← All cardsDOSSIER · General AI assistant · SOLID · VERIFIED 2026-06-09
DeepSeek50SolidBenchmark pendingLow-cost reasoning and coding assistant / API model optionVerified 2026-06-09

Dossier · General AI assistant

DeepSeek

Low-cost reasoning and coding assistant / API model option · last verified 2026-06-09

Solid
General AI assistant
DeepSeek
50/100
ROLELow-cost reasoning and coding assistant / API model option
Editorial fit
50
Source quality
58
Citation honesty
54
Privacy controls
22
Value for money
45
Speed
68
FREEVerified 2026-06-07
FREEFree
CUSTOMPro
SolidVerified 2026-06-09VP·METHOD

VERDICT HISTORY

  • DeepSeek is the low-cost reasoning/API card. This pass sharpens the key facts: DeepSeek’s API is OpenAI- and Anthropic-compatible, current V4-Flash/V4-Pro pricing is extremely low, context length is listed at 1M, and legacy deepseek-chat / deepseek-reasoner names are being deprecated. The low price is real, but the editorial warning is bigger than usual: privacy, residency, institutional policy, and geopolitical availability need explicit review.
  • Converted imported Notion research into a full flagship-ready dossier template with metrics, panes, pricing deck, scenarios, benchmark rows, and comparison slices.

EDITOR'S NOTE

Desk hands-on 2026-06: The models can be fine on price; the consumer web app is what we would not recommend — trust, transparency, and platform polish lag the API story by a mile.

AT A GLANCE

Low-cost consumer chat and API model provider for reasoning, coding, and OpenAI-/Anthropic-compatible model routing, with unusually low token prices and important privacy, residency, and institutional-risk caveats.

Role: Low-cost reasoning and coding assistant / API model optionCategory: General AI assistantEditorial · hands-on

DeepSeek flagship-ready dossier: Low-cost reasoning and coding assistant / API model option.

PUBLIC FACTS · vendor & repo

List prices and pay-as-you-go entry points we can cite without running our own bench. Each tile links to a source when possible.

General AI assistantSetVerdictPal card
synthesizedEvidenceQuality gate
8Sources checkedCard sources
2026-06-07Pricing checkedQuality gate

Limits & product surface

Non-price vendor claims — multipliers, caps, and API scope. Detailed matrices live in subscription and SDK sections below.

Primary surface
consumer-appCard identity
Modalities
text, codeCard identity
Workflow roles
[object Object], [object Object], [object Object]VerdictPal editorial
Alternatives tracked
ChatGPT, Claude, Gemini, OpenRouter, QwenVerdictPal comparison set

HOW IT WORKS · agent loop

The public positioning for this product — the loop we score against on VerdictPal.

Start with the wedge

Cost-sensitive API experiments where token price matters more than ecosystem polish

Run the representative task

Run 20 reasoning/coding tasks on V4-Flash, V4-Pro, ChatGPT, Claude, Gemini, and OpenRouter routes.

Check the failure modes

Privacy/residency may be unacceptable institutionally

Compare before recommending

Compare against ChatGPT, Claude, Gemini before shipping advice.

METRIC LAB · 16 DIMENSIONS

Click a tile for the editorial note. Color follows score: coral, yellow, mint.

Shape

Avg 58 · 22–98

50Editorial fit

Weighted roll-up across 13 dimensions for General AI assistant; pending desk verification if rescored from public facts.

By group

Fit53
  • Editorial fit50
  • Wedge task fit48
  • Feature depth62
Cost51
  • Free-tier utility55
  • Cost-to-value45
  • Opportunity cost52
Trust64
  • Source grounding54
  • Privacy posture22
  • Failure transparency88
  • Evidence strength58
  • Transparency98
Workflow58
  • Integration reach55
  • Setup friction68
  • Reliability48
  • Competitive position52
  • Data portability65

SEARCH MODES · 4 lenses

One input box, many retrieval postures. Filter by tier.

01Free

Default

Default path for DeepSeek — verify limits on the live product.

02Pro

Deep work

Deep work path for DeepSeek — verify limits on the live product.

03Pro

Focused task

Focused task path for DeepSeek — verify limits on the live product.

04Free

Export & share

Export & share path for DeepSeek — verify limits on the live product.

BENCHMARK LEDGER

Public rows are vendor or third-party claims we logged with a date. Desk rows are reserved for VerdictPal self-run results.

