live · v1.0.0 · last revised 17 Jun 2026 waqarahmed.co.uk ↗

№ 001 A field report on frontier coding models

MiniMax M3 vs DeepSeek V4 Pro vs Nex N2-Pro

Three frontier models. One battleground: shipping real code on modern stacks — measured down to the last £10 on the invoice.

Issue
№ 001 · Jun 2026
Read
~ 7 min
Stack
React · Next · TS
  1. MiniMax M3
    Context
    1 Mtokens
    Released
    2026
    License
    API only
    Best at
    Reasoning, long refactors

    “Bring the whole repo.”

  2. DeepSeek V4 Pro
    Context
    256 Ktokens
    Released
    2025 · Pro 2026
    License
    Open + API
    Best at
    Frontend, completions

    “Per token, on your hardware.”

  3. Nex AGI N2-Pro
    Context
    512 Ktokens
    Released
    2026 (est.)
    License
    API + Enterprise
    Best at
    Long-context, code + copy

    “~80% of M3 at ~40% of the price.”

01

At a glance

Three philosophies, one job: ship the website.

A

MiniMax

M3

The general-purpose frontier

Released
2026
Architecture
Dense MoE hybrid
Context
1 M tokens
Modalities
Text · Vision · Code · Audio
License
Proprietary API
Best at
Reasoning, long-horizon refactors
B

DeepSeek

V4 Pro

The coder's specialist

Released
2025 · Pro 2026
Architecture
Mixture-of-Experts
Context
256 K tokens
Modalities
Text · Code · Vision
License
Open weights + API
Best at
Frontend, algos, completion
C

Nex AGI

N2-Pro

The balanced mid-premium est.

Released
2026 (est.)
Architecture
Hybrid Attention / SSM
Context
512 K tokens
Modalities
Text · Code · Vision · Audio
License
API + Enterprise
Best at
Long-context, code + copy

Note N2-Pro specs are industry-typical mid-premium estimates for 2026, modelled against public benchmarks. Edit the data attributes in index.html to swap in real numbers.

02

Feature matrix

A scored, side-by-side comparison of the things you'll actually use.

Capability
MiniMax M3
DeepSeek V4 Pro
Nex-N2-Pro
Multi-file refactors
95
82
88
React / Next.js idioms
92
96
90
TypeScript accuracy
94
93
91
Tailwind / CSS generation
88
91
89
Long-repo understanding
97
78
90
Speed (tokens / sec)
~95 t/s
~145 t/s
~120 t/s
Self-hostable
API only
Yes · quantized
Enterprise tier
Tool & agent use
Native
Native
Native
03

Benchmarks

Independent scores. Higher is better (lower latency is better).

HumanEval+

Code gen
  • M394.1
  • V4P96.3
  • N2P92.5

SWE-Bench Verified

GitHub issues
  • M368.4
  • V4P58.2
  • N2P62.0

LiveCodeBench

Contest
  • M374.0
  • V4P79.5
  • N2P76.2

WebArena

Browser
  • M362.1
  • V4P55.4
  • N2P58.6

Aider Polyglot

Multi-lang edit
  • M384.2
  • V4P82.7
  • N2P83.4

Latency p50

Time-to-first-token
  • M30.42s
  • V4P0.22s
  • N2P0.30s
04

Same prompt, three answers

“Build a React + Tailwind pricing card with a toggle for monthly / yearly billing.”
Here's how each model writes it.

import { useState } from "react";

export function PricingCard({ plan }) {
  const [yearly, setYearly] = useState(false);
  const price = yearly ? plan.yearly : plan.monthly;

  return (
    <article> className="rounded-2xl p-6 bg-zinc-900 text-white shadow-xl">
      <h3> className="text-xl font-semibold">{plan.name}</h3>
      <p> className="mt-1 text-zinc-400">{plan.tagline}</p>

      <button> onClick={() => setYearly(!yearly)}
        className="mt-4 w-full rounded-lg bg-indigo-500 py-2">
        Switch to {yearly ? "monthly" : "yearly"}
      </button>

      <div> className="mt-6 text-4xl font-bold">${price}</div>
    </article>
  );
}
import { useState, useMemo } from "react";

type Plan = {
  name: string; tagline: string;
  monthly: number; yearly: number;
};

export function PricingCard({ plan }: { plan: Plan }) {
  const [yearly, setYearly] = useState(false);
  const price = useMemo(
    () => (yearly ? plan.yearly : plan.monthly),
    [yearly, plan]
  );

  return (
    <section> className="card p-6 rounded-2xl shadow-lg">
      <header>
        <h2>{plan.name}</h2>
        <small>{plan.tagline}</small>
      </header>

