend.

3. Pro: Novita is the wrong answer, and we'll say so

Pro inverts the picture. DeepSeek's off-peak rate on Pro is genuinely competitive, and Novita is not:

Option	Input /1M	Output /1M	Monthly	Note
OpenRouter deepseek-v4-pro-0813 Sign up →	$0.66	$1.98	$115.50	cheapest host is DeepSeek itself
DeepSeek Official — off-peak Sign up →	$0.66	$1.98	$115.50	133 h/week
Novita AI Sign up →	$1.60	$3.20	$240.00	2.1x the cheapest
DeepSeek Official — peak Sign up →	$1.32	$3.96	$231.00	35 h/week

Nothing here recommends Novita for Pro, so we won't pretend otherwise. If your workload is Pro-heavy, run the official endpoint and shift what you can into off-peak, or route through OpenRouter — where DeepSeek's own host currently lists at $0.66/$1.98, meaning you pay the off-peak rate at 15:00 on a Wednesday.

Novita is also worth a look on Pro if you need dated-build parity — but compare first: its deepseek-v4-pro-0813 entry is $1.32/$3.96, a pass-through of DeepSeek's peak rate. Details in §4.

4. Why the price quote you found is probably wrong

Here is the part that makes cross-platform comparison genuinely hard, and that almost no pricing page mentions: both OpenRouter and Novita list each DeepSeek V4 model twice — an undated entry and a dated build — and the two are priced differently on both platforms, in opposite directions.

On Novita, the dated entries are the expensive ones

Novita entry	Input /1M	Output /1M	What it actually is
`deepseek-v4-flash`	$0.14	$0.28	Novita's own hosting
`deepseek-v4-flash-0731`	$0.44	$1.32	Pass-through of DeepSeek's peak rate
`deepseek-v4-pro-0813`	$1.32	$3.96	Pass-through of DeepSeek's peak rate
`deepseek-v4-pro`	$1.60	$3.20	Novita's own hosting

Those dated figures are recognisable: $0.44/$1.32 is DeepSeek's Flash peak rate and $1.32/$3.96 is its Pro peak rate, to the cent. If you landed on Novita's pricing page, sorted by model name, and saw $0.44 for Flash, you would reasonably conclude Novita is expensive. On Flash it isn't — you were looking at a pass-through resale of DeepSeek's undiscounted peak price. The cheap entry on Novita is the undated one. On Pro the pass-through is $1.32/$3.96, which is cheaper than Novita's own $1.60/$3.20 — and both sit above DeepSeek's off-peak rate (§3).

On OpenRouter, it is the other way round

OpenRouter's dated Flash build is the cheaper of its two Flash entries ($0.05/$0.16 against $0.0679/$0.168) and sees more context — 1,310,720 tokens against 1,048,576. And the undated ID is not "the latest model": OpenRouter's own canonical_slug field shows deepseek-v4-pro is pinned to deepseek-v4-pro-20260423, the April build, three months older than -0813.

This pattern is not a DeepSeek quirk. Across the OpenRouter catalogue, dated entries run from 81% cheaper (qwen3-235b-a22b-2507) to twice the price with a quarter of the context (gpt-3.5-turbo-0613). A comparison quoting a model name without an ID cannot be checked, so treat it as unverified.

5. Flat rate vs routing marketplace

The honest trade-off, since it decides which of the two cheap options suits you:

OpenRouter is cheaper and less predictable. It is a routing marketplace: many hosts, one price each, and the router picks. We priced deepseek-v4-pro-0813 on 2 Sep 2026 and found its cheapest host at $0.5808/$1.742. Checking again on 3 Sep, that price was gone — the cheapest host is now DeepSeek's own at $0.66/$1.98, and the host that was cheapest had moved to $1.122/$3.366. Twenty-four hours.
Novita is a flat rate. $0.14/$0.28 for Flash, at every hour, on every host. You can forecast it, and you don't have to re-check it monthly.
DeepSeek direct is cheapest on Pro if your traffic can wait. 79% of the week is already off-peak, so bursty or batch workloads may need no scheduling at all — check your own case.
Cached input dwarfs all of this. DeepSeek charges $0.007–$0.044/M on cache hits. If your prompts repeat, that matters more than which platform you pick.

