AI Model Pricing Comparison 2026: GPT-5, Claude, Gemini, DeepSeek, MiniMax M3, Qwen, GLM & Grok

Stop Overpaying for AI. Here’s the June 2026 Math.

Updated June 15, 2026. The AI pricing landscape moves every few days—most “2026 pricing” tables online are already wrong. Just in the last few weeks: MiniMax M3 shipped (May 31), Qwen 3.7 Max and Gemini 3.5 Flash launched (May 19), Claude Opus 4.8 arrived (May 28), GLM‑5.2 went live (June 13), Grok 4.3 undercut the field, and NVIDIA Nemotron 3 pushed hosted open models down to $0.04/1M input.

The spread today is staggering: from $0.125/1M tokens (NVIDIA Nemotron 3 Nano) to $17.50/1M (GPT‑5.5) for the 50/50 blended rate—a 140x difference. Most teams pick one vendor and overpay by 60–95%.

Here’s the current, verified pricing and the routing formula that captures the savings.

AI Model Pricing Table — Per 1 Million Tokens (June 2026)

Model	Input	Output	Blended (50/50)	vs Cheapest
GPT‑5.5 (premium)	$5.00	$30.00	$17.50	140x ❌
Claude Opus 4.8	$5.00	$25.00	$15.00	120x
Claude Sonnet 4.6	$3.00	$15.00	$9.00	72x
GPT‑5.2	$1.75	$14.00	$7.88	63x
Gemini 3 Pro (≤200K)	$2.00	$12.00	$7.00	56x
Gemini 3.5 Flash	$1.50	$9.00	$5.25	42x
Qwen 3.7 Max (promo)	$1.25	$3.75	$2.50	20x
Kimi K2.7 Code	$0.95	$4.00	$2.48	20x
Grok 4.3	$1.25	$2.50	$1.88	15x
GLM‑5	$0.60	$1.92	$1.26	10x
MiniMax M3 (launch promo)	$0.30	$1.20	$0.75	6x
DeepSeek V4 Pro	$0.44	$0.87	$0.66	5.3x
Grok 4.1 Fast	$0.20	$0.50	$0.35	2.8x
NVIDIA Nemotron 3 Super 120B	$0.10	$0.50	$0.30	2.4x
DeepSeek V4 Flash	$0.14	$0.28	$0.21	1.7x
NVIDIA Nemotron 3 Nano 30B	$0.05	$0.20	$0.125	Baseline ✅

Standard pay-as-you-go list rates. Batch mode (OpenAI, Anthropic, Google) is ~50% cheaper; prompt caching cuts repeated-context input by 80–98%. Tiered models (Gemini 3 Pro, Grok 4.3) charge ~2x above 200K context. MiniMax M3 and Qwen 3.7 Max rates are launch promos (list ~$0.60/$2.40 and $2.50/$7.50 respectively).

Key insight: GPT‑5.5 at $17.50 vs Nemotron 3 Nano at $0.125 is a 140x price difference—and DeepSeek V4 Flash or MiniMax M3 match frontier quality on most routine work for under 1% of the premium cost.

What Changed in 2026 (Why Your Old Pricing Sheet Is Wrong)

Feb 11 — GLM‑5 (Z.ai/Zhipu): $0.60/$1.92, 202K context.
Apr 24 — DeepSeek V4 replaced V3.2 entirely. V4 Flash ($0.14/$0.28) and V4 Pro ($0.44/$0.87), with a 98% context-cache discount on Flash.
May 19 — Qwen 3.7 Max ($1.25/$3.75 promo, list $2.50/$7.50) and Gemini 3.5 Flash ($1.50/$9.00, ~25% cheaper than 3.1 Pro on coding).
May 28 — Claude Opus 4.8: $5/$25, with Fast Mode down to $10/$50 (was $30/$150).
May 31 — MiniMax M3: 1M-token context, 59% SWE‑Bench Pro, at $0.30/$1.20 launch promo (list $0.60/$2.40).
June 2 — Microsoft MAI‑Code‑1‑Flash: beats Claude Haiku 4.5 on SWE‑Bench Verified using up to 60% fewer tokens.
June 13 — GLM‑5.2: 1M context, coding-first; rolling out on Coding Plan tiers (~$18/mo Lite) with standalone API pricing publishing late June.
Grok 4.3 ($1.25/$2.50, cached $0.20) plus Grok 4.1 Fast ($0.20/$0.50) — among the lowest frontier-tier rates, with $175/mo free developer credits.
NVIDIA Nemotron 3 (Super 120B $0.10/$0.50, Nano 30B $0.05/$0.20) — hosted open models now cheaper than any closed frontier API.

The lesson: don’t hard-code prices or model names. Build routing that reads a pricing config so a new release or price cut is a one-line change.

Real Cost Scenarios (Recomputed at June 2026 Rates)

Scenario 1: Customer Support (High Volume)

Setup: 1,000 queries/day · 50K input + 10K output each → 50M input, 10M output daily.

Strategy	Daily	Annual	Savings
All GPT‑5.2	$227.50	$83,038	—
All Gemini 3 Pro	$220.00	$80,300	3%
All MiniMax M2.7	$27.00	$9,855	88%
All DeepSeek V4 Flash	$9.80	$3,577	96%
Smart routing (70% DeepSeek / 20% Gemini / 10% GPT‑5.2)	$73.61	$26,868	68% + safety net

Scenario 2: Code Generation (Developer Tools)

Setup: 10,000 requests/day · 20K input + 50K output → 200M input, 500M output daily.

