Which is cheapest: Grok 4.5, Gemini 3.6 Flash, or DeepSeek V4?

DeepSeek V4 is cheapest at ~$0.43/$0.87 per MTok on API (cheaper self-hosted). Gemini 3.6 Flash is next at Flash-tier pricing, and Grok 4.5 is $2/$6. But Grok 4.5's ~2x token efficiency can close the gap on a per-task basis.

Which cheap model is best for coding agents?

Gemini 3.6 Flash delivers near-Pro coding at Flash-tier cost and Grok 4.5 is strong on agentic tool use with high token efficiency. DeepSeek V4 is the best if you want open weights to self-host, though it is text-only.

Is Grok 4.5 or Gemini 3.6 Flash better for agents?

Grok 4.5 emphasizes agentic efficiency — solving tasks in fewer steps — while Gemini 3.6 Flash cuts token costs up to 65% on long-horizon engineering tasks and scores well on computer use (83% OSWorld-Verified). Both are strong; Grok is cheaper per token on output, Gemini is cheaper overall.

Quick Answer

Grok 4.5 vs Gemini 3.6 Flash vs DeepSeek V4 (July 2026)

Published: July 30, 2026

The Short Answer

For cheap agentic work in July 2026, all three are strong but differ in kind. DeepSeek V4 is the cheapest and open-weight (but text-only). Gemini 3.6 Flash is the best cheap coding model with big token-efficiency gains. Grok 4.5 is the closed flagship with the best raw agentic-efficiency pitch — fewer steps per task.

Head-to-Head

	Grok 4.5	Gemini 3.6 Flash	DeepSeek V4
Maker	xAI	Google	DeepSeek
API in/out ($/MTok)	$2 / $6	Flash-tier	~$0.43 / $0.87
Context	500K	1M	1M
Weights	Closed	Closed	Open (MIT-style)
Multimodal	Yes	Yes (video-strong)	Text-only
Agentic pitch	~2x token efficiency, fewer steps	65% token cut on long tasks	Cheapest at the floor
EU availability	No (AI Act)	Yes	Yes

Where Each Wins

Cheapest at scale + open weights → DeepSeek V4. Priced at the absolute floor with a 1M context window and a permissive license. The catch: it’s text-only and has peak-hour surge pricing on the official API. Best when you can self-host or batch.
Cheap coding + computer use → Gemini 3.6 Flash. Google’s July-21 “workhorse”: near-Pro coding at Flash-tier cost, ~83% on OSWorld-Verified, and up to 65% fewer tokens on long-horizon engineering tasks. Best all-round cheap coder, and it’s available in the EU.
Cheap flagship agents + tool use → Grok 4.5. $2/$6 with ~2x token efficiency, so effective cost-per-task is lower than the sticker. Strong agentic tool use. Downside: 500K context (down from Grok 4.3’s 1M) and no EU availability.

Cost-Per-Task, Not Per-Token

Grok 4.5 looks pricier than DeepSeek V4 and Gemini Flash on paper, but its efficiency (fewer steps, fewer output tokens) narrows the real gap. Gemini 3.6 Flash’s token-efficiency gains do the same. Always benchmark completed-task cost on your own workload — the per-token leader isn’t always the per-task leader.

The Smart Play

Route by difficulty: DeepSeek V4 or Gemini 3.6 Flash for the cheap default, Grok 4.5 when you need stronger agentic tool use, and escalate to a frontier model (Claude Opus 5 / GPT-5.6 Sol) only on the hardest tasks.

Sources

xAI — Grok 4.5: x.ai/news/grok-4-5
Google — Gemini 3.6 Flash: blog.google/innovation-and-ai/models-and-research/gemini-models
VentureBeat — Gemini 3.6 Flash token savings: venturebeat.com
DeepSeek — API pricing: api-docs.deepseek.com