6. FAQ

What is the cheapest DeepSeek V4 API?
At 100M input / 25M output tokens a month: OpenRouter, at $9.00 for Flash and $115.50 for Pro. Both are cheapest-route prices checked 3 Sep 2026, and both will move.

Is Novita cheaper than DeepSeek's official API?
For Flash, yes, at every hour: Novita's flat $0.14/$0.28 is 45% below DeepSeek's off-peak $0.22/$0.66 and 73% below peak $0.44/$1.32. For Pro, no — Novita is $1.60/$3.20, roughly double DeepSeek's off-peak $0.66/$1.98. Use the official endpoint or a router for Pro.

Why do I see two different prices for the same DeepSeek model?
Because it is listed twice: an undated entry and a dated build. On Novita the dated entries are pass-throughs of DeepSeek's peak rate and cost more; on OpenRouter the dated Flash build is cheaper and sees more context. They are different builds, not duplicates — see §4.

Does the undated model ID mean "latest version"?
No. On OpenRouter it is pinned to a specific build — deepseek-v4-pro resolves to the 23 April 2026 release. The dated -0813 entry is newer. Check canonical_slug before assuming.

Do I need to schedule around off-peak hours to save money?
Probably not. 79% of the week is already off-peak, weekends included, applied per request with no opt-in. Scheduling only pays if a large share of your traffic lands in the 35 peak hours — and on Flash, alternatives undercut the off-peak rate anyway.

How often should I re-check these prices?
Route prices daily, if the number matters to a decision you're making now. Flat rates like Novita's only need checking when the provider announces a change. Every figure on this page carries a 3 Sep 2026 verified date and is answerable only to that date.

Bottom line
→ Flash workload, and you want a rate that won't move? Start with Novita AI — $0.14/$0.28 flat, 45% under DeepSeek's own off-peak rate.
→ Pro workload? Skip Novita and go direct to DeepSeek, or route through OpenRouter at $0.66/$1.98 all day.

Sources

Every figure on this page comes from one of these, fetched and snapshotted 3 Sep 2026:

DeepSeek Models & Pricing — official peak and off-peak rates, and the peak-window definition.
OpenRouter model list (JSON) — IDs, canonical_slug, context lengths.
OpenRouter per-host endpoints (JSON) — cheapest-route prices and host counts.
Novita AI Pricing — the flat rates quoted above.

Related on my site

DeepSeek V4 API cost calculator — plug in your own token volume.
DeepSeek context caching explained — cached input is $0.003–$0.044/M, which usually beats any provider switch.
Off-peak pricing, windows in your timezone — half price on the official API for 133 of 168 hours.

Originally published on ai-info.fastget.link — more DeepSeek pricing and comparison pieces there.

Two Months Running an AI-API Reference Site Solo: Google Punished the Launch, Bing Didn't Care

daniel — Fri, 09 Oct 2026 05:55:36 +0000

Two Months Running an AI-API Reference Site Solo: Google Punished the Launch, Bing Didn't Care

I run ai-info — an independent reference site for AI API decisions: pricing breakdowns, model-vs-model comparisons, and guides, aimed at developers who search things like "cheapest coding agent plan" or "is X worth it." No affiliate links (frozen on purpose), nothing for sale in this post. Everything below is a snapshot as of October 8–9, 2026; index states and rankings drift daily, so treat the dates as part of the data.

What shipped

Two months, 42 published pieces (independent slugs on the sitemap as of Oct 8), six sections covering DeepSeek, model comparisons, coding agents, local models, and news. The first batch of 19 posts went out in one burst — deliberately, to give crawlers something to chew on.

The Google arc, in three acts

Act one, the honeymoon. The site got indexed fast — exactly one page: the DeepSeek hub. Not the homepage, not any article. One navigation page. It even earned a brief stretch of impressions. Encouraging, in a narrow way.