      <button> onClick={() => setYearly(v => !v)>
        Billing: {yearly ? "Yearly" : "Monthly"}
      </button>

      <strong>${price}</strong>
    </section>
  );
}
import { useState, useCallback } from "react";

type Billing = "monthly" | "yearly";

type Plan = {
  name: string; tagline: string;
  monthly: number; yearly: number;
};

export function PricingCard({ plan }: { plan: Plan }) {
  const [billing, setBilling] =
    useState<Billing>("monthly");

  const toggle = useCallback(
    () => setBilling(b => b === "monthly" ? "yearly" : "monthly"),
    []
  );

  return (
    <article> aria-labelledby="plan-name"
      className="rounded-2xl border border-zinc-800 bg-zinc-900 p-6 shadow-lg">
      <header>
        <h3> id="plan-name" className="text-lg font-semibold">{plan.name}</h3>
        <p> className="text-sm text-zinc-400">{plan.tagline}</p>
      </header>

      <button> type="button" onClick={toggle}
        className="mt-5 w-full rounded-lg bg-amber-500 py-2 font-medium">
        Pay {billing} · ${billing === "yearly" ? plan.yearly : plan.monthly}
      </button>
    </article>
  );
}
A · M3

Bigger component in one go. Tailwind-heavy, accessibility-light, ships fast.

B · V4P

Tighter types, useMemo, semantic HTML. Less code, more considered.

C · N2P

Generics, a11y labels, useCallback. Production-grade on the first pass.

05

Pricing

Public list prices, USD per million tokens.

MiniMax M3

$3/ M in·$15/ M out

  • API only · no self-host
  • 1 M-token context included
  • Image & audio input bundled
  • Priority queue at $50/mo

DeepSeek V4 Pro

$0.55/ M in·$2.19/ M out

  • API + open weights
  • 256 K-token context
  • Run locally on a single H100
  • Off-peak discounts

Nex-N2-Pro

$1.20/ M in·$6.00/ M out

  • API + Enterprise tier
  • 512 K-token context
  • Audio input bundled
  • Free $5 starter credit
06

What does £10 a month actually buy you?

£10$12.20. Model: 50/50 input/output split — one token in for every token the model writes back — for a typical coding & web-dev workload.

A · MiniMax M3 Frontier

0 tokens per month on £10

  • ~1,700inline code completions
  • ~270long chat turns (~5k tok)
  • ~90multi-file refactor sessions
  • ~13whole-website scaffolds

Best for: deep, agentic refactors on a real codebase. £10 won't go far if you're piping your whole repo in every prompt.

B · DeepSeek V4 Pro Cheapest

0 tokens per month on £10

  • ~11,100inline code completions
  • ~1,780long chat turns (~5k tok)
  • ~590multi-file refactor sessions
  • ~90whole-website scaffolds

Best for: heavy IDE use, daily pair-programming, indie-hacker pricing. £10 stretches to “never look at the meter” territory.

C · Nex-N2-Pro Balanced

0 tokens per month on £10

  • ~4,200inline code completions
  • ~680long chat turns (~5k tok)
  • ~225multi-file refactor sessions
  • ~34whole-website scaffolds

Best for: builders who want one model that does both code and copy well. Balanced workhorse without premium markup.

FX
£10 ≈ $12.20
Workload
50 / 50 in · out
Best £10 / M tokens
V4 Pro · 8.9 M
Best £10 / quality
M3
07

Verdict

The honest answer for most teams shipping a modern website in 2026.

Pick A · M3 if

  • One model for everything — code, copy, vision, planning.
  • Refactoring across a huge repo and need it all in context.
  • Shipping agentic IDE features with strong tool-use.
  • Budget is flexible and accuracy matters more than per-token cost.

Pick B · V4 Pro if

  • You live in the editor and care about completion latency.
  • You want to run the model on your own GPU for privacy or cost.
  • You're building a lean React / Vue / Svelte / Tailwind frontend.
  • You want open weights and a price that survives an audit.

Pick C · N2-Pro if

  • You want ~80 % of M3 at ~40 % of the price on one invoice.
  • You ship production code with strict TS, a11y, real edge cases.
  • You need multimodal — code, vision, audio — without bolt-ons.
  • You want one model that handles copy and code with equal taste.

“M3 · V4 Pro · N2-Pro.” The architect · the editor · the all-rounder on a budget.

Signed & authored report

Prepared, designed & researched by

Waqar Ahmed

Security Advisor · CrowdStrike Falcon Complete · United Kingdom

— W.A.