MeasureResultSource
Run 20 reasoning/coding tasks on V4-Flash, V4-Pro, ChatGPT, Claude, Gemini, and OpenRouter routescorrectness, tool-call quality, code pass rate, cost per solved task, latency, retry rate.PlannedVerdictPal benchmark plan
Test a controlled 200K / 500K / 1M-token retrieval and summarization workloadfactual retention, quote accuracy, output completeness, latency, cost, failure rate.PlannedVerdictPal benchmark plan
Swap OpenAI and Anthropic SDK configurations to DeepSeek base URLsintegration friction, streaming, tool calls, JSON mode, error semantics, deprecation warnings.PlannedVerdictPal benchmark plan
Compare privacy, terms, residency, and policy constraints against ChatGPT, Claude, Gemini, and Mistraldata-use clarity, subprocessor clarity, enterprise controls, jurisdiction risk, allowed-use constraints.PlannedVerdictPal benchmark plan

PRICING DECK · checked 2026-06-09

Pro

Custom

month

  • Fallback tier inferred from the imported pricing summary; verify live checkout before publication.
  • Verify live regional checkout before publishing procurement advice.

RAW · checked 2026-06-09: Verified 2026-06-07 from official DeepSeek API docs: V4-Flash $0.0028/1M cache-hit input, $0.14/1M cache-miss input, $0.28/1M output; V4-Pro $0.003625/1M cache-hit input, $0.435/1M cache-miss input, $0.87/1M output. Both list 1M context, max 384K output, JSON output, tool calls, chat prefix beta, FIM beta in non-thinking mode, and OpenAI/Anthropic base URLs. Concurrency limits: 2500 Flash, 500 Pro. deepseek-chat and deepseek-reasoner deprecate 2026-07-24 15:59 UTC.

RESEARCH LOG · desk notes

What we learned while building this dossier — not vendor copy.

Notion Card pipeline

Notion research pass

DeepSeek is the low-cost reasoning/API card. This pass sharpens the key facts: DeepSeek’s API is OpenAI- and Anthropic-compatible, current V4-Flash/V4-Pro pricing is extremely low, context length is listed at 1M, and legacy deepseek-chat / deepseek-reasoner names are being deprecated. The low price is real, but the editorial warning is bigger than usual: privacy, residency, institutional policy, and geopolitical availability need explicit review.

VerdictPal git

Flagship-ready structure

Converted imported research into metrics, panes, scenarios, benchmark rows, comparison notes, and pricing deck.

COMPETITIVE LENS

Where DeepSeek wins for cited research — and where a rival still belongs in the stack.

DeepSeek is stronger when cost-sensitive api experiments where token price matters more than ecosystem polish; ChatGPT may still win for narrower fit, procurement, or specialist depth.

DeepSeek

  • Cost-sensitive API experiments where token price matters more than ecosystem polish
  • Coding and reasoning comparisons against ChatGPT, Claude, Gemini, Mistral, and OpenRouter routes

ChatGPT

  • Your institution bars sending research data to China-based AI services
  • You need clear enterprise data-residency, retention, or no-training guarantees for consumer chat

UNDER THE HOOD

Vendors named on the product about page — useful for procurement and privacy reviews.

low-cost reasoning
DeepSeek
coding assistance
ChatGPT
API model routing
Claude

DEEP PANES · 6 LENSES

Editorial lenses only. Subscription and API pricing live in their own sections above.

TEST SCENARIOS · hands-on lab

How we exercised the product. Step through each run before you trust the scores.

Test run

Step 1 of 4

Start from the persona in best-for item 1.

BEST FOR

  • Cost-sensitive API experiments where token price matters more than ecosystem polish
  • Coding and reasoning comparisons against ChatGPT, Claude, Gemini, Mistral, and OpenRouter routes
  • Developers who want OpenAI-/Anthropic-compatible endpoints for inexpensive model routing tests
  • Users who can keep sensitive or unpublished research out of consumer chat and standard API calls

AVOID IF

  • Your institution bars sending research data to China-based AI services
  • You need clear enterprise data-residency, retention, or no-training guarantees for consumer chat
  • You want the most mature assistant ecosystem, integrations, admin controls, or procurement story
  • You cannot monitor model-name deprecations and API compatibility changes