Act two, the fade. Sometime around September 18, impressions went to zero and the hub dropped out of the index. No manual action, no message in Search Console, just silence.

Act three, the present. As of the October 8 GSC reading: 3 pages indexed — homepage, news hub, local hub. Zero article pages. Forty-two pieces of content, and Google trusts exactly the skeleton.

Whatever the right label is (scaled-content suspicion, new-domain distrust, or just the post-hangover every search doc warns about), the practical read is the same: this is a site-level verdict, not a per-article one.

What I decided not to do: panic-slow the output to "fix quality." With Bing and direct traffic independently viable and 40+ pieces sitting as a starting library, the play is to keep publishing on a daily floor and point the backlink push at the skeleton pages — homepage and hubs — rather than deep articles. Hubs are what Google trusted once; that's the best signal I have.

The Bing surprise

On September 29 I cross-posted a cost-math piece to dev.to with the canonical pointing back to my original. Bing indexed the copy the same day and counted the backlinks. Google, as expected, never indexed the copy at all — which is exactly how canonical folding is supposed to behave, so no complaint there. It just underlines that the two engines are playing very different games with the same site right now.

That post carried 7 dofollow links back to the site, re-crawled on October 8 and still dofollow. So the channel works as a link source. It does not work as a ranking shortcut — Google never indexed the copy, so none of its on-page text helped my pages. Expect reach ≈ zero, expect credit = the links themselves. Budget your effort accordingly.

The workflow that makes 42 posts survivable

Volume only works if per-post quality has teeth. Three rules doing the heavy lifting:

Narrow verification. Every post gets 3–5 hard numbers checked directly against official sources. Background figures cite the outlet they came from, by name. Verification stamps only go on numbers actually reconciled.
The Chinese-media translation gap. I monitor Chinese tech outlets and vendor announcements. A DeepSeek pricing change announced in Chinese typically gets real English coverage a day or two later. Writing for that 6–48 hour window is the closest thing to an unfair advantage a solo operator gets.
Dated caveats everywhere. Every post carries an "as of" statement. I'm only accountable for correctness at publish time — prices and plans drift, and pretending otherwise is how a reference site loses trust.

Honest scoreboard

Two months in: Google 0 (3 skeleton pages, no impressions), Bing 1 (same-day indexing, backlinks counted), direct and referral traffic small but real and growing. The bet underneath all of it: decision-intent queries around AI APIs are still under-served in English, and backlinks plus time still compound for genuinely useful pages.

Check back in six months. Or don't — I'll be posting the numbers either way.

Snapshot caveats: index counts from GSC on Oct 8–9, 2026; sitemap slug count from a live crawl on Oct 8, 2026; dev.to link attributes re-verified Oct 8, 2026. All of this changes — verify anything you'd cite.

Your DeepSeek V4 Bill Can Swing 2x by the Hour — the Actual Math Across Four APIs

daniel — Tue, 29 Sep 2026 03:45:05 +0000

Full disclosure up front: I run ai-info.fastget.link, where I track AI API pricing and publish the DeepSeek V4 cost calculator this post is based on. All rates below were checked against official pricing pages and live dashboards on 2 Sep 2026 — if you're reading this later, treat them as a snapshot, not gospel. No affiliate links anywhere.

The pricing page is lying to you (sort of)

DeepSeek doesn't have one price for V4. It has four, and which one you pay depends on decisions most developers never consciously make.

Take V4 Flash, the cheap workhorse model. DeepSeek's official list rate is $0.33 input / $0.99 output per million tokens — except that's not what you'll actually pay. DeepSeek runs peak/off-peak billing: off-peak hours cost ×0.67 of list ($0.22/$0.66), peak hours cost ×1.33 ($0.44/$1.32). So the same API call costs $0.22/M input at 3 AM and $0.44/M at 2 PM. Same model, same tokens, 2× apart. (The exact windows and how they map to your timezone are in my off-peak pricing breakdown.)