Strategy	Daily	Annual	Savings
All Claude Opus 4.8 (top coding)	$13,500	$4.93M	—
Kimi K2.7 Code	$2,190	$799K	84%
All MiniMax M3 (59% SWE‑Bench Pro)	$660	$240,900	95%
Hybrid (90% MiniMax M3 / 10% Claude Opus)	$1,944	$709,560	86% + quality net

Scenario 3: Document Processing (Enterprise)

Setup: 1,000 docs/day · 100K tokens each (mixed) → 100M tokens daily at blended rate.

Strategy	Daily	Annual	Savings
All GPT‑5.2	$788	$287,620	—
GLM‑5 (long-context specialist)	$126	$45,990	84%
DeepSeek V4 Flash	$21	$7,665	97%

The Savings Formula

Step 1 — Categorize tasks:

Routine (70%): predictable, high-volume, lower stakes
Complex (20%): nuanced, needs stronger reasoning
Critical (10%): high stakes, needs maximum reliability

Step 2 — Map models to tiers:

Routine → cheapest viable: NVIDIA Nemotron 3, DeepSeek V4 Flash, MiniMax M3, Grok 4.1 Fast
Complex → mid-tier: Gemini 3.5 Flash, Qwen 3.7 Max, Claude Sonnet 4.6, GLM‑5
Critical → premium: Claude Opus 4.8, GPT‑5.2/5.5, Gemini 3 Pro

Step 3 — Route from a pricing config, not hard-coded names:

PRICES = {  # blended $/1M, June 2026
    "nemotron-3-nano":   0.125,
    "deepseek-v4-flash": 0.21,
    "minimax-m3":        0.75,
    "gemini-3.5-flash":  5.25,
    "claude-opus-4.8":  15.00,
    "gpt-5.2":           7.88,
}

def route(task):
    if task.criticality == "high":
        return "claude-opus-4.8"     # quality first
    if task.complexity == "high":
        return "gemini-3.5-flash"    # strong, mid-cost
    return "deepseek-v4-flash"       # cheapest viable

Route 70% of traffic at $0.21, 20% at $5.25, 10% at $15 → blended ≈ $2.70/1M vs $7.88 all-GPT‑5.2 = 66% saved with better task-fit quality.

Hidden Costs to Watch

Token efficiency — a verbose model can emit 50% more output tokens for the same answer. Track output tokens per task type, not just sticker price.
Failure/retry rate — a $0.21/1M model with a 10% retry rate is effectively $0.23/1M; still far below a $7.88 model at 1% retries. Measure effective cost.
Caching — DeepSeek V4 Flash (98% cache discount), Kimi (~85%), and Anthropic/OpenAI prompt caching can dwarf headline-rate differences for reused system prompts.
Switching/integration time — saving $100K/year is worth one engineer-month of routing work many times over.

Frequently Asked Questions

What is the cheapest AI model API in 2026? NVIDIA Nemotron 3 Nano 30B at $0.05 input / $0.20 output per million tokens (≈$0.125 blended) is the cheapest hosted model. Among frontier-class options, DeepSeek V4 Flash at $0.14/$0.28 is the value leader—up to 140x cheaper than GPT‑5.5.

How much does MiniMax M3 cost? MiniMax M3 (released May 31, 2026) is $0.30/$1.20 per million tokens on a 50% launch promo (list $0.60/$2.40), with a 1M-token context window and 59% SWE‑Bench Pro—strong coding at a fraction of frontier cost.

What is GLM‑5.2’s API pricing? GLM‑5.2 (launched June 13, 2026) is a 1M-context, coding-first model rolling out on Z.ai Coding Plan tiers (~$18/mo Lite), with standalone per-token API pricing publishing late June. GLM‑5 remains available at $0.60/$1.92.

How much is Qwen 3.7 Max? Qwen 3.7 Max (May 19, 2026) is $1.25/$3.75 per million tokens on a 50% promo (list $2.50/$7.50), with explicit cache reads at $0.125/1M.

How much is DeepSeek V4? DeepSeek V4 Flash is $0.14/$0.28 and V4 Pro is $0.44/$0.87 per million tokens. V4 replaced V3.2 on April 24, 2026, with a 98% cache discount on Flash.

What does Claude Opus 4.8 cost? Claude Opus 4.8 (May 28, 2026) is $5/$25 per million tokens (Fast Mode $10/$50). Claude Sonnet 4.6 is $3/$15.

How much is GPT‑5 and Grok in 2026? GPT‑5.2 is $1.75/$14; premium GPT‑5.5 is $5/$30. Grok 4.3 is $1.25/$2.50 (cached $0.20), and Grok 4.1 Fast is just $0.20/$0.50.

Can I really save 90% on AI API costs? Yes—for routine and high-volume workloads. Routing predictable traffic to NVIDIA Nemotron 3, DeepSeek V4 Flash, or MiniMax M3 while reserving premium models for critical tasks typically cuts spend 60–95% with a quality safety net.

Action Plan (90 Days)

Week 1 — Audit: track current spend, categorize tasks (routine/complex/critical), measure tokens by task type.
Week 2 — Test: run parallel evals (current model vs cheaper alternatives) on quality, token efficiency, retry rate.
Week 3 — Route: send 20% of routine traffic to a budget model; monitor quality metrics.
Month 2–3 — Optimize: tune routing, add fallbacks, and move prices into config. Expected: 40–70% reduction in 90 days.