The short answer
For low-cost finance research-agent loops in 2026, DeepSeek is the cheaper hosted API on verified numbers: DeepSeek V4-Flash is $0.14 input / $0.28 output per Mtok with a 1M-token window and default caching. Mistral's exact per-token rates are not cleanly published, so confirm them in your own console. The cost lever is tokens-per-step, not the vendor.
For low-cost finance research-agent loops in 2026, DeepSeek is the cheaper hosted API on verified numbers: DeepSeek V4-Flash is $0.14 / Mtok cache-miss input and $0.28 output, with a 1M-token context window and default automatic caching. Mistral's hosted API spans Large 3, Medium 3.5, Small 4, and the Ministral 3 series, but its exact per-token API rates are not cleanly published on the official pricing page (it renders client-side and shows consumer plans), so treat Mistral's per-token figures as the one input you must confirm in your own console. On a finance agent loop, the cost lever is tokens-per-step and steps-per-loop more than the vendor. Model the envelope with the Agent Cost Envelope Calculator.
TL;DR
| Vendor | Model | Input $/Mtok | Output $/Mtok | Context | Verification |
|---|---|---|---|---|---|
| DeepSeek | V4-Flash | $0.14 (miss) / $0.0028 (hit) | $0.28 | 1M | Official docs |
| DeepSeek | V4-Pro | $0.435 (hit-adj.) | $0.87 | 1M | Official docs (promo) |
| Mistral | Large 3 / Medium 3.5 / Small 4 / Ministral 3 | see note | see note | per model | Lineup verified; per-token rates not machine-readable on official page |
DeepSeek's numbers are verified end to end. Mistral's model lineup is verified; its per-token API rates are the one fact to confirm directly before you budget.
DeepSeek: verified, and cheap
DeepSeek's current hosted models are deepseek-v4-flash and deepseek-v4-pro, both with a 1M-token context window and 384k max output. The legacy aliases deepseek-chat and deepseek-reasoner now route to the non-thinking and thinking modes of V4-Flash respectively, and are scheduled for deprecation on 2026-07-24.1
Verified per-Mtok pricing:1
- V4-Flash cache-miss input $0.14, cache-hit input $0.0028, output $0.28.
- V4-Pro cache-miss input listed under a 75%-off promotion through 2026-05-31 (after which it adjusts to one quarter of the original price), output $0.87.
The cache-hit input rate is the standout: at $0.0028 / Mtok it is ~2% of the cache-miss rate, and automatic context caching is on by default. For a research agent that re-sends a large fixed instruction block every step, that boilerplate is nearly free after the first step. For where V4 actually holds up on financial tasks (and where the cheap price hides quality gaps), see DeepSeek V4 for finance 2026.
Mistral: verified lineup, unverified per-token rates
Mistral's current hosted lineup, verified from the official model documentation, is Mistral Large 3 (open-weight general multimodal), Magistral Medium 1.2 (premier reasoning), Mistral Medium 3.5 (agentic/coding), Mistral Small 4, and the Ministral 3 series (14B / 8B / 3B).2
What is not cleanly verifiable from Mistral's official pricing page (as of May 2026) is the exact per-Mtok API rate per model: the pricing page renders client-side and surfaces consumer Le Chat plans rather than a machine-readable API rate card. Third-party trackers quote figures (and disagree among themselves), which is precisely why this article does not state a hard Mistral per-token number as verified. Confirm the live rate in the Mistral console or API pricing section before you budget against it.3 The verified facts: the lineup above, and per-token billing with cache-read discounts on most models.
Why the loop shape beats the vendor
For a finance research agent, the dominant cost driver is not the per-token rate but the loop:
- Tokens per step how much context each tool-call + reasoning step carries.
- Steps per loop how many iterations before convergence.
- Markets per day how many independent loops you run.
A 2x reduction in steps-per-loop or a disciplined context budget per step moves the bill more than swapping a $0.14 model for a $0.28 one. The cheapest vendor on a bloated loop loses to a disciplined loop on a mid-priced model.