And that's just DeepSeek's own portal. The same model ID is resold elsewhere:

Provider	Flash (in/out per M)	Pro (in/out per M)
DeepSeek official (list)	$0.33 / $0.99	$0.99 / $2.97
DeepSeek official, off-peak (×0.67)	$0.22 / $0.66	$0.66 / $1.98
DeepSeek official, peak (×1.33)	$0.44 / $1.32	$1.32 / $3.96
OpenRouter (cheapest host)	$0.05 / $0.16	$0.58 / $1.74
Novita	flat rate, see dashboard	flat rate, see dashboard

Yes, you read the OpenRouter row right. The cheapest host on OpenRouter serves deepseek-v4-flash-0731 at $0.05/$0.16 — roughly a quarter of DeepSeek's off-peak rate. For Pro (deepseek-v4-pro-0813), the cheapest route is Alibaba at $0.58/$1.74, which actually undercuts DeepSeek's own off-peak rate.

So which number ends up on your invoice? That depends on three choices:

1. Official vs. reseller is a cache trade, not just a price trade

The per-token table says "reseller wins." But look at cache pricing before you switch: DeepSeek official sells cache hits at $0.007–$0.044/M — dramatically cheaper than anything resellers offer. If your app reuses long prompts (RAG with a fat system prompt, agents with big tool schemas, long-running chat threads), your real bill at DeepSeek official can come in far below the naive per-token estimate, sometimes below the reseller's headline rate.

Rule of thumb I've settled on: high cache-hit workload → official; stateless, prompt-diverse workload → the cheapest reseller route.

2. Peak windows are a scheduling problem, and scheduling is free

DeepSeek's off-peak discount is the cheapest optimization in the entire AI API market, because it costs zero code quality. Batch jobs, eval runs, embeddings backfills — none of them care whether they run at 3 AM. If half your Flash tokens are batch-shaped, moving them off-peak cuts that half's cost by half ($0.22 vs $0.44 per million input — a 2× spread).

The Pro model makes this more dramatic: official Pro ranges from $1.98/M output off-peak to $3.96/M at peak. Your nightly reasoning batch and your midday interactive traffic are not the same product and shouldn't share a price.

3. Watch the model ID, not the marketing name

One trap worth its own section: on OpenRouter, the undated deepseek-v4-pro ID routes from $0.87/$1.74 — above DeepSeek's own off-peak rate — while the dated deepseek-v4-pro-0813 starts at $0.58/$1.74 via Alibaba. Same model, different ID, different routing pool, different floor price. Pick by model ID and check which host actually serves the route; OpenRouter's "cheapest host" changes and per-route pricing isn't uniform.

Running your own numbers

The calculator on my site takes your monthly input/output token volumes and returns totals for both models across providers, with the peak/off-peak midpoint math built in. To give you scale: at a fairly typical solo-project load of 10M input / 5M output tokens per month on Flash, the spread between OpenRouter's cheapest host and DeepSeek's peak-hour official rate is the difference between roughly $1 and $11 a month. On Pro, the same spread is tens of dollars.

One caveat on the table: Krater.ai is subscription-based ($20/mo for 1,500 credits across 350+ models), so it's excluded from per-token comparisons. If your usage is spiky rather than steady, a subscription can beat all the per-token rates above — but that's a different math problem.

The short version

The "DeepSeek V4 price" is a range, not a number: ×0.67 off-peak to ×1.33 peak on official, and resellers undercut even the off-peak rate by 4× on Flash.
Cache-heavy workloads belong on DeepSeek official ($0.007–$0.044/M cache hits); stateless workloads should shop routes.
Schedule batch work into off-peak windows — it's the only API optimization that costs nothing.
On OpenRouter, quote the dated model ID and check the host.

Prices shift frequently on every one of these providers — I re-verify against official pages and dashboards regularly, and the calculator gets updated when they move. If a number here looks stale, the calculator page is the live copy.

Cross-posted from ai-info.fastget.link, my site on AI API pricing and decisions — the canonical version of this post lives there.