STRENGTHS

  • OpenAI- and Anthropic-compatible API
  • base URLs for both formats
  • V4-Flash and V4-Pro
  • thinking/non-thinking modes
  • 1M context
  • 384K max output
  • JSON output
  • tool calls
  • chat prefix completion beta
  • FIM completion beta in non-thinking mode
  • supported by agent/coding tools like Claude Code, GitHub Copilot, and OpenCode as backend model
  • low token pricing
  • high concurrency limits

WEAKNESSES

  • Privacy/residency may be unacceptable institutionally
  • Consumer and API terms differ
  • Geopolitical/export-control context can affect access
  • Very high max output can cause runaway generation if max_tokens is not set
  • Legacy model names deprecate 2026-07-24
  • DeepSeek reserves right to adjust prices; docs need monitoring

HOW IT COULD IMPROVE

  • Institutional policy friction could be reduced with a local/on-prem deployment option, or at minimum with clear data-handling documentation that compliance teams can review without an NDA.
  • Editorial fit scores 50/100, the weakest dimension on this dossier. Addressing this would meaningfully raise the editorial fit.
  • Wedge task fit scores 48/100, the weakest dimension on this dossier. Addressing this would meaningfully raise the editorial fit.
  • Cost-to-value scores 45/100, the weakest dimension on this dossier. Addressing this would meaningfully raise the editorial fit.
  • Opportunity cost scores 52/100, the weakest dimension on this dossier. Addressing this would meaningfully raise the editorial fit.
  • Source grounding scores 54/100, the weakest dimension on this dossier. Addressing this would meaningfully raise the editorial fit.
  • Privacy posture scores 22/100, the weakest dimension on this dossier. Addressing this would meaningfully raise the editorial fit.
  • Reliability scores 48/100, the weakest dimension on this dossier. Addressing this would meaningfully raise the editorial fit.
  • Competitive position scores 52/100, the weakest dimension on this dossier. Addressing this would meaningfully raise the editorial fit.

EDITORIAL EVIDENCE · 4 ENTRIES

TASK

Run 20 reasoning/coding tasks on V4-Flash, V4-Pro, ChatGPT, Claude, Gemini, and OpenRouter routes.

Pending.

correctness, tool-call quality, code pass rate, cost per solved task, latency, retry rate.

TASK

Test a controlled 200K / 500K / 1M-token retrieval and summarization workload.

Pending.

factual retention, quote accuracy, output completeness, latency, cost, failure rate.

TASK

Swap OpenAI and Anthropic SDK configurations to DeepSeek base URLs.

Pending.

integration friction, streaming, tool calls, JSON mode, error semantics, deprecation warnings.

TASK

Compare privacy, terms, residency, and policy constraints against ChatGPT, Claude, Gemini, and Mistral.

Pending.

data-use clarity, subprocessor clarity, enterprise controls, jurisdiction risk, allowed-use constraints.

This tool is in the VerdictPal Citation Fidelity protocol — frozen question bank, dual-reviewer grading, results still pending. Citation Fidelity v0.1 →

PRIVACY DEEP-DIVE · checked 2026-06-09

DeepSeek consumer chat and Open Platform/API are separate governance surfaces. Privacy policy notes downstream apps built on the open platform are controlled by their developers. Review DeepSeek privacy, open-platform terms, model/training disclosures, and local policy before sending unpublished or regulated data.

Training on your dataNot stated
EU data residencyNot documented
SOC 2 attestationNot documented
Local-first by defaultLocal option available

WORKFLOW ROLES

How this tool fits into a composed research stack:

low-cost reasoningcoding assistanceAPI model routing

QUALITY GATE · SOLID

Evidence ready
Benchmark pending
  • desk hands-on 2026-06-15: headline and metrics adjusted after desk trial.
  • score-revision 2026-06-09: differentiated from public facts; pending desk verification
  • Flagship-quality promotion 2026-06-04: source-backed dossier promoted; benchmark and public recommendation remain locked until evidence packet is complete.
  • Next action: Run cost/quality, long-context stress, compatibility smoke, explicit max-output, and institutional-risk benchmarks.
  • Has API: true
  • Public recommendation: false
  • evidence-override: Notion source-of-truth sync; evidence remains source-backed until desk benchmark or hands-on pass.
  • Flagship-ready structure generated from Notion deep research and imported git fields on 2026-06-09.

Related dossiers

Browse the atlas

The Pack · Editorial newsletter

New cards in your inbox. Free.

One short email when a card ships or changes status. No tracking, no third-party analytics. Unsubscribe in one click.