Verified engine output
The block below runs the Agent Cost Envelope Calculator on a 150-market-per-day research loop (6 steps/loop, a convergence check, business-calendar cadence) using the engine's cheapest in-table frontier-class model as the reference floor. DeepSeek V4-Flash sits below that floor on a per-token basis, so the engine number is a conservative upper bound on what the same loop costs on DeepSeek: the real DeepSeek bill is lower. The output is computed live from the shipped bundle, not typed by hand; the DeepSeek and Mistral list prices in the table above are the verified inputs to do the substitution yourself.
Decision guidance
- Cheapest verified hosted API for a finance agent loop DeepSeek V4-Flash, especially with the default cache amortizing a fixed instruction prefix.
- Need a thinking/reasoning mode DeepSeek V4-Flash thinking mode (via the reasoner alias until deprecation) or V4-Pro; on Mistral, Magistral Medium.
- EU data-residency or open-weight self-host requirement Mistral's open-weight models are the relevant option; confirm the hosted per-token rate before assuming a price.
- Budget against either substitute the verified per-token rate into your own loop shape; the loop discipline matters more than the headline rate.
Connects to
- Agent Cost Envelope Calculator: per-loop economics for a research agent.
- Cheapest LLM for SEC Filings 2026: DeepSeek V4-Flash in the filing-extraction context.
- Claude vs GPT-5 vs Gemini for Financial Analysis 2026: when the loop needs a frontier reasoning tier.
- Best LLM for Financial Analysis 2026: the task-tiered pillar.
References
Footnotes
-
DeepSeek. "Models & Pricing." api-docs.deepseek.com, verified 2026-05-25. https://api-docs.deepseek.com/quick_start/pricing ↩ ↩2
-
Mistral AI. "Models Overview." docs.mistral.ai, verified 2026-05-25. https://docs.mistral.ai/getting-started/models/models_overview/ ↩
-
Mistral AI. "Pricing." mistral.ai, verified 2026-05-25 (page renders client-side; per-token API rates not machine-readable). https://mistral.ai/pricing ↩
Verified engine output
Show the recompute-verified inputs and outputs
| model_id | claude-haiku-4-5 |
|---|---|
| input_tokens_per_step | 4000 |
| output_tokens_per_step | 800 |
| steps_per_loop | 6 |
| convergence_check_pct | 50 |
| markets_per_day | 150 |
| target_monthly_usd | 500 |
| calendar_mode | business |
| model › id | claude-haiku-4-5 |
|---|---|
| model › provider | anthropic |
| model › name | Claude Haiku 4.5 |
| model › tier | economy |
| model › input usd per mtoken | 1 |
| model › output usd per mtoken | 5 |
| model › cache read usd per mtoken | 0.1 |
| model › context window | 200000 |
| model › notes | Cheap filtering / pre-processing. |
| steps (7 items) | [...] |
| cost per loop | 0.052000000000000005 |
| tool use subtotal | 0.048 |
| convergence cost | 0.004 |
| cost per day | 7.800000000000001 |
| cost per month | 171.60000000000002 |
| days per month | 22 |
| tokens per loop | 31200 |
| blended usd per1 ktokens | 0.0016666666666666668 |
| within budget | true |
| budget utilization | 0.34320000000000006 |
Computed live at build time.
Frequently asked questions
- Is DeepSeek or Mistral cheaper for financial analysis in 2026?
- DeepSeek is the cheaper verified hosted API — V4-Flash at $0.14/$0.28 per Mtok with a 1M context and near-free cache hits. Mistral's exact per-token API rates are not cleanly published on its official pricing page, so a precise comparison requires confirming Mistral's live rate.
- What is DeepSeek's context window?
- Both V4-Flash and V4-Pro carry a 1M-token context window with 384k max output.
- What happened to deepseek-chat and deepseek-reasoner?
- They are now legacy aliases routing to the non-thinking and thinking modes of V4-Flash, scheduled for deprecation on 2026-07-24.
- What are Mistral's current models?
- Mistral Large 3, Magistral Medium 1.2, Mistral Medium 3.5, Mistral Small 4, and the Ministral 3 series (14B/8B/3B), per the official model documentation.
- Does the vendor or the loop shape matter more for cost?
- The loop shape — tokens per step and steps per loop — usually moves the bill more than the per-token rate. A disciplined loop on a mid-priced model beats a bloated loop on the cheapest model.